Crop type discrimination using Geo-Stat Endmember Extraction and machine learning algorithms

Abstract

The identification of crop diversity in today’s world is very crucial to ensure adaptation of the crop with changing climate for better productivity as well as food security. Towards this, Hyperspectral Remote Sensing (HRS) is an efficient technique based on imaging spectroscopy that offers the opportunity to discriminate crop types based on morphological as well as physiological features due to availability of contiguous spectral bands. The current work utilized the benefits of Airborne Visible Infrared Imaging spectrometer- New Generation (AVIRIS-NG) data and explored the techniques for classification and identification of crop types. The endmembers were identified using the Geo-Stat Endmember Extraction (GSEE) algorithm for pure pixels identification and to generate the spectral library of the different crop types. Spectral feature comparison was done among AVIRIS-NG, Analytical Spectral Device (ASD)-Spectroradiometer and Continuum Removed (CR) spectra. The best-fit spectra obtained with the Reference ASD-Spectroradiometer and Pure Pixel spectral library were then used for crop discrimination using the ten supervised classifiers namely Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), Support Vector Machine (SVM), Minimum Distance Classifier (MDC), Binary Encoding, deep learning-based Convolution Neural Network (CNN) and different algorithms of Ensemble learning such as Tree Bag, AdaBoost (Adaptive Boosting), Discriminant and RUSBoost (Random Under Sampling). In total, nine crop types were identified, namely, wheat, maize, tobacco, sorghum, linseed, castor, pigeon pea, fennel and chickpea. The performance evaluation of the classifiers was made using various metrics like Overall Accuracy, Kappa Coefficient, Precision, Recall and F1 score. The classifier 2D-CNN was found to be the best with Overall Accuracy, Kappa Coefficient, Precision, Recall and F1 score values of 89.065 %, 0.871,87.565%, 89.541% and 88.678% respectively. The output of this work can be utilized for large scale mapping of crop types at the species level in a short interval of time of a large area with high accuracy.

1. Introduction

The enormous growth in population and availability of limited resources has led the researchers to precisely monitor the health of the crops (Massawe et al., 2016). Thus, among many attributes of land cover, crop type is found to be very effective as far as mapping and assessment of crop diversity in terms of spectral discrimination analysis is considered. Crop diversity would help farmers in growing and developing improved crop varieties with better yield (Renard and Tilman, 2019). The primary objective of crop diversity is sustainable agriculture by ensuring adaptation of the crop to the climate, enabling enough production, so that poverty is reduced, nutritional standards can be achieved and thus safeguarding the food security (Kehlenbeck and Maass, 2004). The role of remote sensing in identifying crop types is very crucial. Multispectral broadband remote sensing data have been used for this purpose for the last three decades. Multispectral data are generally used for estimating crop area, in-season monitoring and large-area vegetation mapping. This data has known issues and limitations as far as their limited spectral bandwidth information is considered (Dhumal et al., 2013). These limitations lead to prominent uncertainties in the classification of crops and their health monitoring.

The hyperspectral remote sensing or imaging spectrometer offers the opportunity to discriminate crop types based on morphological as well as physiological features characteristics through contiguous spectral bands. The lowest bandwidth will provide the finest details of the crops (Lu and Weng, 2007). Past studies by (Thenkabail et al., 2013, Mariotto et al., 2013) have specified the use of various types of sensors such as space-borne Hyperion, HIS (Hyperspectral Instrument) and airborne CASI (Compact Airborne Spectrographic Imager), AVIRIS-NG (Airborne Visible/Infrared Imaging Spectrometer-Next Generation), AISA (Airborne Imaging Spectrometer for Applications) and HyMap for the discrimination of crop types (Singh et al., 2020a). The potential of hyperspectral sensors was investigated for identifying the mustard rot disease (Bhattacharya and Chattopadhyay, 2013). Most of the classification problems are affected by the low signal-to-noise ratio (SNR) and limited data availability with a smaller number of bands (Malhi et al., 2020). A possible solution was proposed through the AVIRIS-NG campaign in 2012, which is an international collaboration between ISRO and NASA.

There are several supervised as well as unsupervised classification techniques were available for the discrimination of crop types. The most widely used supervised classifiers are SVM and SAM (Galvão et al., 2018). The SAM classifier measures the spectral similarity between the reference and targeted spectra (Christian and Krishnayya, 2009) and SVM is a kind of non-parametric classifier that works on the concept of statistical learning (Srivastava et al., 2012). Other techniques such as SID, Minimum distance and Binary encoding were also found to be useful in discriminating various crop types (Zhang et al., 2013, Kuching, 2007). Deep learning classifier CNN was introduced as the most useful classifier because of its superior performance (Lawrence et al., 1997). It can learn features automatically from images through several convolutional blocks and provides perfect classification (Gao et al., 2018, Ramprasath et al., 2018). In a study, performance of most popular algorithms of Ensemble learning such as Tree Bag, AdaBoost, Discriminant and RUSBoost were compared and documented (Dietterich, 2000). Tree Bag classifier provides a flowchart like structure, which generally used to develop a relationship between the input predictor of the dataset and target class (Quinlan, 1987). AdaBoost classifier improves the accuracy using an iterative process (Mounce et al., 2017). The discriminant algorithm is based on the estimation of covariance matrix (Deng et al., 2019). While, RUSBoost algorithm is mainly designed to improve classification performance when training data is imbalanced (Seiffert et al., 2009).

