# Spectral and Spatial Methods for Hyperspectral and Thermal Image-Analysis to Estimate Biophysical and Biochemical Properties of Agricultural Crops

Authored by: Yafit Cohen , Victor Alchanatis

# Biophysical and Biochemical Characterization and Plant Species Studies

Print publication date:  December  2018
Online publication date:  December  2018

Print ISBN: 9781138364714
eBook ISBN: 9780429431180

10.1201/9780429431180-3

#### Abstract

The ultimate goals of hyperspectral and thermal sensing in precision agriculture are to estimate biophysical and biochemical properties of agricultural crops (BB-PACs) and to delineate and characterize homogeneous management zones for optimal agricultural management like such fertilization, irrigation, or other agro-technical operations. This chapter concentrates on three characteristics of hyperspectral images: (1) their unique spectral properties, (2) the spatial attribute of hyperspectral images, and (3) the state-of-the art algorithms for hyperspectral image processing that show the added value of spatial information when combined with spectral information for mapping plant biophysical and biochemical properties of agricultural crops (BB-PACs). In addition, thermal panchromatic imaging is presented as an image type complementary to the hyperspectral images. Numerous hyperspectral vegetation indices (VIs) and multivariate spectral models have been developed to estimate crop status. Thus, to address the observed interchangeability of wavelengths and indices along crops, cultivars, growth stages, years, and sites, use of relative spectral/thermal indices is suggested for delineation of management zones and creation of prescription maps for variable-rate application.

#### 3.1  Introduction

The ultimate goals of hyperspectral and thermal sensing in precision agriculture are to estimate biophysical and biochemical properties of agricultural crops (BB-PACs: phonetically pronounced as bee-bee-pax) and to delineate and characterize homogeneous management zones for optimal agricultural management such fertilization, irrigation, or other agrotechnical operations. On the face of it, the use of hyperspectral and thermal remote sensing for precision agriculture seems to be similar to their use for natural vegetation. But for natural vegetation, remote sensing is widely used for classification of natural vegetation types, while in precision agriculture it is aimed at quantification of BB-PACs. The different goals are pursued by adapted analysis approaches. This chapter concentrates on three characteristics of hyperspectral (HS) images: First, their unique spectral properties, namely the narrow bandwidths and the plethora of the bands, as opposed to wider and limited number of bands in other broad-band spectral sensing systems such most multi-spectral satellite images; Second, the spatial attribute of hyperspectral images, as opposed to point spectral measurements of other spectral systems; and third, the state-of-the-art algorithms for hyperspectral image processing that show the added value of spatial information when combined with spectral information for mapping plant BB-PACs. In addition we present thermal panchromatic imaging as an image type complementary to the hyperspectral images.

Modern agricultural crop production relies on close monitoring of the crop status. This enables efficient management of available resources for profitable and environmentally friendly agricultural practice. Widely used monitoring tools are mainly based on point sampling of biophysical and biochemical properties of the crop. Numerous crop properties have been studied over the years and act as indicators of the crop condition. Local and global growth protocols have been developed based on these measured biophysical and biochemical properties. For example, irrigation management of cotton is widely based on the height measurement of the plants at selected points; this is a biophysical property that can be easily measured by simple means, but it is labor intensive and is based on selected sampled spots. Another example is fertilization management in potatoes, where nitrate content in the petiole is used as an indication for the fertilizer requirement. Table 3.1 lists some important biophysical and biochemical properties that are used in agricultural crop growing protocols.

### Table 3.1   Biophysical and Biochemical Properties of Crops That Serve as Indicators for Agricultural Crop Management

Property (BB-PAC)

Example Crops

Agrotechnical Management Parameter

Biophysical

Biomass [kg m−1]

Wheat, rice, corn

Fertilization

Leaf area index/crop cover [no units/%]

Wheat, soybean, corn, cotton

Fertilization

Crop height [m]

Cotton, wheat

Irrigation, application of growth regulators

Canopy volume [m3]

Orchards, wheat

Irrigation, fertilization

Yield [kg m−1]

Wheat, corn, cotton

Stomata conductance [mmol s−1]

Vineyards

Irrigation

Leaf/stem water potential [MPa]

Cotton, orchards, vineyards

Irrigation

Flowering intensity [relative units]

Orchards

Growth regulators, mechanical thinning

Biochemical

Nitrogen content [%N]

Corn, wheat, potatoes

Fertilization

Chlorophyll content [μg cm−2]

Corn, wheat, cotton

Fertilization

Salinity [mmol]

Cotton

Water quality management; not used in practice

Leaf water content [%]

Wheat, potato

Irrigation

Leaf macro-elements such as phosphorus (P) and potassium (K) [mg kg−1]

Olives

Fertilization, not used in practice

These examples illustrate the great importance of monitoring biophysical and biochemical properties of agricultural crops. The desire to upgrade from point measurements to maps with high density of data has brought remote sensing to the front of the technologies that can fulfill such a mission.

A number of other crop health conditions related to crop protection, such as pest damage, plant diseases, and weed infestation, are also expressed through changes in the biophysical and biochemical properties of the crop. Several reports in the literature show the contribution of remote sensing techniques in the detection of plant diseases [1,2], pest damage [3,4], and weed infestation [5–7]. All studies report that hyperspectral remote sensing can detect the phenomena assuming that they are the factor that causes the anomalies in the field. This chapter will focus on sensing plant properties related to manageable agricultural resources such as irrigation and fertilization and will not discuss the issues of sensing plant properties related to plant diseases, pest damage, and weed infestation.

This chapter is divided into four main parts. The first (Section 3.2) describes the general characteristics of spectral and thermal sensing of agricultural crops. The second (Sections 3.3 and 3.4) reviews the most prominent methods of hyperspectral and thermal data processing to model and enhance quantification of BB-PACs. The third part (Section 3.5) describes how these methods are applied to predict specific biophysical and selected biochemical properties of agricultural crops. The last part of the chapter, presents approaches to integrate the hyperspectral data with hyperspatial attributes of the hyperspectral images to enhance their potential to delineate management zones. Additionally, in a few places in this chapter, the complementary characteristics of the VIS–NIR–SWIR and the thermal ranges are described and discussed.

#### 3.2.1  VISible and Near-Infrared (VIS–NIR–SWIR) Range

Spectral characteristics of green vegetation have very prominent features: two valleys in the visible portion of the spectrum are determined by the pigments contained in the plant. Chlorophyll absorbs strongly in the blue (450 nm) and red (680 nm) regions, also known as the chlorophyll absorption bands. This is the reason for the human eye perceiving healthy vegetation as green. When the plant is subjected to stress that hinders normal growth and chlorophyll production, there is less absorption in the red and blue regions and the amount of reflection in the red waveband increases. In some cases where stress is severe, the stress can be sensed by human eyes.

The spectral reflectance signature has a dramatic increase in the reflection for healthy vegetation at around 700 nm. In the near-infrared (NIR) between 700 and 1300 nm, a plant leaf typically reflects between 40% and 60%, of the incident radiation; the rest is transmitted, with only about 5% being adsorbed. For comparison, the reflectance in the green range reaches to 15%–20% of the incident radiation.

This high reflectance in the NIR is due to scattering of the light in the intercellular volume of the leaves’ mesophyll. Structural variability in leaves in this range allows one to differentiate between species, even though they might look the same in the visible region. Beyond 1300 nm the energy incident upon the vegetation is largely absorbed or reflected with very little transmission of energy. Water absorption bands are mostly noted at around 760, 970, 1200, 1470, 1900, and 2870 nm can be used for plant water content estimation.

#### 3.2.2  Far or Thermal Infrared (TIR)

The water pathway from the stem to leaf evaporation sites is essential for maintaining leaf water balance, allowing stomata to stay open, and resulting in carbon capture. Evapotranspiration is the process in which water stored in the soil or vegetation is converted from the liquid into the vapor phase and is transferred to the atmosphere. Evapotranspiration decreases plant temperature. Stomatal regulation plays a key role in plant response to water stress. As plant stomata close, evapotranspiration rate decreases; the energy heat balance between the vegetation and its environment is changing and leaf temperature rises. Thus, leaf temperature may be used as an indicator of plant water status and plant health. Leaf temperature can be sensed by measuring the far-infrared or thermal infrared (8–14 μm) radiation they emit. First attempts to apply canopy temperature for assessing plant water status and plant health status were made in the 1960s [8]. The availability of thermal cameras led to a significant evolution of the thermal remote sensing in the 2000s [9].

#### 3.3  Spectral Analysis Methods

Spectral and thermal remote sensing provide important information on agricultural crops. The link between the biophysical and biochemical properties of the crops and the sensing data is based on intensive data processing of the remotely sensed data using a variety of methods. A large number of processing methods have been developed over the last decades that differ in their underlying physical assumptions, the mathematical models, and how direct or indirect is the link between the data and the property. It is of great importance to understand well the basic methodology of data processing in order to ensure that the limitations and the advantages of each of method are used properly in interpretation and application of real situations. In this section, we review the most prominent methods of hyperspectral and thermal data processing to model and enhance quantification of biophysical and biochemical properties of agricultural crops. This section does not address the use of specific bands for specific properties but describes the underlying methodology for building specific relationships.

A number of methods are commonly used for analyzing spectral data to extract BB-PACs. The source of the data may be a point spectral sensor as well as a hyperspectral imaging camera. In the latter case, each pixel is regarded as a single point measurement. In both cases, there are hundreds of narrow spectral bands, with bandwidth around 1–10 nm. There are three main methods for spectral analysis: (a) bands selection, (b) spectral indices, (c) linear and nonlinear multivariate statistics and models.

Selection of individual/set of bands and the use of spectral indices were mainly developed in the arena of remote sensing, whereas multivariate statistical methods are mainly developed in the chemometrics arena.

#### 3.3.1  Spectral Bands Selection

Spectral bands selection comprises a methodology for choosing hyperspectral bands that provide sufficient, but not redundant, information to classification or prediction algorithms, using practical amount of computational resources. There are two conceptually different approaches to band selection: unsupervised and supervised. Unsupervised methods order the spectral bands without training, based on generic information evaluation approaches. They are usually very fast and computationally efficient, and can provide information for clustering an image to classes of common spectral signatures. Supervised methods require training data in order to build an internal predictive model. They are usually more computationally intensive than unsupervised methods and can provide quantitative models for predicting BB-PACs [10].

Unsupervised methods for spectral band selection include the use of such methods as principal components analysis [11] and band–band correlation [12]. Supervised methods include the use of such methods as correlation of the spectral bands with the BB-PAC studied [13], and stepwise discriminant analysis [11] to extract the number of independent wavelengths that can explain the variability of the measured BB-PAC. Both methods result in an optimum number of spectral bands that contain unique information.

