Crop Classiﬁcation for Agricultural Applications in Hyperspectral Remote Sensing Images

: Hyperspectral imaging (HSI), measuring the reﬂectance over visible (VIS), near-infrared (NIR), and shortwave infrared wavelengths (SWIR), has empowered the task of classiﬁcation and can be useful in a variety of application areas like agriculture, even at a minor level. Band selection (BS) refers to the process of selecting the most relevant bands from a hyperspectral image, which is a necessary and important step for classiﬁcation in HSI. Though numerous successful methods are available for selecting informative bands, reﬂectance properties are not taken into account, which is crucial for application-speciﬁc BS. The present paper aims at crop mapping for agriculture, where physical properties of light and biological conditions of plants are considered for BS. Initially, bands were partitioned according to their wavelength boundaries in visible, near-infrared, and shortwave infrared regions. Then, bands were quantized and selected via metrics like entropy, Normalized Difference Vegetation Index (NDVI), and Modiﬁed Normalized Difference Water Index (MNDWI) from each region, respectively. A Convolutional Neural Network was designed with the ﬁner generated sub-cube to map the selective crops. Experiments were conducted on two standard HSI datasets, Indian Pines and Salinas, to classify different types of crops from Corn, Soya, Fallow, and Romaine Lettuce classes. Quantitatively, overall accuracy between 95.97% and 99.35% was achieved for Corn and Soya classes from Indian Pines; between 94.53% and 100% was achieved for Fallow and Romaine Lettuce classes from Salinas. The effectiveness of the proposed band selection with Convolutional Neural Network (CNN) can be seen from the resulted classiﬁcation maps and ablation study.


Introduction
Due to advancements in remote sensing image acquisition mechanisms and the growing availability of rich spectral and spatial information by using a variety of sensors, hyperspectral imaging has gained importance. In particular, Hyperspectral Image (HSI) classification has become a prominent source for practical applications in fields like agriculture, environment, forestry, mineral mapping, etc. [1][2][3][4][5].
The present paper focuses on analyzing and using HSI in the agriculture field. Accurate information about growing crops with different climate conditions and agricultural resources and with different timestamps (before, during, and after cultivation) is extremely important and useful for agricultural development. Traditional methods, like field surveys and other statistical-based analyses, are very time-consuming. Advanced remote sensing

•
There are different successful methods in the literature for HSI classification. However, not all methods are suitable for all the available application areas to perform classification. In the present paper, we designed a band selection model for crop classification based on the physical and biological properties of plants.

•
A new framework for informative band selection is proposed by partitioning the original hyperspectral cube based on the reflectance nature in the visible, near-infrared, and short wave infrared regions of the electromagnetic spectrum. This further uses measures such as entropy, NDVI, and NDWI, respectively, for band quantization. • A two-dimensional convolution network for hyperspectral image classification is designed and implemented for the accurate classification of agriculture crops with the selected bands. Detailed analysis of the results in crop classification is showcased. The following is a breakdown of the paper's structure: materials and methods are presented in Section 2, which consists of a technical description of the proposed method. Dataset description is described in Section 3, followed by Section 4, which includes HSI classification, experimental results, and analysis. Finally, the conclusions are presented in Section 5. ing-based approaches, relative entropy, and information entropy-based approaches [25][26][27][28]. The optimal clustering framework was introduced in [29] for band selection and successfully applied the novel objective function with several constraints. In these methods, bands are clustered initially and then ranked according to different measures to select the representative bands from the image. All these methods work well for selecting the informative bands and hence produce high classification accuracy. However, these methods may not be suitable for application-specific classification problems. For example, in the present context, we consider crop classification as our target. It is important to adopt the band section strategy in the view of agricultural phenology, where biophysical properties of plants are also taken into consideration. This is shown in Figure 1 (Source for Figure 1a: Vegetation analysis: using vegetation indices in ENVI [16]). The details of the measures used for the band selection, mathematical equatio range of the measures, and BS procedure are discussed in the Section 2.
The contributions of the paper are shown below: • There are different successful methods in the literature for HSI classification. Ho ever, not all methods are suitable for all the available application areas to perform classification. In the present paper, we designed a band selection model for crop classification based on the physical and biological properties of plants.

•
A new framework for informative band selection is proposed by partitioning the original hyperspectral cube based on the reflectance nature in the visible, near-infrared, and short wave infrared regions of the electromagnetic spectrum. T further uses measures such as entropy, NDVI, and NDWI, respectively, for band quantization.