Many experiments were conducted in past using AVIRIS data for the land use/cover mapping (Petropoulos et al., 2012). Several studies have utilized the concept of crop phenology (Chaube et al., 2019), machine learning techniques for crop variables estimation (Gupta et al., 2015, Gupta et al., 2018, Singh et al., 2019), crop type classification (Kumar et al., 2017), land use land cover mapping and classification (Suman and Srivastava, 2016, Pandey et al., 2018, Fragou et al., 2020), chlorophyll prediction (Srivastava et al., 2021), Soil and water quality assessment (Pandey et al.), hyperspectral imagery for urban vegetation cover extraction and soil contamination monitoring (Petropoulos et al., 2015). The current study focused on the crop type discrimination using the following objectives- (i) Establishment and implementation of GSEE on AVIRIS-NG for endmember extraction, (ii) Development of spectral libraries for the crop type classification using AVIRIS-NG, GSEE processed AVIRIS-NG and resampled ASD-Spectroradiometer data (iii) Spectral feature comparison among all developed spectral Libraries using CR (iv) Development and optimization of state of the art machine learning classifiers such as SAM, SID, SVM, MDC, Binary Encoding, CNN and different algorithms of ensemble learning such as Tree Bag, AdaBoost, Discriminant and RUSBoost for crop discrimination (iv) Extensive validation of the outputs using field campaign ground truth datasets.

2. Study Area

In this study, AVIRIS-NG airborne flight data of the Anand district of Gujarat state was used. It is located in the western part of India between 72°00′ to 72°99′N and 22°46′ to 22°56′E at an elevation of 39 meters above mean sea level. Anand district covers 2,951 sq. km and is also known as to be the milk capital of India being the centre of the largest milk dairy. Average annual rainfall of 692 mm was recorded between 1901 to 2014. The highest annual rainfall was recorded as 1518 mm in the year 2005 and the lowest annual rainfall was recorded as 209 mm in 1918 (Priyan, 2015). The soil types in the Anand district are mainly of two types, Goradu (Gravelly) and black (Deckers et al., 2002). The land coverage is predominantly of agriculture type that is suitable for crops such as wheat, maize, tobacco, groundnut and various types of pulses. Tobacco is the most valuable crop produced which is grown on thousands of acres in some taluka of Anand district (General, 2001). AVIRIS-NG airborne flight scene and sampling points for the Anand district of Gujarat can be seen in Figure 1.

Fig. 1. Geographical location of the study area.

3. Datasets

Rigorous field sampling was conducted during the flight overpass of AVIRIS-NG in the year 2016. In situ data points and ASD-Spectroradiometer Reflectance data for various agricultural sites were collected using Garmin eTrex 10 GPS corresponding for different crop types over 26 locations at the campaign site. Nine crops were identified during the sampling for the Anand district study area. The level-2 surface reflectance data from the AVIRIS NG flight in February 2016 was downloaded from VEDAS-SAC (source: https://vedas.sac.gov.in/aviris/) geoportal. AVIRIS-NG measures the data at 5 nm intervals with high SNR >2000 @600nm and >1000 @ 2200 nm. With the availability of a large number of bands (425), the most relevant and accurate information can be retrieved. The accuracy of AVIRIS NG is 95% with a FOV of 34° and IFOV of 1 m rad (https://aviris-ng.jpl.nasa.gov/). The resolution of ground sampling distance (GSD) differs from 4-8 m at a flight elevation of 4-8 Km for a swath of 4-6 Km. VEDAS site provides data in both radiance as well as reflectance format. The presence of various gases and aerosols in the atmosphere makes the data to be susceptible to distortions that need to be handled carefully using atmospheric correction methods. The data from AVIRIS NG was pre-processed from the atmospheric correction point of view for generating L0 and L1 products (radiance to reflectance) (Bhattacharya et al., 2019, Srivastava et al., 2020). The acquired data has a spatial resolution of 4 m with 425 bands ranging between 350-2500 nm at an interval of 5 nm that are sensitive to the band regions in the VNIR (400-100nm) and SWIR (970-2500nm) regions (Singh et al., 2020b, Malhi et al., 2021). ASD FieldSpec®4 full range spectroradiometer instrument was used to capture leaf reflectance data. The spectra were effectively calibrated with a white reference panel. Spectroradiometer has an instantaneous field of view (IFOV) of 10 mm. The spectral range of the spectroradiometer wavelengths varies from 350-2500 nm. The sampling rate of the device is 0.2 seconds per spectrum. Spectral resolution varies from 3 nm in the short-wave infrared and 10 nm in the far-infrared wavelengths, which are resampled to 1 nm. The instrument records spectra of a total of 2151 bands (Anand et al., 2021, Pandey et al., 2021).