#### 3.3.2  Spectral Indices

Spectral indices assume that the combined interaction between a small number of wavelengths is enough to describe the biochemical or biophysical interaction between light and matter. The simplest form of index is a simple ratio (SR), where the ratio between two wavelengths is indicative for a BB-PAC under investigation. The typical form of a SR index is

$I = R λ i R λ j$
where I is the index value, $R λ i$ and $R λ j$ are the reflectance values in wavelength λ i and λ j respectively.

Enhanced SRs are the normalized difference spectral indices (NDI or NDSI), which also exploit the difference between two distinct wavelengths, but normalize it using the following equation:

$I = R λ i − R λ j R λ i + R λ j$

Another category of spectral indices comprises integrated indices (or derivative indices), where more than two wavelengths are combined to produce a value that is correlated with BB-PACs. Integrated indices are usually specific to a certain BB-PAC and sometimes to the crop that they were developed for. An extensive compilation of all three index categories can be found in Li et al. [14].

#### 3.3.3  Multivariate Methods

Spectral indices that are based on a small number of bands are indicators of irregular conditions and provide evidence that an anomaly is present. Despite their widespread use it has not been possible to design an index that is sensitive only to a desired variable and totally insensitive to all other vegetation parameters [15]. Thus, if the factor or the cause of the anomaly in the field is known, then some of the spectral indices may be able to quantify the level or the severity of the anomaly. The advantage of the whole spectral signature of the crop is that it contains information that can be used to identify the cause for the spectral changes in the light reflected from the canopy as well as to quantify it.

Multivariate statistics assumes that there is an underlying relation between the spectral signature of the crop and its biochemical or biophysical properties. Statistical tools extract this underlying relationship as a model that is often a linear model. The large number of independent variables (wavelengths) together with the high colinearity between the variables (spectral bands) do not permit the use of common multivariate methods such as multivariate linear regression (MLR) based on least squares, before prior selection of the most indicative independent wavelengths. Therefore, methods that overcome these constraints were developed over the years. Increasing numbers of multivariate methods were adopted for processing spectral data and hyperspectral images for agriculture. Here we list a few that have been used recently for BB-PAC estimation and describe in more details the most common used for that purpose. The spectral angle mapper (SAM) algorithm determines the spectral similarity between two spectra by calculating the angle between them, treating them as vectors. The Artificial Neural Network (ANN) is a nonparametric nonlinear model that uses neural networks spreading between layers and simulates human brain receptors and information processing. ANN is a learning classification method based on large labeled (tagged) samples and is affected by the complexities of the network structure and the sample making it prone to over-learning and reducing the ability for generalization. Spectral vector machine (SVM), is a pattern recognition method which is also based on statistical learning theory [16]. Another machine learning algorithm is the random forest (RF) which is designated for classification or regression tasks [17]. For classification tasks, it is operated by constructing a multiple decision trees based on iterative selection of training samples. As the random selection is sensitive to selected feature dimensions (or insensitive to some feature dimensions), the trees can gain accuracy as they grow without suffering from over-fitting.

Additionally, there are two more common methods: principal components regression (PCR), which has a core of unsupervised data extraction, and partial least-squares regression (PLSR), which is a supervised method. Both methods produce a linear model.

Partial least-squares regression is related to both PCR and MLR, and can be thought of as occupying a middle ground between them [18]. PCR finds factors that capture the greatest amount of variance in the predictor variables (spectra). MLR seeks to find a single factor that best correlates predictor variables with predicted variables (BB-PACs). PLS attempts to find factors that both capture variance in the predictor variables (spectra) and achieve correlation while avoiding the colinearity between spectral bands. In other words, PLS attempts to find factors (called latent variables) that maximize the amount of variation explained in the spectra that is relevant for predicting the BB-PAC. This is in contrast to PCR, where the factors (called principal components) are selected solely based on the amount of variation that they explain in spectra. In mathematical terms, the difference between them is the objective function that is used to optimize the calculation of the regression coefficients. Unsupervised methods tend to minimize only the inter-class (between classes) variance based on the spectral curves of the samples. Supervised methods either minimize the variance of the intra-class (within the class) variance or a combination of the inter-class and intra-class variance.

Wavelets are a group of functions that vary in complexity and mathematical properties and that are used to dissect data into different frequency components and then characterize each component with a resolution appropriate to its scale. Wavelet analysis of a reflectance spectrum is performed by scaling and shifting the wavelet function to produce wavelet coefficients that are assigned to different frequency components. By selecting appropriate wavelet coefficients, a spectral model can be established between the coefficients and biochemical concentrations. Hence, wavelet analysis has the potential to capture the information contained within high-resolution spectra and offers the prospect of developing robust, generic methods for pigment determinations [19,20].

It should be noted that most of these methods are confined to classification and detection problems and are not often used for quantitative estimation of crop characteristics from hyperspectral data.

#### 3.4  Thermal Analysis Methods

Currently, thermal cameras provide either panchromatic images or multi/super-spectral images in the range 3–14 μm. This section concentrates on the analysis of panchromatic thermal images in the range of 8–14 μm that are used in most agricultural studies. Since they are panchromatic images there is no dimensionality complexity. For BB-PACs estimation, the core of the thermal image analysis is to convert the surface temperature to meaningful water status indices. Maes and Steppe [9] provide a comprehensive review on ground-based thermal imaging for estimating evapotranspiration and water shortage stress, while this section focuses on thermal imaging analysis for water status estimation and mapping.

#### 3.4.1  Computation of Crop Water Stress Index

The use of canopy temperature as an indicator of plant water status is not new and was popularized by Idso and colleagues [e.g., 21,22]. Since canopy temperature is affected by both plant water status and environmental conditions, water stress indices that calibrate the environmental conditions were developed. The crop water stress index (CWSI) based on canopy temperature [22] has become an acceptable index to map in-field variability of crop water status using thermal images. CWSI is defined as a fraction of the canopy temperature between dry (upper) and wet (lower) baselines under ambient conditions. It can be calculated by:

$CWSI = T canopy − T wet T dry − T wet$
where T canopy is the canopy temperature, T wet is the temperature of a fully transpiring leaf, and T dry is the temperature of a nontranspiring leaf. For irrigation scheduling, CWSI mapping based on aerial thermal images should be simple to compute in order to be used in the routine of irrigation management. Accordingly, there are two main challenges in CWSI computations that researchers have addressed in the last two decades: (i) development of a methodology for accurate extraction of canopy temperature; and (ii) the setup and formulation of baselines (T wet and T dry) that can be accurately measured, extracted, or computed.

#### 3.4.1.1  Canopy Temperature Extraction

An object-oriented methodology for pure canopy temperature extraction using merely thermal images has been suggested for trees [23–25]. The methodology suits some orchard structures that have soil between crop rows, since the canopy temperature is well differentiated from the exposed soil. Nevertheless, in orchards that have grass in between the crop rows, thermal imaging cannot easily be used to differentiate between grass and tree canopy pixels. Fusing a digital surface model or information on the rows’ location with the thermal images is suggested to address this challenge.

For field crops like cotton and wheat, the main challenge is the extraction of mixed pixels of canopy and soil. Methodologies that combine multispectral (MS) images in the VIS-NIR range with thermal images have been developed to extract canopy pixels [26,27]. An empirical methodology for canopy temperature extraction using only thermal images and air temperature was developed for field crops by Meron et al. [28]. This methodology requires only thermal images and air temperature and may be suited for orchards as well.

#### 3.4.1.2  Forms of Wet and Dry Baselines

Empirical and theoretical (analytical) forms of wet and dry baselines have been proposed and used for CWSI calculation and mapping, as summarized in [9,29]. For large scale CWSI mapping, dry baseline temperature was used solely in its empirical form, that is, air temperature + X°C. The canopy–air temperature difference is unique for each crop in each region, but it is relatively stable and indifferent to changes in vapor pressure deficit (VPD). In comparison, wet baseline temperature determination is a greater challenge for researchers as it is highly dependent on vapor pressure; thus it can be found mostly in its empirical (as a function of air temperature and VPD) and theoretical (based on the energy balance [30]) forms. Berliner et al. [31] and Taghvaeian et al. [32] have shown the potential in using a well-watered reference plot to measure the wet baseline temperature, but they used it for canopy and air temperature difference index and not for CWSI. This approach uses the crop as a bioindicator but, instead of using a single leaf as in ground thermal imaging [25], it uses a set of pixels from a field. Another bioindicator wet reference, named statistical or virtual reference, has been suggested more recently, which uses the average temperature of the coolest 5%–10% of the canopy pixels [23,29,33,34]. The statistical reference assumes that at the time of thermal imaging, there are areas in the field that are over-irrigated.

#### 3.4.2  Satellite and Aerial Thermal Images

A trade-off exists between satellite and aerial thermal imaging in terms of cover area and spatiotemporal resolution. Satellites provide images covering large areas at a low cost per area unit and have thus became a common tool used by farmers. Currently, satellite-based images in the VIS–NIR range have a relatively high spatiotemporal resolution, but in the thermal range their finest resolution (60 m in Landsat) and their long revisit time are often not appropriate for irrigation management (Table 3.2). Aerial-based thermal images, which, theoretically, can be acquired on a daily basis, have high spatial resolution but are limited by their cover area (limited capacity), and are thus expensive per area unit. A revival of thermal imaging for water status mapping has been sensed lately with the increasing availability of compact, low-cost uncooled microbolometer-based thermal focal plane arrays. These cameras can be easily mounted on unmanned airborne vehicles (UAVs) (or even integrated into smartphones). However, being noncooled, they suffer from temporally and spatially dependent changes that require constant calibration of both the gain and offset. With the absence of a means of internal calibration, they cannot be used for trustworthy assessment of canopy temperature and thus reliable estimation of the crop water status. Most of the compact thermal cameras are very sensitive (thermal resolution of 0.1°C degrees or more). Yet, only a few have sufficient accuracy (±2°C and better) while most of them lack it (±5°C and worse). Even the more accurate compact cameras suffer from a significant drift of the temperature. To our knowledge, currently there is no compact low-cost camera that has appropriate calibration hardware. New publications may have paved the way for retroactive calibration of such cameras [35,36] but their applicability in real conditions have yet to be proved. Enhancement of the spatial resolution of the aerial thermal images can also be achieved by employing super-resolution algorithms that exploit the vast overlap between sequential images [37]. If applicable, aerial thermal images may be acquired from higher altitudes and cover much larger areas. For larger fields, it was proposed that sharpening methods would be adjusted and applied for thermal images of Sentinel-3, which has a revisit time of a few days [37]. Finally, a novel approach was introduced that fused aerial thermal images with satellite MS images in the VIS–NIR–SWIR range [37]. A similar method was employed to upscale aerial hyperspectral images for natural vegetation monitoring [38].