•
A two-dimensional convolution network for hyperspectral image classification is designed and implemented for the accurate classification of agriculture crops wi the selected bands. Detailed analysis of the results in crop classification is showcased.
The following is a breakdown of the paper's structure: materials and methods

Materials and Methods
Crop classification using HSI consists of two steps. The first is band selection, and the second is classification. Partition-based band selection is proposed in this paper. Initially, the partition is performed based on properties of the vegetation spectrum, and three partitions are created. These are termed VIS, NIR, and SWIR partitions ( Figure 1). Further bands are selected from these three partitions based on three relevant metrics, entropy, NDVI, and MNDWI, respectively. In the second step to perform classification, the concatenated bands are given as input to the designed CNN model. The CNN model consists of a series of convolution and fully connected layers. Finally, crop classification can be achieved for the selected input data. The proposed architecture for the classification of crops using HSI data is shown in Figure 2. Let H be hyperspectral data with h rows, w columns and d represents numbe spectral bands of H. Each denotes one band in the dataset. According to spectral ref tance properties in various regions of the electromagnetic spectrum, HSI data can separated into n partitions Partition#1, Partition#2 …Partition#n with bands d1, d2,… respectively. Here: In the present context, n value is considered as 3 with Partition#1 denoting band the VIS region, Partition#2 denotes bands in the NIR region, and Partition#3 deno bands in the SWIR region. Three metrics are then chosen to select bands from each pa tion.
Information Entropy (IE) is a criterion to measure spatial information in the H bands. For a particular band X , Information Entropy is defined as: where P(x) is the probability of number of grey level of the histogram of the band x the image. Based on IE, each band in the VIS region is quantified, and top m (< d ) bands selected from Partition#1 based on the threshold limit value of δ1.
Normalized Difference Vegetation Index (NDVI) measures plant health in term greenness density, as shown in Equation (3). This is a widely used vegetation index in remote sensing community. The NDVI ranges from +1 to −1. Dead plants have −1 NDVI value, and healthy plants have values between 0.65 and 1.

NDVI = (NIR RED) (NIR + RED)
Based on NDVI, each band in the NIR region is quantified, and top m (< d ) ba are selected from Partition#2 based on the threshold limit value of δ2.
Modified Normalized Difference Water Index (MNDWI) measures the open wa enhanced identification and is computed using Equation (4). This will suppress no generated by vegetation and soil and, at the same time, improve the open water featu MNDWI ranges from +1 to −1.

MNDWI =
(GREEN SWIR) (GREEN + SWIR) Let H be hyperspectral data with h rows, w columns and d represents number of spectral bands of H. Each denotes one band in the dataset. According to spectral reflectance properties in various regions of the electromagnetic spectrum, HSI data can be separated into n partitions Partition#1, Partition#2 . . . Partition#n with bands d 1 , d 2 , . . . , d n respectively. Here: In the present context, n value is considered as 3 with Partition#1 denoting bands in the VIS region, Partition#2 denotes bands in the NIR region, and Partition#3 denotes bands in the SWIR region. Three metrics are then chosen to select bands from each partition.
Information Entropy (IE) is a criterion to measure spatial information in the HSI bands. For a particular band X j , Information Entropy is defined as: where P(x) is the probability of number of grey level of the histogram of the band x in the image.
Based on IE, each band in the VIS region is quantified, and top m 1 (< d 1 ) bands are selected from Partition#1 based on the threshold limit value of δ1.
Normalized Difference Vegetation Index (NDVI) measures plant health in terms of greenness density, as shown in Equation (3). This is a widely used vegetation index in the remote sensing community. The NDVI ranges from +1 to −1. Dead plants have −1 as NDVI value, and healthy plants have values between 0.65 and 1.
Based on NDVI, each band in the NIR region is quantified, and top m 2 (< d 2 ) bands are selected from Partition#2 based on the threshold limit value of δ2.
Modified Normalized Difference Water Index (MNDWI) measures the open water enhanced identification and is computed using Equation (4). This will suppress noise generated by vegetation and soil and, at the same time, improve the open water features. MNDWI ranges from +1 to −1.
Based on MNDWI, each band in the SWIR region is quantified, and top m 3 (< d 3 ) bands are selected from Partition#3 based on the threshold limit value of δ3.
The proposed band selection algorithm is presented as Algorithm 1 below.