Spectral library generation was done using a spectral library builder tool in ENVI 5.2 software. ENVI is an image processing software that has various analysis functions for the study of remote sensing images. This software is useful for image classification, image enhancement and spectral analysis. Spectral feature comparison was done on PRISM software (https://pubs.usgs.gov/of/2011/1155/) specifically for AVIRIS-NG, ASD-Spectroradiometer and identified GSEE pure pixel spectra. The best fit with reference data and spectral signatures of pure pixel features were then used as endmembers for hyperspectral classification. Afterwards, supervised classification algorithms such as SAM, SID, SVM, Minimum distance and Binary encoding were implemented using the classification tool in ENVI 5.2 and machine learning classifiers namely, Ensemble-based Tree Bag, AdaBoost, Discriminant, RUSBoost, through ARTMO (Automated Radiative Transfer Model Operators) software. While the 2D-CNN classification was implemented with the help of Python 3.9.9.

4. Methodology

The current work mainly focuses on crop classification from AVIRIS-NG hyperspectral imagery. The pure pixel is identified firstly on AVIRIS-NG imagery from end members using the GSEE algorithm. Spectral signatures were generated from captured reflectance data of ASD-Spectroradiometer instrument, AVIRIS-NG imagery and also with identified pure pixel values from the hyperspectral image followed by the generation of the spectral library. The identified pure pixels were further used for the classification of the image and the independent set of ground datasets was used for the validation of the classified image. Spectral feature comparison was done among all generated spectral libraries using the continuum removal technique. The best fit spectral library was used for the crop classification. Finally, several classification algorithms, namely SAM, SID, Minimum Distance, Binary Encoding, SVM and various machine learning algorithm based on Ensemble learning such as Tree Bag, AdaBoost, Discriminant, RUSBoost and convolution neural network were applied to the AVIRIS-NG image to classify the crop types existing in the study area. The performance of the classification algorithms is evaluated by using the metrics, Kappa Coefficient, Overall Accuracy, Precision, Recall and F1 score. The flow of the proposed methodology is depicted in Figure 2.

Figure 2. Flow diagram of the methodology.

4.1. Spectral Feature Comparison

The continuum removal technique is also known as the convex hull technique. It is applied over the top of the spectrum which connects spectral maxima and gives background absorption feature values (Huang et al., 2004). Removal of concave shape in spectra is compulsory for the quantification of absorption features. Convex hull transform provides a better comparison among all captured spectra by different types of instruments or techniques. The first and last band values in output continuum removed spectra are equal to 1.0 (Sowmya and Giridhar, 2017). The continuum removal technique is useful to normalize reflectance spectra and provides a comparison of every absorption feature from a common baseline. To connect the local spectral maxima, a convex hull is fitted over the top of the spectrum. The continuum-removed reflectance R’_(λ) can be attained by dividing the value of reflectance R_(λ) and the reflectance level of continuum line R _c(λ) at the consistent wavelength (Rezaei et al., 2008, Mutanga and Skidmore, 2003).

$$R_{(\text{λi})}^{\prime }=\frac{{\text{R}}_{(\text{λi})}}{{\text{R}}_{\text{c}(\text{λi})}}$$ (1)

Continuum removed absorption features such a band depth (BD) and band depth ratio also can be useful parameters to know the depth difference between two spectral values. Band depth can be calculated by a difference of 1 in continuum removed reflectance.

$$B{D}_{(\text{λi})}=1-{\text{R}}_{(\text{λi})}^{\prime }$$ (2)

Where BD is Band depth and R’_(λi) is continuum removed reflectance values. The normalized band depth ratio (BDR) can be estimated by dividing the band depth (D _c) and band center (Filippi et al., 2007, Mutanga and Skidmore, 2003).

$$BD{R}_{(\text{λi})}=\frac{{\text{BD}}_{(\text{λi})}}{{\text{D}}_{\text{c}}}$$ (3)

Continuum removal techniques were performed among all generated spectral libraries. The calculated coefficient of determination (r²) can be used as a measure of the fit between the features. Continuum removal was performed on both spectra using the continuum endpoints (λ1, λ2). PRISM software is useful for the selection of these endpoints. It is established at the points of endpoints of minimum absorption surrounding the absorption feature. Further, the continuum is modelled using a linear function fit. In the continuum-removal tool, a line was drawn over the feature of interest. A GUI (graphical user interface) based continuum removal interface will appear with a plot of the continuum-removed feature.