### Table 3.2   Satellite Thermal Bands, Their Spatial Resolution and Revisit Time

Satellite

Spatial Resolution (meters)

Revisit Time (days)

Landsat 7 and 8

60–100 (30) a

16 (8) b

ASTER

90

16

MODIS

1000

1

Sentinel-3 (SLSTR) c

1000

1

Notes:

a  Bands are acquired at 60 or 100 m resolution (in Landsat 7 and 8, respectively), but are resampled to 30 m in delivered data product.

b  Landsat 8 satellite images the entire Earth every 16 days in an 8-day offset from Landsat 7.

c  Sea and Land Surface Temperature Radiometer.

#### 3.5.1.1  Leaf Area Index, LAI

Green leaf area index (LAI) is a key variable used by crop physiologists and modelers for estimating foliage cover, as well as forecasting crop growth and yield. The exposed area of living leaves plays a key role in various biophysical processes such as plant transpiration and CO2 exchange. Because LAI is functionally linked to the canopy spectral reflectance, its retrieval from remote sensing data has prompted many investigations and studies over the years [19,39–44]. Most of these studies have relied on empirical relationships between the ground-measured LAI and observed spectral responses.

The most common index to estimate LAI and its counterparts, the crop cover and biomass, is the normalized differential vegetation index (NDVI) [15], which expresses the normalized ratio between the reflected energy in the red chlorophyll absorption region and the reflected energy in the NIR mesophyll scattering region. Yet, it is well documented that the NDVI approaches saturation asymptotically under conditions of moderate-to-high above-ground biomass [43], it therefore may be a good predictor only for low to medium LAIs (0–4).

Linear regression analysis of single bands and two-band combinations of pseudo NDVIs (NDSIs) have shown the importance of the red-edge spectral region (700–740 nm) and the short-wave infrared (SWIR) spectral region, and the advantage of narrow bands over traditional broad bands in LAI prediction [13,41,42,45,46]. A major problem in the use of indices to estimate LAI arises from the fact that canopy reflectance, in the visible and NIR, is strongly dependent also on chlorophyll content of the canopy [e.g., 47]. Moreover, both variables have similar effects on canopy reflectance, particularly in the spectral region from the green (550 nm) to the red edge (740 nm). To uncouple the LAI effect, Haboudane et al. [40] developed two indices: the modified triangular vegetation index (MTVI2) and the modified chlorophyll absorption ratio (MCARI1). Prediction algorithms based on these two indices were applied for CASI hyperspectral image over fields of soybean, corn, and wheat and showed excellent agreements between modeled and measured LAI.

Other studies exploit wider ranges of the spectra or even the whole spectrum to improve LAI prediction. Delegido et al. [48] have shown that the spectra between 500 and 750 nm can be fitted with good precision to third-degree polynomials and that there was strong correlation between one of the coefficients and LAI values that ranged from 0 to 7. This is a significant improvement over other methods since it covers the whole range of LAI (0–7) and is not limited to low (0–2) and medium (2–4) LAI. Multivariate and PLS regression models based on selected narrow bands or the whole spectrum, respectively, have been shown to be comparable or better LAI predictors than narrow-band NDIs [41,42]. While narrow-band NDVI had strong correlation in LAI range of 0–3 and explained 80% of the LAI variability, the multivariate regression of wider range had a very high correlation in LAI range of 0–6 and explained 90% of the variability [41,49]. In another study, several multivariate methods were used to predict LAI in soybean, namely, RF, ANN, SVM, and PLS [50], and all methods demonstrated similar performance measures, explaining around 70% of the LAI variability.

#### 3.5.1.2  Biomass

Forecasting and estimating of crop production using remote sensing has great consequence on food provision management and is fundamental to applications of precision agriculture. In-season biomass estimation from remote sensing for yield forecasting and variable rate applications has been a challenge for various studies.

Biomass and LAI have similar effects on spectral characteristics and studies have shown similarities as well as some differences in the estimation of both crop properties using spectral measurements and hyperspectral images. Correlation coefficients between spectral reflectance in discrete narrow bands and LAI and biomass in various crops presented similar shapes [13,45]. Spectral bands that are best suited for characterizing LAI and biomass were determined by Thenkabail et al. [13,45] and no significant differences on their prediction accuracy was found. Yet, while no improvement was achieved by using PLS regression models for LAI estimation, PLS models significantly improved the prediction of biomass by lowering the RMSE by 22%, compared to the best narrow-band indices [42]. Correspondingly, PLS models using the spectral range of 350–2500 nm were found to better predict wheat dry biomass compared to common vegetation indices: R 2 values of 0.80 and 0.50, respectively [51]. A recent study, that used snapshot hyperspectral images from a hyperspectral camera mounted on a UAV, showed no advantage of PLS models over the best narrow-band indices (R 2 = 0.5) to predict biomass of winter wheat canopies, but a significant improvement was achieved where crop heights and the full spectra were combined in a PLS model (R 2 = 0.78) [52].

#### 3.5.1.3  Water Status

Crop water status is a key biophysical property that is used to manage irrigation, as well as to evaluate crop health. In most cases, it is directly associated with water availability in the soil, and when this is not the case (i.e., water availability is not the limiting factor), water status becomes an indicator of crop health. For example, when salinity is a limiting factor of water uptake, crop water status becomes an indicator of salinity stress. Similarly, plant diseases that damage water flow in the plant affect the crop water status, which becomes an indicator of the disease's presence or its severity.

Crop water status is a function of soil water availability, hydraulic resistance along the flow path, plant water capacitance, and meteorological conditions that determine atmospheric evaporative demand [53]. Crop water status can be quantified by measuring either leaf water content or leaf and stem water potential. The spectral characteristics of water can be used to quantify the water content in the leaves. For wavelengths sensitive to water absorption (760, 970, 1450, 1940, and 2950 nm), leaf reflectance decreases as water content increases. Numerous studies have shown the ability of spectral indices to determine leaf relative water content (RWC), for example, the early study of Hunt and Rock [54], the study of Ceccato et al. [55], and more recent studies such as [56–58]. In a few studies, attempts to use indices as algebraic expressions of reflectance values for specific wavelengths did not yield significant relationships at the canopy level [51,59]. Nevertheless, when methods that use the whole spectrum were analyzed, canopy water content could be predicted from remotely sensed data. Namely, PLS models based on the first derivative of the spectrum in the range 350–2500 nm predicted water content with R 2 of 0.87, while spectral indices with exponential model achieved R 2 = 0.2 [51]. PLS models of spectral curves were found best predictors of RWC in comparison to various spectral indices and of other multivariate spectral models like MLR [60]. In addition, when the water absorbance band at 970 nm was considered, leaf water content was successfully predicted based on the slope (first derivative) of the spectral curve at 1015–1050 nm (R 2 = 0.97 for simulated data and 0.68 for field data) [61]. Other methods that consider the entire wavelength spectrum between 700 and 1300 nm showed that nonlinear models based on radial basis functions produce considerably better results than linear regression models (relative error of 4% and 17%, respectively) [62]. This outcome might indicate the existence of a complex dependency relationship between reflectance and leaf water content. It might also explain the poor results obtained by some methods based on indices in other studies.

Leaf water potential (LWP) in crops and stem water potential (SWP) in orchards are important biophysical parameters that indicate the ability of the crop to transfer water from soil to leaf [63]. The reports in the literature show limited ability to remotely estimate them using hyperspectral sensing in the VIS/NIR region since they express the physical status of water potential in the plant tissue [59,64,65]. Nevertheless, they affect the status of the leaves’ stomata, which control the evapotranspiration process and affect leaf temperature. An important consequence of the stomatal closure that occurs when plants are subject to water stress is that energy dissipation is decreased, so leaf temperature tends to rise [30]. As mentioned above (in Section 3.4), the most common and widely utilized thermal index is the CWSI [22]. In the last decade, thermal crop sensing technologies have been widely used as tools for monitoring and mapping crop water status in various orchards [34], grapevines [23,66], olives [67,68], almonds [69], and other various crops [26,29,70,71]. Furthermore, thermal sensors and imaging have been employed for uniform and variable-rate irrigation management [72–76]. There are only a few studies that combined hyperspectral spectral images in the VIS–NIR range with thermal images for the prediction of various BB-PACs [e.g., 77,78]. A recent study integrated thermal imaging with hyperspectral sensors to assess their relationship with water status and grain yield of wheat cultivars [68]. The results show that the normalized relative canopy temperature (NRCT) alone was closely and significantly associated with RWC, with canopy water content and with grain yield (R 2 = 0.81 and R 2 = 0.87). The data fusion model of PLSR based on selected spectral indices improved the yield prediction under three irrigation regimes (R 2 = 0.97). A scientific report* on the fusion of hyperspectral images in the range of 400–980 nm and panchromatic thermal images in the range of 8–14 μm for estimating and mapping nitrogen level and water status has shown that the two ranges have complementary characteristics. In this study, a two-factor experiment of different nitrogen and irrigation treatments was conducted in potato fields, and it was found that (1) the spectral index NDI [79] extracted from the hyperspectral images acquired on two different dates was significantly affected by nitrogen treatments; (2) the water index 900/970 [80] was not affected by irrigation treatments; and (3) canopy temperature was sensitive to irrigation treatments while insensitive to nitrogen treatment. From the plentiful studies that have assessed either thermal images or spectral reflectance sensing and imaging or both to estimate water status, it can be concluded that thermal imaging has the ability to detect minor and mild water stress while spectral sensing technologies in the VIS–NIR range are more capable of detecting water stress in more advanced stages.

#### 3.5.2.1  Chlorophyll Content

The most commonly used biochemical property of crops is chlorophyll content. It reflects the general condition of the crop, since chlorophyll is the producing “factory” of the crop. Changes in chlorophyll may indicate limited availability of important elements, among a wide possibility of options or other biotic or abiotic stresses. Chlorophyll deficiency can be detected by remote sensing, using specific spectral indices. Nevertheless, detection of chlorophyll deficiency is not an indicator of the cause that induced the deficiency.

Chlorophyll-specific spectral indices can be divided into two categories: (a) indices based on chlorophyll absorption in the blue (around 450 nm) and red (around 680 nm) spectral region and (b) indices that are based on the displacement of the red edge inflection point (700–740 nm). Several reports in the literature describe the use of simple and combined spectral indices for leaf chlorophyll estimation [20,81]. Among the set of indices tested, index combinations such as modified chlorophyll absorption ratio index/optimized soil-adjusted vegetation index (MCARI/OSAVI), triangular chlorophyll index/OSAVI, Moderate Resolution Imaging Spectrometer terrestrial chlorophyll index/improved OSAVI (MSAVI), and red-edge model/MSAVI seemed to be relatively consistent and more stable as estimators of crop chlorophyll content [20].