A comparative assessment was done between two continuum-removed features of the spectrum. PRISM software also measures the spectral similarity between two continuum removed spectral features by using a linear regression algorithm. The r² fit value ranges vary from 0 to 1, nearby values to 1 show the better matches with high fit numbers and the values equal to 1 show perfect agreement. To know the spectral similarity between reference and observed spectra continuum parameters were applied, which are available in the continuum removal interface. The dotted line shows the continuum-removed spectral feature of the reference spectrum and the thick solid line presents the continuum-removed spectral feature of the observed spectrum. In the lower-left corner, the fit parameter (r²) is shown that can be calculated using linear regression. Values of the feature band center parameters and feature depths parameters for reference and observed spectrum were also calculated, which can be seen in the lower part of the figure. First, center wavelength position of the quadratic function for the reference spectral feature (λ_CquadRef) and observed spectral feature (λ_CquadObs) and second feature wavelength position of the channel for the reference spectral feature (λ_CchanRef) and observed spectral feature (λ_CchanObs). Values of Band depths of the features also can be seen in the lower right corner of the plot. First, depth values are based on the quadratic function for scaled reference spectral features (D_SQref) and observed spectral features (D_SQobs). Second, values are based on the channel center of the feature for reference spectral features (D_Cref) and observed spectral features (D_Cobs). (Figure 3)

Figure 3. The spectral feature comparison plot, including fit value and other parameters of wheat.

4.2. Geo-Stat Endmember Extraction algorithm

Endmember extraction is found to be very useful in extracting pure pixels from hyperspectral images. There are two methods of endmember extraction namely, statistical and geometrical methods, which were used in this study. Statistical methods are generally based on inferencing techniques while convex geometry is being used in geometrical methods. Both methods have some advantages and disadvantages. The statistical method is found to be efficient in mixed scenarios and geometrical methods are used in simple scenarios. By using the advantages of both the methods, a GSEE is implemented that extracts pure pixels more efficiently. The GSEE algorithm uses covariance between bands as a statistical feature and convex related points as geometrical features (Shah et al., 2020). The combination extracts pure pixels that existed in the data (Shah and Zaveri, 2019b). Let us assume hyperspectral image as Y of size N*Z. N represents the number of bands in the hyperspectral image and Z represents the number of pixels. The hyperspectral image generally contains many mixed pixels due to mixtures of materials and it is important to identify the pure pixels. The approach to extracting pure pixels from the hyperspectral image is known as Hyperspectral Endmember Extraction (HEE) (Shah and Zaveri, 2019a). Generally, HEE approaches take input as Y and the number of pure materials (P). The number of pure materials/endmembers (P) can be generated in-prior through ground-based measurement or using well-known algorithms (Chang, 2018). Hence, the output of HEE is a pure material matrix (M) of size N*P with P locations. The block diagram for the HEE approach is shown in Figure 4.

Figure 4. Block diagram of Hyperspectral Endmember Extraction.

The main features of GSEE algorithm is given below.

1. Co-variance between bands as a statistical feature

   $$\mathrm{cov}\left(y_i-y_j\right)=E\left[\left(y_i-E\left(y_i\right)\right)\left(y_j-E\left(y_j\right)\right)\right]$$ (4)

   where y_i and y_j represent the i^th and j^th band data. Operator E[.] represent averaging operator. The GSEE algorithm finds cov (y_i, y_j) between all the possible bands without any repetition. Using the statistical features, the algorithm optimizes the data which will be used in the convex geometry. The optimized two band data is stored in S= (y_m, y_n), In this case, y_m and y_n are optimized using co-variance-based features.
2. Convex points as the geometrical features

   $$conv(S)=\left\{\sum _{x=1}^P{w}_x{s}_x\mid \sum _{x=1}^P{w}_x=1,\forall w_x\in R_{+}\right\}$$ (5)

   In the above equation, w_x represents the weights associated with each input s_x. R₊ is a positive real number set and P is the number of pure materials.

*The code for GSEE is available on https://codeocean.com/capsule/4686606/tree

4.3. Automated Radiative Transfer Model Operator (ARTMO)

ARTMO is a very useful software package developed on MATLAB (Rivera et al., 2014). It encompasses both leaf and canopy level RTMs (Radiative Transfer Model) and also provides the user interface for RTM execution and plotting of the RTM’s output for specified optical sensor range between 400-2500 nm. It can be download from https://artmotoolbox.com/. ARTMO provides several toolboxes for the retrieval of biophysical and biochemical parameters and a classification toolbox for the mapping of remote sensing images. Classification toolbox encompasses various types of machine learning classification algorithms such as Artificial Neural Network (ANN), Discriminant analysis, Random Forest, Ensemble learning, Naïve Bayes, Nearest Neighbors, Decision Tree, Pattern Recognition Network and classification Tree. Output values of the kappa coefficient and overall accuracy can be used for the assessment of classification accuracy. Schematic representation of the classification toolbox can be seen in Figure 5.

Figure 5. Layout of Classification toolbox in ARTMO.

4.4. Classification Techniques

The spectral library of different crops was used as an input in the crop type classifiers. Several classification algorithms, namely SAM, SID, Minimum Distance, Binary Encoding, SVM and CNN with different classification algorithms of Ensemble learning classifier such as Tree Bag, AdaBoost, Discriminant and RUSBoost were applied for classification of the crop types in the study area. The SAM classification algorithm uses an n-dimensional angle to match the likeness between the spectra of that pixel and reference class spectra. Smaller angles depict a better resemblance to the reference spectrum (Yuhas et al., 1992, Kuching, 2007). In the present study, SAM classifier was implemented on AVIRIS-NG image using a single value of 0.3 radians as the maximum thresholding value for all the classes. SID, also known as a spectral classification method measures the divergence to match pixel to reference spectra. Minimum divergence offers more accurate results and provides the maximum number of similar pixels. The threshold value is found to be very useful in mapping the pixels (Rajashekararadhya and Shivakumar, 2017). A single value of 0.05 Maximum Divergence threshold was used as the maximum classification threshold value for all classes. Another useful method, the minimum distance classifier was also used in the current study that measures the Euclidean distance value from an unknown pixel by using the mean vectors of each of the endmembers. Every pixel is classified to the nearest class after specifying the threshold value which sometimes resulted in some pixels being unclassified. In the present study, minimum distance error was selected for classification of all pixels Another supervised classifier binary encoding was used that encrypts the data into categories such as zero and one according to the value of the spectral mean. The pixels are classified with the highest number of matched bands subject to the specification of the minimized threshold value (Mazer et al., 1988). Minimum encoding threshold 0.3 was selected to classify all the pixel values. The most commonly used classifier, SVM was also used to evaluate the performance of data that follows statistical learning theory. SVM determine a hyperplane that separates two classes for training data and validation data. It produces good classification results from complex and noisy data (Pal and Mather, 2003, Srivastava et al., 2012). The classification probability threshold value of 0.2 and radial basis function kernel was used to classify the image.

Ensemble methods are based on multiple learning algorithms with improved accuracy performances (Mounce et al., 2017). The bagging algorithm uses different bootstrap samples to train the number of base learners, which is also known as the base learning algorithm. Boosting algorithms have many variants but AdaBoost was introduced as the most famous algorithm. It generates another base learner by calling the base learning algorithm in T times. Final learner is derived by weighted voting of the T base learner. Weighting of the learner is generally determined during the training process. Ensemble Discriminant Analysis classifier is mainly used to classify patterns between two classes. It draws hyperplane and projects the data in such a way to maximize the separation of two classes. The RUSBoost (Random Under Sampling) algorithm one of the famous ensemble-based algorithms is designed to classify when a single class has many more observations. RUSBoost algorithm with X denotes the feature space and D is the weights (Blackard and Dean, 1999).

With the recent advancements in classification algorithms, deep neural networks (DNNs) have been widely used in multiple fields including, medical, environmental, artificial intelligence etc. DNN has proved its efficiency in pixel-based classification in several studies and applied to different land-use types (Zeng et al., 2014, Kong et al., 2018). But its classification performance especially for the hyperspectral data having rich spatial information in terms of spectral bands is yet to be tested in heterogeneous settings (Li et al., 2018, Jia et al., 2016). As a deep neural network inspired by the biological brains, CNN is a multi-layer model that can learn non-linear complex relationships between the input parameters. Especially in the case of hyperspectral data having hundreds of input bands, the CNN architecture can play an important role in feature extraction and classification. Presently, a 2D-CNN based hyperspectral image classification is used to classify different crop types present in a diverse land-use setting. A detailed CNN architecture can be seen in Figure 6.

Figure 6. CNN Architecture.

Before implementing any neural network technique, the data needs to be normalized, as the Hyperspectral data represents layers in terms of bands, the data should be normalized layer by layer. The multiple interconnected layers of the neural network help in the data abstraction with the help of artificial neurons. The artificial neurons receive the input values and pass them through their specific weights estimated through the optimization process. Thereafter, the weighted sum is delivered through the non-linear activation function that further passes it to the next layer. These steps help the network to learn and help in maximizing the performance. In the final layer, the values are passed through the SoftMax function (Zhang et al., 2018) which transforms them into probability as shown in Equation 6.

$${\widehat{p}}_z=\sigma (s(x){)}_z=\frac{\exp \left(s_z(x)\right)}{\sum _{j=1}^z\exp \left(s_j(x)\right)}$$ (6)

where, z is the number of classes, s(x) is representing the weight of each class for the instance of x and σ(s(x))_z is the estimated probability of x belonging to class z.

The feature extraction is done using convolution, the hyperparameters involved in the CNN are dependent upon its size and the number of a kernel that is chosen randomly (Riese et al., 2020). The feature extraction is then followed by downsampling by pooling layers and then the fully connected layer is passed through the ReLU activation function (Ide and Kurita, 2017) (also known as rectifier linear unit) which is mathematically represented as,

$$f(x)=x^{+}=\max (0,x)$$ (7)

where x is the input neuron.

The performance evaluation of the models is estimated by tuning the learning parameters, kernels and their weights via loss function through forwarding propagation and updating these parameter values by optimizing the algorithm either by backpropagation or gradient descent. The Rectified Linear Unit (ReLU) activation function and Adam optimizer were used with 2D-CNN. The kernel size is set to 3x3 and kernel filters used are 16→32→64→128 respectively. The layer description is shown below-

Input(5×5×n) →Conv2D Layer (Filters:16→32→64→128) →Flatten →NN Layer (1000 units, Dense) → Dropout (0.3) →NN Layer (m units, Dense) [Output]

where n is the number of bands in hyperspectral data and m denotes no. of land cover classes The mathematical form of all the classification algorithms is listed in Table 1.