Chlorophyll content was also estimated using wavelet decomposition on hyperspectral data. In the context of remote sensing of foliar chlorophyll, wavelet analysis has the potential to capture much more of the information contained with reflectance spectra than previous analytical approaches that use a small number of optimal wavebands. This approach was found to be more reliable than simple linear regression analysis when linking chlorophyll to the reflectance measured. This was observed both for leaf-level measurements as well as top of canopy measurements (peach trees) [82]. The wavelet-based approach outperformed models based on untransformed spectra (such as stepwise derivative) and a range of existing spectral indices. While wavelet-based models yielded 1:1 relationships between measured and predicted chlorophyll content in the range of 0–60 μg cm−2 (with R 2 of 0.88), other methods (including indices and first derivative) saturated above 30 μg cm−2 [83]. These findings indicate that wavelet analysis warrants further investigation as a method for extracting meaningful quantitative information from hyperspectral data.

Refinements in the technique for quantifying chlorophyll could explore the use of new wavelet functions or combinations of functions, multiple scales of wavelet coefficients, alternative methods for calculating derivatives prior to wavelet decomposition and different approaches to the selection of wavelet coefficients during model calibration. The value of wavelet analysis of spectra for quantifying leaf chlorophyll in principle has been demonstrated; it is now important that this is tested in practice and that the generality of the technique for hyperspectral remote sensing of vegetation is explored, particularly at the canopy and landscape scales [83].

The approach providing the highest predictive accuracy was that using multiple regression models based on wavelet coefficient energy feature vectors. This was closely followed by multiple regression models derived from the energy feature vectors of the nth-largest wavelet coefficients, which in turn was closely followed by stepwise regression models based on wavelet coefficients. The predictive accuracy of the stepwise regression models derived from narrow-band reflectance was substantially lower than that of the wavelet-based approaches, and the simple ratio and normalized difference ratio spectral indices had the poorest performance by some margin [84]. A number of techniques have been developed for red edge position extraction in the past. A more recent one suggested by Dong et al. [85] is a wavelet-based technique.

Leaves that suffer from chlorophyll shortage may have discolored spots, resulting in a range of spectral properties from a single leaf. Liu et al. [86] have demonstrated that the entropy, standard deviation of the spectral properties used, and spatial features were very good indicators of the leaf chlorophyll content. They concluded that spatial information can be used to retrieve chlorophyll content, with an accuracy equivalent to that of spectral information, and can provide information that spectral reflectivity cannot provide.

#### 3.5.2.2  Nitrogen Content

Nitrogen deficiency is one of the most important conditions to be detected, since it directly affects the productivity of the crop. An additional reason that makes the detection of nitrogen deficiency very important is the fact that nitrogen leaches under the root zone when irrigation or water management is not appropriate, creating conditions that are suboptimal for crop growth.

Nitrogen (N) indices can be divided into indices that are based on wavelengths in the visible and the NIR region, and indices that include specific nitrogen absorption wavelengths in the SWIR. The additional value of using SWIR-based indices has been shown in studies on wheat in which a firm advantage was revealed for the proposed SWIR-based indices in their ability and sensitivity to predict N content in potato leaves [87].

Many hyperspectral vegetation indices (VIs) have been developed to estimate crop nitrogen status at leaf and canopy levels. They have been evaluated for different growth stages and years using data from both nitrogen experiments and farmers’ fields. Furthermore, to identify alternative promising hyperspectral VIs, evaluation of all possible two-band combinations of SRs and NDIs has been performed. The results indicated that best-performing published and newly identified VIs included simple ratios in the red edge region and in the blue region [11,14]. Red edge and NIR bands were more effective for nitrogen estimation at early growing stage, but visible bands, especially ultraviolet, violet, and blue bands, were more sensitive at later growing stage.

Across sites, years, cultivars, and growth stages, the combination of wavelengths in the blue range (370 and 400 nm), as either simple ratio or an NDI, performed most consistently in both experimental and field data for wheat [14]. Together with green, red, NIR, and red edge, the blue range was found sensitive to nitrogen content in rice in another extensive study that integrated ground based spectral data from different sites and years [88]. Yet, in the same study, it was found to be insensitive to nitrogen content when Hyperion images were utilized. For Hyperion images, only the red edge and NIR were found sensitive.

In their study for detecting nitrogen stress in two potato cultivars, Tyler et al. [89] reported that canopy-scale spectral data can distinguish between N treatments better than tissue samples and that among several spectral indices Medium Resolution Imaging Spectrometer (MERIS) terrestrial chlorophyll index (MTCI) [90] was the best spectral index to be used for variable rate nitrogen prescriptions in potatoes. MTCI extracted from Hyperion satellite images exhibited the best relation (logarithmic) to N content also in rice in comparison to more than 50 published two- and three-bands indices [88]. In terms of prediction ability, MTCI performed slightly better than other published indices but significantly worse than two and three-band indices proposed by Tian et al. [88]. Moreover, the two new indices performed well using ground spectra, modeled airborne visible/infrared imaging spectrometer (AVIRIS) spectra, Hyperion spectra and acquired Hyperion images. Despite that, a newer study that analyzed the spatial variability of chlorophyll and N content of rice from Hyperion imagery in India found different relationships between N content and the indices suggested by Tian et al. [91]. Moreover, their modified index has shown significantly wider range of predicted N content than the index suggested by Tian et al. and was thus better for mapping the spatial variability of N content.

The majority of the indices predicting N content are based on indirect indicators, mostly chlorophyll content, which is proven to be physiologically linked to N content. Herrmann et al., explored the performance of new N spectral indices dependent upon the SWIR (1200–2500 nm), and particularly the 1510 nm band because it is related to N content [92]. The results revealed a firm advantage for the SWIR-based indices in their ability to predict and in their sensitivity to N content. The best index is one that combines information from the 1510 and 660 nm bands, but no significant differences were found among the new SWIR-based indices.

Beyond the differences between crops, sites, and years, growth stage also had a significant influence on the performance of different vegetation indices and on the selection of sensitive wavelengths for leaf nitrogen estimation. The observed interchangeability of wavelengths and indices along growth stages and cultivars may be addressed by multivariate methods, which make use of the whole spectrum and not only selected wavelengths. For instance, multivariate methods were used to estimate leaf nitrogen content based on narrow-band spectral data in potatoes. PLSR analysis has resulted in a stronger correlation between predicted and measured leaf nitrogen content (R 2 = 0.95) than the nitrogen-specific transformed chlorophyll absorption reflectance index (TCARI) (R 2 = 0.82), even though in both models data from narrow bands was used [93]. Moreover, with PLS the improved correlation was achieved with a single model for both the vegetative and the tuber-bulking periods, while the TCARI yielded a different model for each period [93]. In the same study, when the number of wavelengths was reduced from 400 to 11, and the bands’ bandwidth was broadened from 1.3 to 20–40 nm, in order to simulate the Venμs satellite data, the accuracy of the spectral model was decreased (R 2 = 0.78), yet still included both vegetative and tuber-bulking periods. Similar results were obtained for nitrogen prediction in winter wheat. Models based on NDVI had an exponential characteristic, which implies saturation for high nitrogen values, and low coefficient of determination (R 2 = 0.15). When the derivative of the spectrum between 350 and 2500 nm was used in conjunction with PLSR models, the coefficient of determination was significantly better (R 2 = 0.82).

Another approach was suggested to address the observed interchangeability of wavelengths and indices along crops, cultivars, growth stages, years, and sites. A nitrogen sufficiency index (NSI) was applied to leaf N concentration and spectral indices/models to normalize them for comparative purposes between spectral indices and PLS prediction models [89]. Applying the NSI formula to spectral data made it insensitive to external factors such as cultivar and growth stage. In practice, it means that for a proper use of hyperspectral data the farmer should keep N-rich areas within the field [94] to be used for NSI estimation and N prescription maps for variable N application rate. Adapting the N-rich areas concept to commercial production practices might seem to be straight forward but the N-rich plants are likely to develop differently from the remaining field and do not represent the normal canopy. The virtual reference concept uses a histogram to characterize and display the spectral data from which the vegetation index of adequately fertilized plants can be identified [95]. As described in Section 3.5.3, this approach was also suggested and successfully tested for wet-temperature reference for water status estimation.

#### 3.5.3  Suggested Approaches in Predicting BB-PACs

Key BB-PACs like LAI, chlorophyll level, and water status have major effect on transition zones of the spectra reflectance curve such as the red edge and water absorption bands. Thus, the narrow band widths of hyperspectral data allows for better estimation of crop properties compared with the relatively coarse bandwidths acquired with multispectral scanners. While hyperspectral images in the VIS-NIR range provide a tool for estimating and mapping various BB-PACs in the fields, they are limited in assessing and mapping crop water status parameters that are essential for irrigation management. To that end, integration of images from the thermal range is required.

With the advances in hyperspectral technologies, practical issues related to data volumes and data-processing emerged. The processing complexity and the statistical concerns of colinearity and over-fitting entailed in spectral analysis have led to the widespread adoption of the dimensionality reduction approach. Various narrow-band indices were developed and were shown to improve the broad-band indices. Step-wise discriminant analysis was used in many studies to select a few optimal bands for characterizing agricultural crop variables. In general this type of analysis had demonstrated the importance of the red, the red edge, and the SWIR regions and, to lesser extent, the blue, green, and NIR regions. These findings together with the high cost of hyperspectral systems and the analytical complexity promoted the development of super-spectral platforms such as the Rapid-Eye, the World-View2, and the Venμs (launched on July 2017).

Despite the similarities found in the literature, the selected bands were not identical for the same crop property in different sites nor were they identical to different crop properties in the same site. Moreover, beyond the differences between crops, sites, and years, growth stage also has been shown to have a significant influence on the performance of different vegetation indices. Finally, beyond issues related to calibration, accuracies, and operational characteristics of the sensors, the leaf or canopy reflectance, in the visible–NIR–SWIR ranges, is highly dependent on both biophysical and biochemical properties. Moreover, several properties have similar effects on canopy reflectance. It means, for example, that while N content is the desired property, information provided by analyzed spectral measurements would be biased by factors other than N, such as water status, stand density, and pests. In view of this, we doubt the utility of pursuing the approach of the best set of bands or the best index. In other words, in our opinion, there is no global set of bands or global index for predicting any BB-PACs. For this reason, we do not provide a table that summarizes specific bands, spectral indices, and spectral ranges for the various key BB-PACs. Instead we present Figure 3.1, which generally and very coarsely shows single bands and band ranges that have been used in the cited studies for estimating LAI and biomass; water content; and nitrogen and chlorophyll levels. Rather than using a set of single bands or indices, this overview strongly demonstrates the advantage of hyperspectral systems that provide contiguous spectra using multivariate analysis techniques. With recent developments in compact hyperspectral sensors and compact uncooled thermal cameras, combined with available UAV that can carry them [36,96,97], new horizons for hyperspectral and thermal data are opened. Snapshot hyperspectral cameras were used to create radiometrically calibrated hyperspectral data [98,99] and even provide 3D hyperspectral information for vegetation monitoring [99]. To make these systems affordable for agricultural stakeholders, methodologies that fuse hyperspectral or thermal aerial imaging with MS satellite imaging should be developed [37,38].