Table 1. Supervised classification algorithm used in this study.

Classifier	Equation	Reference
Spectral Angle mapper (SAM)	where u and v are two n-dimensional spectra and (α) is the angle between two spectra.	(Girouard et al., 2004, Kruse et al., 1993)
Spectral Information Divergence (SID)	${\varnothing }_i={\displaystyle \sum _{j=1}^L{p}_j(\log \left(\frac{p_j}{q_j}\right)+\sum _{j=1}^L{q}_j\log \left(\frac{q_j}{p_j}\right)}$ where Ø is SID error vector for all endmembers and probability vectors p and q are defined for L-dimensional two spectra.	(Nascimento and Dias, 2005)
Minimum Distance	$d_k^2={\left(x-{\mu }_k\right)}^t\left(x-{\mu }_k\right)$ where D is Euclidean distance, X is the vector of image (n bands), μk is the mean of kth class.	(Richards & Richards, 1999)
Binary Encoding	$\overrightarrow{X_{ij}}={\left[X_{ij}(1),X_{ij}(2),\dots \dots X_{ij}(l),\dots X_{ij}(L)\right]}^T$ where, L is the number of spectral channels and (i, j) is the indices belonging to the spatial location of the pixel in the scene.	(Mazer et al., 1988)
Support Vector Machine (SVM)	$K\left(x_i,x_j\right)=\exp \left(-\gamma {\Vert \left(x_i,x_j\right)\Vert }^2\right),\gamma >0$ where k is the number of samples, γ is the width of the kernel function, xi is an n-dimensional vector and xj is the label of each class.	(Srivastava et al., 2012, Petropoulos et al., 2010)
Convolution Neural Network (CNN)	$norm\left(x_{i,j}^k\right)=\frac{x_{i,j}^k-\min \left(x_{i,j}^k\right)}{\max \left(x_{i,j}^k\right)-\min \left(x_{i,j}^k\right)}$ Where x is the pixel location at ith row and jth column of band k.	(Zhang et al., 2018)
Ensemble Learning [Tree] [Bag]	$H(x)=argma{x}_{y\in Y}\sum _{t=1}^{T}1\left(y=h_t(x)\right)$ Where D is the training dataset, T is the number of learning rounds and ht is the base learner.	(Zhou, 2009)
Ensemble Learning [Tree] [AdaBoost]	$H(x)=sign(f(x))=sign\sum _{t=1}^T{\alpha }_t{h}_t(x)$ Where D is the training dataset, T is the number of learning rounds and ht is the base learner.	(Zhou, 2009)
Ensemble Learning [Discriminant]	$g(x)={\omega }^T+{\omega }_0$ where ω is weigth of the linear classifier and ω0 is bias of the classifier	(Onishi and Natsume, 2013)
Ensemble Learning [RUSBoost]	$H(x)=argma{x}_{y\in Y}\sum _{t=1}^T{h}_t(x,y)\log \log \frac{1}{{\alpha }_t}$ Where α is update weight paramter, t is the number of learning rounds and ht is the base learner.	(Mounce et al., 2017)

4.5. Accuracy assessment

The performance of the classification algorithms was evaluated using Kappa Coefficient, Overall Accuracy, Precision, Recall and F1 score. Kappa Coefficient was calculated using equation 8 (Bishop et al., 2007).

$$k=\frac{N\sum _{i=1}^r{X}_{ii}-\sum _{i=1}^r\left(x_{i+}\right)\left(x_{+i}\right)}{N^2-\sum _{i=1}^r\left(x_{i+}\right)\left(x_{+i}\right)}$$ (8)

Precision is a metric that calculates the number of correct true or positive predictions using the terms of confusion matrix results with percentage values for every class. Precision is calculated using equation 9. The recall value is calculated using equation 10 (Flach and Kull, 2015).

$$precision=TP/(TP+FP)$$ (9)

$$TPR/Sensitivity/Recall/ProbabilityofDetection=\frac{TP}{TP+FN}\times 100$$ (10)

where TP is the true positive value of classification, FP is a false positive value and FN is a False-negative value. TPR is also known as sensitivity, probability of detection or recall. FPR is also known as the probability of false alarm. The TPR provides the percentage of correctly predicted instances of crop types other than the required crop type whereas specificity provides the percentage of correctly predicted instances of crop types. The F1 score value is the harmonic mean of the precision and recall and also known as Dice similarity coefficient. F1 score is calculated using equation 11 (Tharwat, 2018).

$$F1=\frac{2\ast precision\ast recall}{precision+recall}$$ (11)

F1 score is generally used when false negative or false positive is also of significance. This is a measure of the test’s accuracy. This is an important metric in terms of it penalizing the extreme values. This founds its applicability in multiclass problems and misbalancing scenarios.