Figure 3.1   Spectral bands and spectral ranges that were used in various studies to estimate key BB-PACs. Vertical bars refer to single bands and horizontal lines refer to band range.

Whether analyzed contiguous spectra or spectral indices are used, normalization procedures such as the well-fertilized reference plots for N level estimation or statistical wet reference for water status estimation, seem to be inevitable. In the view of the cited studies, it is deduced that absolute estimation of any BB-PACs is unachievable unless a reference area or reference data are available. Based on such references, the approach of a sufficiency index that was introduced for nitrogen variable rate application (nitrogen sufficiency index, NSI) should be utilized for other BB-PACs.

#### 3.6  Spatial Methods

In the early days of hyperspectral imaging, hyperspectral data processing techniques focused on analyzing the spectral data without incorporating information on the spatially adjacent data. In other words, hyperspectral data were usually treated not as images but as unordered listings of spectral measurements with no particular spatial arrangement [100]. The importance of analyzing both spectral and spatial patterns has been identified as a desired goal by many scientists devoted to multidimensional data analysis. This type of processing has been approached from various points of view representing different levels of combination between spectral and spatial information.

Nearly all of the methods combining spectral and spatial information were developed for land cover classification. General reviews and illustrations of spectral–spatial classification methods of hyperspectral imagery can be found in Fauvel et al. [101] and Plaza et al. [100]. In general, the spectral–spatial methods seek to reduce the salt-and-pepper appearance of the classification; to use spatial characteristics (such as entropy and standard deviation (STD)) and spatial features (such as size and shape) to enhance the separation ability between classes; and to perform image segmentation prior to the classification to define a spatial neighborhood for the pixels. There is a basic difference between land use/cover classification and estimation of BB-PACS. Land use/cover types are discrete elements with relatively well-defined borders. Moreover, most of them have relatively distinct spectral signature. In comparison, biophysical and biochemical crop properties are continuous variables, with smooth differences in spectral signature and with amorphous shapes.

The spatial and temporal variability of soil and crop factors within a field is the factual base of precision agriculture [102]. Opportunities exist to use airborne hyperspectral and thermal imaging for mapping the spatial variability of crop properties in agricultural fields. Maps of nitrogen and water status can then be used to delineate management zones for fertilization and irrigation variable-rate application. Delineating management zones involves spatial filtering to reduce effects of noise in measurements of individual factors and removal of excessive details in within-field variability to simplify the shapes and size of the zones. Methods combining spectral and spatial information in studies designated to estimate levels of crop biophysical and biochemical properties and to divide them into homogenous/management zones are scarce. Indeed, spectral models manipulated over hyperspectral images were used to create maps of biophysical and biochemical crop properties [40,103,104] or further to partition fields into management zones based on spectral properties [105]. In the creation of the maps, smoothing operations were applied for reducing the speckle effect. Yet, all of the maps were created merely based on a pixel-by-pixel spectral data without incorporating information on the neighboring pixels. Here we describe potential approaches and illustrate methods for combining spectral and spatial information for segmentation of hyperspectral images based on spectral-based crop properties.

#### 3.6.1  Hyperspectral Data Set

To illustrate the potential that lies in some of the described methods we used two aerial hyperspectral images taken over an experimental potato plot under different nitrogen treatments and over a commercial potato field. For the commercial field, an aerial thermal image was also available. The experimental plot and the commercial field were planted with cv. Desiree in Kibbutz Ruhama, Israel (31.388N, 34.598E). To assess a range of N levels, five treatments were applied (Table 3.3) with four replicates. Each replicate was 18 m wide (six rows) by 50–100 m long. More details on the overall study can be found in [93]. The experimental plot represents a controlled area with a relatively wide range of N levels with known borders. In contrast, the commercial field represents spatial variability that its range and spatial pattern are unknown in advance.

### Table 3.3   Nitrogen Treatments Applied in the Potato Field in Spring 2007

Nitrogen Treatment

N Rate (kg ha−1)

Percentage N Rate Relative to Commercial Rate

Yield

T100%

400

100

52 c

T75%

300

75

52 c

T50%

200

50

52 c

T25%

100

25

45 b

T0%

0

0

36 a

Hyperspectral images above the experimental plot and the commercial field were acquired on May 25, 2007 and April 24, 2012, respectively. AISA Eagle hyperspectral imaging push broom sensor (Spectral Imaging Ltd., Oulu, Finland) in the range of 400–970 nm, with 420 bands with spectral resolution of 1.3 nm. The image was acquired from 500 m height and had 1 m spatial resolution. Preprocessing of the image included selection of every second band of the original 420 bands and smoothing of the 210-band spectra of the new cube, with a 15-points window.

Figures 3.2 and 3.3 are an RGB and false color images of the experimental plot derived from the narrow-band hyperspectral image. The false color is overlaid by the borders of the N treatments.

Figure 3.2   RGB (670 nm, 550 nm and 420 nm) image of the experimental plot.

Figure 3.3   False color image of IR (750 nm), red (670 nm) and green (550 nm) bands of the experimental plot overlaid by the borders of the N treatments.

#### 3.6.2  Spatial Information as a Preprocessing Tool

Individual spectra of the same object or property taken from neighboring pixels in the hyperspectral image present relatively high variability. Figure 3.4 shows individual spectra for regions of interest (ROIs) taken from sub-plots 75% and 0% along with the mean spectra (thick black line). These hyperspectral data are rather noisy in comparison to spectra collected using a spectrometer [106]. This is primarily a result of how hyperspectral data are collected. In most spectrometers, a single measurement is actually the mean of several independent spectra that were collected over a small area, which greatly reduces the noise in the spectra.

Figure 3.4   Individual spectra for ROIs taken from sub-plots of T75% (left) and T0% (right) along with the mean spectra (thick black line).

Reduction of the noise is essential for calibrating a spectrally based model. However, in an aerial hyperspectral image, each pixel of a hypercube is a single spectrum of a relatively wide area. To reduce the spectral noise in calibrating models, Lawrence et al. [106] suggested using a spatially averaged ROI spectrum and then applying the model on a pixel-by-pixel basis. Lawrence [106] used manual selection of ROI spectra from a close range hyperspectral image to calibrate a PLS spectral model for contaminant detection on poultry carcasses. Manual selection of homogenous ROIs from an aerial hyperspectral image of a field is problematic and might suffer from subjectivity.

#### 3.6.3  Spatial Information to Improve Spectral Classification

Spatial context was suggested as a second step for the refinement of results obtained by spectrallybased techniques. This approach consists of three parts: (1) a pixel-by-pixel spectral classification; (2) definition of a pixel neighborhood (surrounding each pixel); and (3) performance of a local operation so that if there is strong evidence that individual spectra of pixels in a neighborhood are spectrally homogenous they are included in the same cluster. This approach was developed for MS images and extended by Jimenes et al. [107] to airborne hyperspectral sensors. The developed classifier is an unsupervised modification of the supervised extraction and classification of homogenous objects (ECHO). Based on the dataset, the developed algorithm, called UnECHO, automatically estimates the required threshold of homogeneity level of the entire neighborhood without input from the human analyst [107]. When applied to urban and rural areas, the UnECHO successfully uncovered spatial structures and significantly improved spectral classifications (C-means or maximum likelihood [ML]).

Another example of this approach is the Markov random field (MRF) in which spatial characterization is performed by modeling the spatial neighborhood of a pixel as a spatially distributed random process. The MRF attempts to make regularization via the minimization of an energy function using known land covers and their prior probabilities. Similarly to Jimenes et al. [107], Plaza et al. [100] developed an unsupervised version of this methodology. They used a neuro-fuzzy classifier to perform classification in the spectral domain and to compute a first approximation of the posterior probabilities of classes. The output of this step is then fed to the MRF spatial analysis stage, which was performed using a maximum likelihood probabilistic reclassification. The performance of the MRF in classifying urban land cover types was compared with the results of the first stage, that is, a neuro-fuzzy classifier. Similar classification accuracies were achieved mainly because the spatial analysis stage reassigned only border pixels to different classes.

In both classification methods—the UnECHO and the modified MRF—the neighborhoods are determined in advance and do not account for the real size and shape of the objects in the image. This kind of division might not be suitable for the gradual change of biophysical and biochemical properties of crops over the field. Several segmentation approaches were suggested to be performed prior to the hyperspectral image classification. The segmentation techniques such as partitional clustering and hierarchical segmentation [108,109] partition an image into homogeneous regions with different sizes based on a homogeneity criterion [101]. These nondeterministic approaches may be more suitable for mapping homogeneous zones of biophysical and biochemical properties of crops in a field (an illustration of such a technique is provided in the next section). Another alternative that may be adopted for segmentation of homogeneous zones in a field is the geostatistical approach [110,111]. Lark [110] suggested a spatially weighted averaging of the class memberships within a local neighborhood based on the variogram. Although the generation of spatially coherent regions (SCR), developed by Lark [110], was initially applied for limited dimensionality of nonspectral data (multitemporal yield data), it can be adapted to hyperspectral images.

#### 3.6.3.1  Analysis of Hyperspectra Images of an Experimental Plot

To initially investigate the ability to generate SCRs in the spectral domain, we realized and applied the SCR for the HS images of the experimental plot [112]. A fuzzy C-means classification into seven classes was applied (Figure 3.5). Assuming that in the experimental plot the main effect on the reflectance is the nitrogen level, the classes were labeled with a nitrogen level based on a visual inspection and prior knowledge of the N treatments. Despite the noisy result, the fuzzy classification captured differences in N levels: similar N levels in individual pixels were assigned with similar classes. Yet, the classes do not fully match with N treatments. Major difference exists between the two lowest N treatments (0% and 25%) and the other three treatments (50%–100%) while minor or no difference exists between the 50%, 75% and 100% treatments [93]. They also did not differ in their yield (Table 3.2). Additionally, it seems that the western part of the experimental plot is populated with more vital plants than the eastern part.

Figure 3.5   A fuzzy C-means classification of the 210-band HS image of the experimental plot, 7 classes.