5. Results & Discussions

5.1. Identification of Pure Pixel and Class Separability Analysis

As specified earlier in the last section, the GSEE algorithm is used as an endmember extraction algorithm. This algorithm works by evaluating the covariance between every band to compute the most effective bands. The covariance matrix evaluated through the GSEE algorithm is depicted in Figure 7. As the input hyperspectral data consists of 372 bands, the size of the resulting covariance matrix was 372*372. The value bar is represented on the right side of Figure 3. The yellow colour represents the higher values whereas the blue colour represents the lower values. The most effective bands identified were 51 and 52 as the covariance value between them was found as 0.999. A total of twenty-five endmembers were identified with the help of the GSEE algorithm as depicted in Figure 8 (a). To identify the pure pixels, in situ sampling data was used. The pure pixels were identified using row and column values in GSEE. The pixels with the most matches as per crops were identified as pure pixels. A total of 9 pure pixels were identified for a total number of 9 crops. As every pixel has a unique row and column value, the same is used for generating the spectral library for each of the crops Figure 8 (b).

Figure 7. Covariance matrix for the studied hyperspectral data.

Figure 8.

(a)Pure pixel after GSEE implementation (b) Class Separability plot using pixels values between two bands (10 and 11).

It is very important from the computational cost and complexity perspective that the objects of different classes should be discriminated through some metrics. One such metric is the concept of separability between classes. As depicted in Figure 8 (b), the class separability plot was prepared using pixels values between the two bands to evaluate any mixed pixel effects. These two bands were chosen randomly, for eg here bands 10 and 11 were plotted. Jeffries-Matusita and Transformed Divergence method were used to estimate the separability values of each class. It is clear from the plot that the separable values between the classes are approaching 2, which showed that the classes are quite separated from each other and the training pixels chosen were pure for classification.

5.2. Spectral matching of reference and AVIRIS-NG

Spectral feature comparison was done using continuum removal interface in PRISM. First generated SPECPR files with the same wavelength records has been selected for the continuum removal process using PRISM software. Further, Spectral feature comparison was done among continuum removed reference, Pure Pixel and AVIRIS-NG image extracted spectral libraries. All continuum removed plots are generated and displayed in Figure 9(a)-9(i). The comparative assessment was done for all crops separately. Five crops such as wheat, maize, tobacco, sorghum and fennel were shown more than 0.99 fit value (r²) between CR applied spectral libraries of pure –pixel and observed data. Tobacco crop is considered a dominant crop for the Anand district. Tobacco and Sorghum crops have maximum coverage area for the year 2016 at Anand district, Gujrat. Highest fit values were obtained for the Tobacco and Sorghum crops. Both have observed fit value (r²) 0.998 between CR applied pure –pixel and observed spectral libraries, while with AVIRIS-NG and observed spectra, values of 0.483 and 0.465 were obtained. Fit values obtained for the other crops is shown in Table 2. One of the main problems in crop classification is mixed pixel problem. Mixed pixel has more than one mixture of materials. GSEE algorithm removes these challenges and provides pure pixel of the crops. According to the results of spectral feature comparison of all crops and identified pure pixel using GSEE algorithm-based spectra, it can be seen that the best correlation was observed with ASD-spectroradiometer spectra. Best fit pure pixel spectral library was further used for crop classification. Other parameters such as feature band center and feature depth parameter were also calculated for the reference and observed data along with the center and depth values of spectrum selected for continuum removed spectra (Figure 9).

Figure 9.

(a-i) Spectral feature comparison among the spectral library of Raw AVIRIS-NG imagery, Computed Pure Pixel and ASD- Spectroradiometer data using Continuum Removal Technique

Table 2. Best fit values among CR applied spectral libraries.

Crop	Best fit value (r2) (Reference – Pure Pixel)	Best fit value (r2) (Reference - AVIRIS)
Wheat	0.990	0.558
Maize	0.995	0.582
Tobacco	0.998	0.483
Sorghum	0.998	0.465
Chickpea	0.946	0.480
Fennel	0.992	0.430
Linseed	0.975	0.333
Castor	0.955	0.398
Pigeon pea	0.987	0.410

5.3. Supervised classification

The obtained pure pixels were used for the generation of the spectral library for different crops as shown in Figure 10. It is clear from figure that different crops are having different spectral signatures. The reflectance of maize and wheat crops in NIR regions were observed as 43 % and 39 % respectively. One interesting observation is spectral pattern similarity with different spectral signature values were found in the crops like chickpea, wheat, sorghum and maize. Many classification algorithms were applied to the spectral library of the crops and the classified maps are generated that are displayed in Figure 11(a)-11(j). With SAM classifier, the maximum area was found to be covered by maize crop at 17.851 % while the minimum area was found to be occupied by castor at 3.172 %. The maximum area was found to be covered by linseed crop at 18.084 % and the minimum area was found to be covered by tobacco crop at 2.587 % with SID classifier. With the minimum distance classifier, the maximum area was found to be covered by sorghum crop at 21.804 % and the minimum area was found to be covered by fennel crop at 1.293 %. The maximum area was found to be covered by linseed crop at 20.976 % and the minimum area was found to be covered by fennel crop at 1.317 % with binary encoding classifier. SVM is one of the most popular classifiers in providing results with better accuracy. Here, the classifier offered the maximum area to be covered by sorghum crop at 26.729 % and the minimum area to be covered by fennel crop at 0.292 %. Among all ensemble-based classification algorithms, AdaBoost has provided better results and showed that the maximum area was covered by wheat crop 23.062% and minimum area covered by fennel crop 2.317%. In CNN classification map, the maximum area was covered by the wheat crop (25.03%) and the minimum area was covered by the fennel crop (1.06%).

Figure 10. Spectral profile of different crops from AVIRIS-NG.

Figure 11. (a-j) Classified images using AVIRIS -NG through supervised classifiers.

5.4. Performance Assessment

The performance of the classification algorithms were evaluated based on kappa coefficient, Overall Accuracy (OA), Precision, Recall and F1 score. The CNN and SAM classifiers were found to be better performing than other classifiers as far as overall accuracy is considered. The CNN and SAM showed the OA of 89.065 % and 87.068 %, respectively. The kappa coefficient of these two algorithms were found to be 0.871 and 0.851, respectively. For proper diagnosis, other performance evaluation metrics were also calculated such as precision, recall and F1 score, which yielded values of 87.565%, 89.541% and 88.678%, respectively with CNN (Table 2). From Figure, it is clear that CNN offered the best performance values than the other classifiers.

The results showed that AVIRIS-NG data is found to be very promising in delivering accurate and reliable results (Mäkisara et al., 1997). Several works on agriculture mapping, reflect the true potential of hyperspectral sensors. The author (Kuching, 2007) conducted their research with some classification techniques on hyperspectral data. Data pre-processing was done using CALIGEO software. The classifiers, Decision tree and SAM performed in the best manner with an overall accuracy of 50.67% and 48.83%. Another study was conducted using AVIRIS-NG data in the agricultural site of Maddur, Karnataka and Anand (Gujrat). The data pre-processing techniques, PCA (Principal Component Analysis), MNF (Maximum Noise Fraction) and pixel purity index were utilized for dimensionality reduction. The application of these pre-processing techniques resulted in the improvement of the accuracy where the SAM classifier showed an accuracy of 77.7% and kappa coefficient 0.75. Utilizing AVIRIS data, research on hyperspectral agricultural mapping was conducted with testing of an already developed endmember extraction algorithm based on a support vector machine. The algorithms, SVM-BEE, N-FINDER and SMACC (Sequential Maximum Angle Convex Cone) were used for extracting endmembers. The accuracy of the classifiers was computed and SVM-BEE was found to be performing in the best manner with an accuracy of 62.36 % than other classifiers (Filippi et al., 2009). A thorough comparison was done between previous work and the current work. The comparison is made on the accuracy of the classifiers used. It indicates that the current work utilizing GSEE outperforms the previous work.

6. Conclusions

Hyperspectral Remote Sensing (HRS) is an efficient technique based on imaging spectroscopy and now used in many applications. In this technique, for the identification of material, spectral signature of the materials were first developed and then used for mapping or prediction through some algorithms. The HRS provides an opportunity to reap the benefits offered by advanced remote sensing to develop an improved understanding of algorithms and techniques and deliver a state-of-the-art solution that could benefit society as a whole. The current work utilized the benefits of HRS through AVIRIS-NG data and explored the effective solutions for classification of crop types or identification of crop diversity in an area. Firstly, endmembers were identified using GSEE algorithm followed by identification of pure pixels for crop type discrimination analysis using AVIRIS-NG data. The next step was the development of spectral libraries for the crop type classification and spectral feature comparison was done among all developed spectral libraries using continuum removal algorithm. Finally, validation of the classifiers using field campaign ground truth datasets was conducted in this work. The classifiers, CNN and SAM were found to be better performing than the other classifiers. Based on performance evaluation metrics, Overall Accuracy, Kappa Coefficient, Precision, Recall and F1 score, the CNN was found to be best performing with values of 89.065 %, 0.871,87.565, 89.541 and 88.678 respectively. In comparison with the previous work, it was also observed that the application of the GSEE algorithm has improved the accuracy of the classification of the crop types.

Table 2. Performance assessment of different types of supervised classifiers.

Classifiers	OA (Overall Accuracy)	Precision	Recall	F1 Score
SAM	87.068	85.371	88.44	86.879
SID	76.263	70.815	73.369	72.069
BIN	37.278	30.664	35.282	32.811
MIN	80.776	72.122	84.175	77.683
SVM	84.783	70.887	78.176	74.353
Ensemble (Bagging)	76.264	72.338	73.579	73.158
Ensemble (Boosting)	79.862	75.674	69.91	74.929
Ensemble (Discriminant)	62.356	59.674	60.241	54.702
Ensemble (RUS Boost)	79.459	77.821	72.573	75.471
CNN	89.065	87.565	89.541	88.678