Following the spectral classification, the variogram half-range was calculated and used as the neighborhood radius for refining the spectral classification (Figure 3.6). The resulted SCR significantly reduced the speckle effect by uncovering most of the spatial structures of the sub-plots with different N levels. If the affecting factor responsible for the variability in the field is known in advance, the resulting SCRs can be used as management zones for variable-rate application. If not, they can be used both for selecting spectra free of noise for calibrating the spectral model [106] or for implementing a validated model for crop properties estimation.

Figure 3.6   Spatially coherent regions of the fuzzy C-means.

Since the neighborhood is not determined by geometrical shapes, the borders between sub-plots are not crisp but rather fuzzy. This result implies that this type of flexible neighborhood definition is more suitable for the real situation in the field where changes in N levels are gradual and not sharp.

#### 3.6.3.2  Analysis of Hyperspectral and Thermal Images of a Commercial Field

The spectral vegetation indices—NDVI and NDI [79]—were calculated from the aerial HS images (Figure 3.7). The spectral indices and the thermal images (Figure 3.7) exhibit some variability in the commercial field. The eastern side has higher NDVI and NDI values with lower temperature. The SCR methodology was implemented on NDVI and NDI maps and on the raw thermal image of the commercial potato field to examine its ability to delineate homogeneous zones of N levels and water status levels. Similarly to the analysis of the experimental plot hyperspectral image, the image of the commercial field was classified by the fuzzy C-means classification. Since, unlike the manipulated experimental plot, the variability in a commercial field is unknown in advance, the optimized number of classes was calculated based on the change in goodness variance of fit (GVF) with the increasing number of classes. Following the spectral classification, the variogram was calculated to determine the neighborhood radius for refining the spectral classification.

Figure 3.7   Aerial thermal image and NDVI and NDI maps of the commercial potato field. The two latter maps were extracted from an aerial HS image.

Despite the visual resemblance, the number of classes and the variogram of each image were different, leading to different homogeneous zones. The NDVI, NDI and canopy temperature were classified into 2, 3, and 4 classes, respectively.

Figure 3.8 presents the major segments following the omission of very small segments. The SCR algorithm seems to capture the variability in the field but did not necessarily follow the borders between zones perceived by visual inspection of the images. The NDVIs of the two classes were very high, with minor differences. In this range, NDVI saturates and can hardly be associated with potato crop parameters [113]. The NDI image showed relatively high values [89] yet the differences between the lowest and the highest classes may be of consequence for nitrogen level [89]. The thermal image was divided into four classes and, according to the CWSI values, two of them were over-irrigated and two of them had the optimal CWSI value before irrigation. The results show that while there were differences in nitrogen levels the main source of the variability in this field was water status. The main conclusions are as follow: (1) hyperspectral image in the VIS–NIR range and thermal images are complementary; (2) integrating the spatial attribute to the analysis contributes to reveal the variability in BB-PACs and is required for delineating homogeneous zones for variable-rate fertilization and irrigation. However, some of the variability that may be revealed by the image analysis might be of no consequence to the variable-rate application and further expert inspection is required prior to the creation of prescription maps.

Figure 3.8   Homogeneous zones based on the NDVI, NDI and canopy temperature images. Red lines present borders of the homogeneous zones.

#### 3.6.4  Fusion of Spectral and Spatial Information

The previous approaches separate spatial from spectral information, and thus the two types of information are not treated simultaneously. Plaza et al. [100] suggested incorporating spatial context into the SVM spectral classifier. In this method, a pixel entity is redefined simultaneously both in the spectral and spatial domains by applying some feature extraction to its surrounding area, which yields spatial (contextual) features such as the mean or standard deviation per spectral band. These separated entities lead to two different kernel matrices that can be summed and introduce cross-information features in the formulation. When applied for land cover classification in an agricultural area, the contextual SVM showed classification accuracy of 95%. It outperformed a spectral classifier based on Euclidean distance and performed much better than other methods like ECHO that use spectral and spatial information to classify homogeneous objects. Similarly to the UnECHO and the modified MRF, the contextual SVM is based on a predefined neighborhood of N × N windows.

Another approach for fusing spectral and spatial information is a multiscale or hierarchical segmentation. Hierarchical segmentation is based on sequential optimization to produce a hierarchical data-driven decomposition of the picture with no restriction on segment shapes [114]. Beamlet analysis is a framework for multiscale image analysis in which line segments play a role analogous to the role played by points in wavelet analysis [115]. The beamlet-decorated recursive dyadic partitioning (BD-RDP) is one realization of the beamlet analysis. While partitioning with basic RDP is limited to square elements, the BD-RDP allows that some of its squares (optionally) are decorated by a beamlet, that is, can be partitioned not only by squares but also by other geometrical shapes. In comparison to the basic RDP, this additional flexibility allows the BD-RDP to approximate an image more accurately with much fewer segments. The BD-RDP was originally designated to one-dimensional images by implementing two main steps: a spreading phase, where the image is partitioned into its smallest parts according to a quad-tree structure and a folding phase, where the tree is folded up according to a target function. The target function has to serve the idea of minimum variation between the original and reassigned values with a penalty for the number of different segments. The BD-RDP does not require a priori knowledge of the number of segments.

Levi et al. [116] introduced two enhancements of the algorithm, which were illustrated using the hyperspectral of the experimental plot by Cohen et al. [112]. First, it was modified to suit multidimensional images based on the Euclidian distance between vectors, and second, a merging neighbor's phase was added that checks the possibility of merging segments that belong to different dyadic squares using the target function. The multidimensional three-step BD-RDP was applied to the 210-band hyperspectral image of the experimental plot and the segmentation result is shown in Figure 3.9. As it is not a classification, the color of each segment was determined by its average reflectance in an IR band (750 nm) to partially demonstrate the differences in N levels.

Figure 3.9   Multidimensional beamlet-decorated recursive dyadic partitioning of the 210-band hyperspectral image of the experimental plot. The value of each segment is the average of the reflectance in an IR band (750 nm).

The multiscale segmentation successfully uncovered the spatial structures in the image according to differences in N levels. Unlike the SCR classification, the beamlet analysis results in homogeneous segments and needs further analysis to classify the segments according to their N level. For that a fuzzy C-means or a calibrated PLS model can be applied.

#### 3.6.5  Integrating Spectral and Spatial Attributes: A Summary

Hyperspectral images are distinguished from point spectral measurements by their added spatial aspect. Nevertheless, hyperspectral images are usually not treated as images but as lists of spectral measurements with no particular spatial arrangement. Estimation of crop properties using hyperspectral images was based only on spectral information. Precision agriculture necessitates the partition of the field into homogeneous zones. Similar to land use segmentation and classification, delineation of homogeneous zones would benefit from incorporating spectral and spatial information. In this chapter we have introduced different levels of integration of spatial information in estimating crop properties. The ability to integrate and fuse spatial analysis with spectral information was initially demonstrated using an hyperspectral image of an experimental potato plot.

In general, the suggested methods were effective in classifying N levels and uncovered spatial structures that coincided with the blocks of the different treatments, but no quantitative evaluation was done. In other studies, spatial information derived from hyperspectral images was found to be valuable in land use classification. The contribution of spatial analysis for BB-PACs estimation is yet to be studied. For that, research is needed to investigate existing methods and adjust them or develop new methodologies to incorporate spatial and spectral information.

While integrated spectral/spatial algorithms hold great promise for in-season management zone delineation using hyperspectral images, they also introduce computational challenges. Since agrotechnical decisions are made routinely by the farmer once or twice a week, a temporal aspect should also be taken into consideration. With rapid developments, satellites constellations like PlanetLab are already providing daily high–spatial resolution MS images with NIR and RGB bands. Other satellites like the Sentinel2 and the Venμs provide more than 10 bands in the VIS–NIR range in a 2- to 5-day revisit time with reasonable spatial resolution. We may foresee that in the near future frequent hyperspectral images will also be available from satellites. In the meantime, a multiscale approach may address the limited resolutions (temporal and spatial) of the hyperspectral and thermal satellite images. UAV hyperspectral and thermal imaging systems could be used as sampling systems, directed to spots that set wide ranges of the targeted BB-PACs. Thereafter, the associations between the UAV and MS satellite images can be used to extrapolate to wider areas.

In order to fully exploit hyperspectral images, processing methods that can take advantage of their enhanced spectral, spatial and temporal features are required. Parallel processing hardware has necessarily become a requirement to speed up processing performance and to satisfy high computational requirements. As a result, the future potential of hyperspectral image processing methods will also be largely defined by their suitability for being implemented in parallel [100]. The hierarchical segmentation approach that haas been presented in this chapter is suitable for parallel processing hardware and thus the routine of in-season management zone delineation utilizing this approach may be speeded up to meet the timeline requirements of the agrotechnical applications.

#### 3.7  Concluding Remarks

Hyperspectral remote sensing systems enable the collection of several hundred spectral bands in a single acquisition, thus producing detailed spectral and spatial data. The narrow band widths of hyperspectral data allow for better estimation of crop properties compared with the relatively coarse bandwidths acquired with multispectral scanners. While hyperspectral images in the VIS–NIR range provide a tool for estimating and mapping various BB-PACs in the field, they are limited in assessing and mapping crop water status parameters that are essential for irrigation management. For that end, integration of images from the thermal range is required. The literature on the use of hyperspectral and thermal imaging for predicting BB-PACs is massive and cannot be fully summarized in any framework. In this chapter we have reviewed a small portion of these studies and tried to draw guiding principles on how to extend the spectral and spatial properties of hyperspectral and thermal data for management zone delineation toward variable-rate applications.

1. Rather than using sets of single bands or indices, hyperspectral systems that provide contiguous spectra are highly advantageous when analyzed with multivariate techniques.
2. Thermal range and the VIS–NIR–SWIR spectrum should be integrated to decouple the effect of water status and chlorophyll or nitrogen level.
3. Compact hyperspectral sensors and compact uncooled thermal cameras suited for UAVs are already available, opening new possibilities. But to make these systems affordable for agricultural stakeholders, methodologies that fuse hyperspectral or thermal aerial imaging with MS satellite imaging should be developed.
4. Analysis of hyperspectral and thermal images should integrate both their spectral and spatial attributes for the delineation of management zones.

#### References

1
Mahlein, A.K. , Spectral signatures of sugar beet leaves for the detection and differentiation of diseases. Precision Agriculture, 11(4) p. 413–431.
2
Liu, Z.Y. , H.F. Wu , and J.F. Huang , Application of neural networks to discriminate fungal infection levels in rice panicles using hyperspectral reflectance and principal components analysis. Computers and Electronics in Agriculture, 72(2) p. 99–106.
3
Yang, Z. , Differentiating stress induced by greenbugs and Russian wheat aphids in wheat using remote sensing. Computers and Electronics in Agriculture, 2009. 67(1–2) p. 64–70.
4
Reisig, D.D. and L.D. Godfrey , Remotely sensing arthropod and nutrient stressed plants: a case study with nitrogen and cotton aphid (Hemiptera: Aphididae). Environmental Entomology, 2010. 39(4) p. 1255–1263.
5
López-Granados, F. , Weed detection for site-specific weed management: mapping and real-time approaches. Weed Research: 51, p. 1–11.
6
Shapira, U. , Weeds detection by ground-level hyperspectral imaging. in 10th International Conference on Precision Agriculture, 2010.
7
Thorp, K.R. and L.F. Tian , A review on remote sensing of weeds in agriculture. Precision Agriculture, 2004. 5(5) p. 477–508.
8
Fuchs, M. and C.B. Tanner , Infrared thermometry of vegetation1. Agronomy Journal, 1966. 58(6) p. 597–601.
9
Maes, W.H. and K. Steppe , Estimating evapotranspiration and drought stress with ground-based thermal remote sensing in agriculture: A review. Journal of Experimental Botany, 2012. 63(13) p. 4671–4712.
10
Bajcsy, P. and P. Groves , Methodology for hyperspectral band selection. Photogrammetric Engineering and Remote Sensing, 2004. 70(7) p. 793–802.
11
Jain, N. , Use of hyperspectral data to assess the effects of different nitrogen applications on a potato crop. Precision Agriculture, 2007. 8(4–5) p. 225–239.
12
Thenkabail, P.S. , Accuracy assessments of hyperspectral waveband performance for vegetation analysis applications. Remote Sensing of Environment, 2004. 91(3–4) p. 354–376.
13
Thenkabail, P.S. , R.B. Smith , and E. De Pauw , Evaluation of narrowband and broadband vegetation indices for determining optimal hyperspectral wavebands for agricultural crop characterization. Photogrammetric Engineering and Remote Sensing, 2002. 68(6) p. 607–621.
14
Li, F. , Evaluating hyperspectral vegetation indices for estimating nitrogen concentration of winter wheat at different growth stages. Precision Agriculture, 2010. 11(4) p. 335–357.
15
Tucker, C.J. , Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of Environment, 1979. 8 p. 127–150.
16
Cortes, C. and V. Vapnik , Support-vector networks. Machine Learning, 1995. 20(3) p. 273–297.
17
Breiman, L. , Random forests. Machine Learning, 2001. 45(1) p. 5–32.
18
Wold, S. , The collinearity problem in linear-regression—the partial least-squares (PLS) approach to generalized inverses. Siam Journal on Scientific and Statistical Computing, 1984. 5(3) p. 735–743.
19
Baret, F. and G. Guyot , Potentials and limits of vegetation indexes for LAI and APAR assessment. Remote Sensing of Environment, 1991. 35(2–3) p. 161–173.
20
Haboudane, D. , Remote estimation of crop chlorophyll content using spectral indices derived from hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 2008. 46(2) p. 423–437.
21
Idso, S.B. , Normalizing the stress-degree-day parameter for environmental variability. Agricultural Meteorology, 1981. 24(C) p. 45–55.
22
Jackson, R.D. , Canopy temperature as a crop water stress indicator. Water Resources Research, 1981. 171 p. 133–138.
23
Baluja, J. , Assessment of vineyard water status variability by thermal and multispectral imagery using an unmanned aerial vehicle (UAV). Irrigation Science, 2012. 30(6) p. 511–522.
24
Cohen, Y. , Use of aerial thermal imaging to estimate water status of palm trees. Precision Agriculture, 2012. 13(1) p. 123–140.
25
Zarco-Tejada, P.J. , V. González-Dugo , and J.A.J. Berni , Fluorescence, temperature and narrow-band indices acquired from a UAV platform for water stress detection using a micro-hyperspectral imager and a thermal camera. Remote Sensing of Environment, 2012. 117(0) p. 322–337.
26
Rud, R. , Crop water stress index derived from multi-year ground and aerial thermal images as an indicator of potato water status. Precision Agriculture, 2014. 15(3) p. 273–289.
27
Jones, H.G. , Use of infrared thermography for monitoring stomatal closure in the field: Application to grapevine. Journal of Experimental Botany, 2002. 53(378) p. 2249–2260.
28
Meron, M. , Crop water stress mapping for site-specific irrigation by thermal imagery and artificial reference surfaces. Precision Agriculture, 2010. 11(2) p. 148–162.
29
Cohen, Y. , Mapping water status based on aerial thermal imagery: Comparison of methodologies for upscaling from a single leaf to commercial fields. Precision Agriculture, 2017. 18(5) p. 801–822.
30
Jones, H.G. , Use of infrared thermometry for estimation of stomatal conductance as a possible aid to irrigation scheduling. Agricultural and Forest Meteorology, 1999. 95(3) p. 139–149.
31
Berliner, P. , D.M. Oosterhuis , and G.C. Green , Evaluation of the infrared thermometer as a crop stress detector. Agricultural and Forest Meteorology, 1984. 31(3–4) p. 219–230.
32
Taghvaeian, S. , Conventional and simplified canopy temperature indices predict water stress in sunflower. Agricultural Water Management, 2014. 144 p. 69–80.
33
Alchanatis, V. , Evaluation of different approaches for estimating and mapping crop water status in cotton with thermal imaging. Precision Agriculture, 2010. 11(1) p. 27–41.
34
Gonzalez-Dugo, V. , Using high resolution UAV thermal imagery to assess the variability in the water status of five fruit tree species within a commercial orchard. Precision Agriculture, 2013. 14(6) p. 660–678.
35
Klapp, I. , S. Papini , and N. Sochen , Radiometric imaging by double exposure and gain calibration. Applied Optics, 2017. 56(20) p. 5639–5647.
36
Ribeiro-Gomes, K. , Uncooled thermal camera calibration and optimization of the photogrammetry process for UAV applications in agriculture. Sensors (Switzerland), 2017. 17(10).
37
Cohen, Y. , Future approaches to facilitate large-scale adoption of thermal based images as key input in the production of dynamic irrigation management zones. Advances in Animal Biosciences, 2017. 8(2) p. 546–550.
38
Asner, G.P. , Progressive forest canopy water loss during the 2012–2015 California drought. Proceedings of the National Academy of Sciences, 2016. 113(2) p. E249–E255.
39
Aparicio, N. , Spectral vegetation indices as nondestructive tools for determining durum wheat yield. Agronomy Journal, 2000. 92(1) p. 83–91.
40
Haboudane, D. , Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: Modeling and validation in the context of precision agriculture. Remote Sensing of Environment, 2004. 90(3) p. 337–352.
41
Lee, K.S. , Hyperspectral versus multispectral data for estimating leaf area index in four different biomes. Remote Sensing of Environment, 2004. 91(3–4) p. 508–520.
42
Hansen, P.M. and J.K. Schjoerring , Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalized difference vegetation indices and partial least squares regression. Remote Sensing of Environment, 2003. 86(4) p. 542–553.
43
Gitelson, A.A. , Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation. Journal of Plant Physiology, 2004. 161(2) p. 165–173.
44
Danner, M. , Retrieval of biophysical crop variables from multi-angular canopy spectroscopy. Remote Sensing, 2017. 9(7).
45
Thenkabail, P.S. , R.B. Smith , and E. De Pauw , Hyperspectral vegetation indices and their relationships with agricultural crop characteristics. Remote Sensing of Environment, 2000. 71(2) p. 158–182.
46
Darvishzadeh, R. , Leaf Area Index derivation from hyperspectral vegetation indices and the red edge position. International Journal of Remote Sensing, 2009. 30(23) p. 6199–6218.
47
Zarco-Tejada, P.J. , Assessing vineyard condition with hyperspectral indices: Leaf and canopy reflectance simulation in a row-structured discontinuous canopy. Remote Sensing of Environment, 2005. 99(3) p. 271–287.
48
Delegido, J. , Retrieval of chlorophyll content and LAI of crops using hyperspectral techniques: Application to PROBA/CHRIS data. International Journal of Remote Sensing, 2008. 29(24) p. 7107–7127.
49
Wang, F.-m. , J.-f. Huang , and Z.-h. Lou , A comparison of three methods for estimating leaf area index of paddy rice from optimal hyperspectral bands. Precision Agriculture, 2011. 12(3) p. 439–447.
50
Yuan, H. , Retrieving soybean leaf area index from unmanned aerial vehicle hyperspectral remote sensing: Analysis of RF, ANN, and SVM regression models. Remote Sensing, 2017. 9(4).
51
Pimstein, A. , A. Karnieli , and D.J. Bonfil , Wheat and maize monitoring based on ground spectral measurements and multivariate data analysis. Journal of Applied Remote Sensing, 2007. 1 p. 16.
52
Yue, J. , Estimation of winter wheat above-ground biomass using unmanned aerial vehicle-based snapshot hyperspectral sensor and crop height improved models. Remote Sensing, 2017. 9(7).
53
Naor, A. , I. Klein , and I. Doron , Stem water potential and apple size. Journal of the American Society for Horticultural Science, 1995. 120(4) p. 577–582.
54
Hunt, E.R. and B.N. Rock , Detection of changes in leaf water-content using near-infrared and middle-infrared reflectances. Remote Sensing of Environment, 1989. 30(1) p. 43–54.
55
Ceccato, P. , Designing a spectral index to estimate vegetation water content from remote sensing data: Part 1—Theoretical approach. Remote Sensing of Environment, 2002. 82(2–3) p. 188–197.
56
El-Hendawy, S.E. , Spectral assessment of drought tolerance indices and grain yield in advanced spring wheat lines grown under full and limited water irrigation. Agricultural Water Management, 2017. 182 p. 1–12.
57
Kim, D.M. , Highly sensitive image-derived indices of water-stressed plants using hyperspectral imaging in SWIR and histogram analysis. Scientific Reports, 2015. 5.
58
Zhang, F. and G.S. Zhou , Estimation of canopy water content by means of hyperspectral indices based on drought stress gradient experiments of maize in the North Plain China. Remote Sensing, 2015. 7(11) p. 15203–15223.
59
Rodriguez-Perez, J.R. , Evaluation of hyperspectral reflectance indexes to detect grapevine water status in vineyards. American Journal of Enology and Viticulture, 2007. 58(3) p. 302–317.
60
Das, B. , Comparison of different uni- and multi-variate techniques for monitoring leaf water status as an indicator of water-deficit stress in wheat through spectroscopy. Biosystems Engineering, 2017. 160 p. 69–83.
61
Clevers, J. , L. Kooistra , and M.E. Schaepman , Estimating canopy water content using hyperspectral remote sensing data. International Journal of Applied Earth Observation and Geoinformation, 2010. 12(2) p. 119–125.
62
Ordonez, C. , Functional statistical techniques applied to vine leaf water content determination. Mathematical and Computer Modelling, 2010. 52(7–8) p. 1116–1122.
63
Jarvis, P.G. , The interpretation of the variations in leaf water potential and stomatal conductance found in canopies in the field. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 1976. 273(927) p. 593–610.
64
Rapaport, T. , Combining leaf physiology, hyperspectral imaging and partial least squares-regression (PLS-R) for grapevine water status assessment. ISPRS Journal of Photogrammetry and Remote Sensing, 2015. 109 p. 88–97.
65
Pôças, I. , Predicting grapevine water status based on hyperspectral reflectance vegetation indices. Remote Sensing, 2015. 7(12).
66
Möller, M. , Use of thermal and visible imagery for estimating crop water status of irrigated grapevine. Journal of Experimental Botany, 2007. 58(4) p. 827–838.
67
Agam, N. , Spatial distribution of water status in irrigated olive orchards by thermal imaging. Precision Agriculture, 2013. 15(3) p. 346–359.
68
Berni, J.A.J. , Mapping canopy conductance and CWSI in olive orchards using high resolution thermal remote sensing imagery. Remote Sensing of Environment, 2009. 113(11) p. 2380–2388.
69
Gonzalez-Dugo, V. , Almond tree canopy temperature reveals intra-crown variability that is water stress-dependent. Agricultural and Forest Meteorology, 2012. 154-155(0) p. 156–165.
70
O'Shaughnessy, S.A. , Using radiation thermography and thermometry to evaluate crop water stress in soybean and cotton. Agricultural Water Management, 2011. 98(10) p. 1523–1535.
71
Tilling, A.K. , Remote sensing of nitrogen and water stress in wheat. Field Crops Research, 2007. 104(1–3) p. 77–85.
72
O'Shaughnessy, S.A. , S.R. Evett , and P.D. Colaizzi , Dynamic prescription maps for site-specific variable rate irrigation of cotton. Agricultural Water Management, 2015. 159 p. 123–138.
73
Rosenberg, O. , Are thermal images adequate for irrigation management?, in 12th International Conference on Precision Agriculture. 2014 Sacramento, California, USA.
74
Osroosh, Y. , Automatic irrigation scheduling of apple trees using theoretical crop water stress index with an innovative dynamic threshold. Computers and Electronics in Agriculture, 2015. 118 p. 193–203.
75
Steele, D.D. , B.L. Gregor , and J.B. Shae , Irrigation scheduling methods for popcorn in the northern Great Plains. Transactions of the ASAE, 1997. 40(1) p. 149–155.
76
Prenger, J.J. , Plant response-based irrigation control system in a greenhouse: System evaluation. Transactions of the ASAE, 2005. 48(3) p. 1175–1183.
77
Berni, J.A.J. , Thermal and narrowband multispectral remote sensing for vegetation monitoring from an unmanned aerial vehicle. Ieee Transactions on Geoscience and Remote Sensing, 2009. 47(3) p. 722–738.
78
Elsayed, S. , Thermal imaging and passive reflectance sensing to estimate the water status and grain yield of wheat under different irrigation regimes. Agricultural Water Management, 2017. 189 p. 98–110.
79
Datt, B. , Visible/near infrared reflectance and chlorophyll content in Eucalyptus leaves. International Journal of Remote Sensing, 1999. 20(14) p. 2741–2759.
80
Penuelas, J. , Estimation of plant water concentration by the reflectance Water Index WI (R900/R970). International Journal of Remote Sensing, 1997. 18(13) p. 2869–2875.
81
Bannari, A. , Potential of hyperion EO-1 hyperspectral data for wheat crop chlorophyll content estimation. Canadian Journal of Remote Sensing, 2008. 34 p. S139–S157.
82
Kempeneers, P. , Generic wavelet-based hyperspectral classification applied to vegetation stress detection. Ieee Transactions on Geoscience and Remote Sensing, 2005. 43(3) p. 610–614.
83
Blackburn, G.A. and J.G. Ferwerda , Retrieval of chlorophyll concentration from leaf reflectance spectra using wavelet analysis. Remote Sensing of Environment, 2008. 112(4) p. 1614–1632.
84
Blackburn, G.A. , Wavelet decomposition of hyperspectral data: A novel approach to quantifying pigment concentrations in vegetation. International Journal of Remote Sensing, 2007. 28(12) p. 2831–2855.
85
Li, D. , WREP: A wavelet-based technique for extracting the red edge position from reflectance spectra for estimating leaf and canopy chlorophyll contents of cereal crops. Isprs Journal of Photogrammetry and Remote Sensing, 2017. 129 p. 103–117.
86
Liu, B. , Combining spatial and spectral information to estimate chlorophyll contents of crop leaves with a field imaging spectroscopy system. Precision Agriculture, 2017. 18(4) p. 491–506.
87
Herrmann, I. , SWIR-based spectral indices for assessing nitrogen content in potato fields. International Journal of Remote Sensing, 31(19) p. 5127–5143.
88
Tian, Y.C. , Assessing newly developed and published vegetation indices for estimating rice leaf nitrogen concentration with ground- and space-based hyperspectral reflectance. Field Crops Research, 2011. 120(2) p. 299–310.
89
Nigon, T.J. , Hyperspectral aerial imagery for detecting nitrogen stress in two potato cultivars. Computers and Electronics in Agriculture, 2015. 112(Supplement C) p. 36–46.
90
Dash, J. and P.J. Curran , The MERIS terrestrial chlorophyll index. International Journal of Remote Sensing, 2004. 25(23) p. 5403–5413.
91
Moharana, S. and S. Dutta , Spatial variability of chlorophyll and nitrogen content of rice from hyperspectral imagery. Isprs Journal of Photogrammetry and Remote Sensing, 2016. 122 p. 17–29.
92
Herrmann, I. , SWIR-based spectral indices for assessing nitrogen content in potato fields. International Journal of Remote Sensing, 2010. 31(19) p. 5127–5143.
93
Cohen, Y. , Leaf nitrogen estimation in potato based on spectral data and on simulated bands of the VEN mu S satellite. Precision Agriculture, 2010. 11(5) p. 520–537.
94
Samborski, S.M. , N. Tremblay , and E. Fallon , Strategies to make use of plant sensors-based diagnostic information for nitrogen recommendations. Agronomy Journal, 2009. 101(4) p. 800–816.
95
Holland, K.H. and J.S. Schepers , Use of a virtual-reference concept to interpret active crop canopy sensor data. Precision Agriculture, 2013. 14(1) p. 71–85.
96
Faiçal, B.S. , An adaptive approach for UAV-based pesticide spraying in dynamic environments. Computers and Electronics in Agriculture, 2017. 138 p. 210–223.
97
Pantazi, X.E. , Evaluation of hierarchical self-organising maps for weed mapping using UAS multispectral imagery. Computers and Electronics in Agriculture, 2017. 139 p. 224–230.
98
Yang, G. , The DOM generation and precise radiometric calibration of a UAV-mounted miniature snapshot hyperspectral imager. Remote Sensing, 2017. 9(7).
99
Aasen, H. , Generating 3D hyperspectral information with lightweight UAV snapshot cameras for vegetation monitoring: From camera calibration to quality assurance. ISPRS Journal of Photogrammetry and Remote Sensing, 2015. 108 p. 245–259.
100
Plaza, A. , Recent advances in techniques for hyperspectral image processing. Remote Sensing of Environment, 2009. 113(Supplement 1) p. S110–S122.
101
Fauvel, M. , Advances in spectral-spatial classification of hyperspectral images. Proceedings of the IEEE, 2013. 101(3) p. 652–675.
102
Zhang, N. , M. Wang , and N. Wang , Precision agriculture: A worldwide overview. Computers and Electronics in Agriculture, 2002. 36(2) p. 113–132.
103
Haboudane, D. , Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. Remote Sensing of Environment, 2002. 81(2–3) p. 416–426.
104
Miao, Y.X. , Combining chlorophyll meter readings and high spatial resolution remote sensing images for in-season site-specific nitrogen management of corn. Precision Agriculture, 2009. 10(1) p. 45–62.
105
Liu, J.G. , Variability of seasonal CASI image data products and potential application for management zone delineation for precision agriculture. Canadian Journal of Remote Sensing, 2005. 31(5) p. 400–411.
106
Lawrence, K.C. , Partial least squares regression of hyperspectral images for contaminant detection on poultry carcasses. Journal of Near Infrared Spectroscopy, 2006. 14(4) p. 223–230.
107
Jimenez, L.O. , Integration of spatial and spectral information by means of unsupervised extraction and classification for homogenous objects applied to multispectral and hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 2005. 43(4) p. 844–851.
108
Tilton, J.C. , Best merge region-growing segmentation with integrated nonadjacent region object aggregation. IEEE Transactions on Geoscience and Remote Sensing, 2012. 50(11) p. 4454–4467.
109
Tarabalka, Y. , J.A. Benediktsson , and J. Chanussot , Spectral and spatial classification of hyperspectral imagery based on partitional clustering techniques. IEEE Transactions on Geoscience and Remote Sensing, 2009. 47(8) p. 2973–2987.
110
Lark, R.M. , Forming spatially coherent regions by classification of multi-variate data: An example from the analysis of maps of crop yield. International Journal of Geographical Information Science, 1998. 12(1) p. 83–98.
111
Nansen, C. , A.J. Sidumo , and S. Capareda , Variogram analysis of hyperspectral data to characterize the impact of biotic and abiotic stress of maize plants and to estimate biofuel potential. Applied Spectroscopy, 2010. 64(6) p. 627–636.
112
Cohen, S. , Combining spectral and spatial information from aerial hyperspectral images for delineating homogenous management zones. Biosystems Engineering, 2013. 114(4) p. 435–443.
113
Herrmann, I. , LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands. Remote Sensing of Environment, 2011. 115(8) p. 2141–2151.
114
Beaulieu, J.M. and M. Goldberg , Hierarchy in picture segmentation: A stepwise optimization appraoch. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1989. 11(2) p. 150–163.
115
Donoho, D. and X. Huo , Beamlets and multiscale image analysis, in Multiscale and Multiresolution Methods, T.J. Barth , T. Chan , and R. Haimes , Editors. 2002, Springer Lecture Notes in Computational Science and Engineering. p. 149–196.
116
Levi, O. , S. Cohen , and Z. Mharaby , Effective hyper-spectral image segmentation using multi-scale geometric analysis, in IADIS Multi Conference on Computer Science and Information Systems 2010. 2010 Freiburg, Germany.

## Use of cookies on this website

We are using cookies to provide statistics that help us give you the best experience of our site. You can find out more in our Privacy Policy. By continuing to use the site you are agreeing to our use of cookies.