Algorithm 1: Proposed Band Selection Approach for Crop Classification
Input: H ∈ R h×w×d be the Hyperspectral image Data, R: Red band, G: Green band, thresholds: δ1, δ2, and δ3 Output: H BS , finer sub cube with informative and selected bands Step 1: Partition the image H with d bands into three sub cubes based on light properties and biological conditions Step 2: Let the number of bands in each of the sub cubes be d 1 , d 2 and d 3 bands respectively from visible, near Infrared, and shortwave infrared regions Step 3: for i:1 to d 1 , Compute the entropy, E i , of each band using Equation (2) Step 4: Generate finer sub cube H vis with m 1 bands for those bands from d 1 whose E i > δ1 Step 5: for i = 1 to d 2 , Compute the Normalized Difference Vegetation Index, NDVI i , using Equation (3) Step 6: Generate finer sub cube H NIR with m 2 bands for those bands from d 2 whose NDVI i > δ2 Step 7: for i = 1 to d 3 , Compute the Modified Normalized Difference Water Index, MNDWI i , using Equation (4) Step 8: Generate finer sub cube H SWIR with m 3 bands for those bands from d 2 whose MNDwI i > δ3 Step 9: Combine the sub cubes H vis , H NIR and H SWIR as H BS = H vis ∪ H NIR ∪ H SWIR which satisfies Equation (5) There are two modules in the proposed methodology for the band selection task known as Partition and Ranking. The partition module is focused on the Biophysical properties of plants in Visible, NIR, and SWIR regions. These regions are named Partition#1, Partition#2, and Partition#3. In the ranking module, agricultural phenology metrics such as Entropy, NDVI, and MNDWI are computed using Equations (2)-(4) for the three partitions, respectively. Then, the computed values and selected representative bands are ranked using an adaptive threshold for each of the measures. The parameter tunning is shown in the experimental section.
Let H BS ∈ R h×w×m denote hyperspectral data cube after the selection of m spectral bands [30][31][32]. This data can be split into two parts. One is for training and the other for testing. Let imagine χ as a training vector that will be input to the CNN model. The first layer is the convolution layer which follows according to Equation (6). Here ⊗ denotes convolution operator, filter is denoted Ғ and (i, j) denotes the corresponding spatial location.
Here σ (.) denotes activation function. For better convergence, the ReLU activation function is used in the present model as in Equation (7). This function gives output as same as input or zero. σ After a series of convolution layers, the feature vectors are converted into a single Flatten Vector (FV), which will be given as input to Fully Connected (FC) layers. In FC layers, two operations, pre-activation and activation, will be performed at every node. All the FC layers use the ReLU activation function. However, the last layer uses the softmax activation function, as shown in Equation (8).
This function gives output probabilities of each crop, and hence classification is possible.
The proposed method performance was tested with state-of-the-art methods (a brief description of the methods is given in Section 4). The experimental works were carried out using MATLAB R 2018b and Python with Google co-laboratory. The hardware utilized for the work was a personal computer with Intel(R) Core(TM) i5-6500 CPU with 3.20 GHz and 8 GB RAM.

Dataset Description
In the present work, experiments were conducted using two popular AVIRIS sensorbased datasets. These datasets are freely available and can be downloaded from [33]. The first hyperspectral data used is Indian Pines. These data were captured in the agricultural area of northwestern Indiana, USA, on 12 June 1992. The number of pixels is 145*145, collected in the form of 224 bands by covering the electromagnetic spectrum in the range of 400-2500 nm. Agriculture crops like corn and soya are covered in this data as 64% area. The vegetation types of grass and pastures are covered with 25% area [34][35][36].
The second hyperspectral data used is Salinas. These data were captured in the agricultural area of the Salinas Valley region, CA, USA, on 9 October 1998. The number of pixels is 512*217, collected in the form of 224 bands by covering the electromagnetic spectrum in the range of 400-2500 nm. Agriculture crops like vineyard fields, broccoli weeds, celery, fallow, and lettuce crops are covered with 100% area [37][38][39]. The class description and pixel samples information for the two datasets is shown in Table 1 [40][41][42].
The major type of classes that exist in both Indian Pines and Salinas are shown in Figure 3. As per our interest for the present paper, different types of crops that exist in both Indian Pines and Salinas are also shown in Figure 4 [43][44][45][46].
The Indian Pines data belongs to the agricultural area of northwestern Indiana, USA. The area includes major portions of the Indian Creek and Pine Creek watersheds. Indiana is the tenth-largest farming state in the USA. More than 80% of the land in Indiana is dedicated to farms, forests, and woodland. The Salinas data belongs to one of the efficacious agricultural areas located in the central coast region of California, called the Salinas Valley, USA. This site is famous for producing most of the agricultural activities in the county due to its rich soil and plentiful underground water supplies [47].
The nomenclature for the individual classes is set according to the type of land, growing type, and spectral properties. For example, the classes "Brocoli_green_weeds_1" and "Brocoli_green_weeds_2" of Salinas's data belong to the same type of land cover representing broccoli weeds, but they have different spectral properties due to their different conditions. The crop "Lettuce_romanine_4wk" refers to the lettuce crop that grows in the fourth week. The same terminology is used for the classes "Lettuce_romanine_5wk", "Lettuce_romanine_6wk", and "Lettuce_romanine_7wk". From the Indian Pines data, "Corn-notill" represents cultivation without tillage and "Corn-mintill" represents cultivation with minimum tillage. The same terminology is used for the classes "Soybean-notill" and "Soybean-mintill". These two classes, along with "Soybean clean", represent different growing periods of the same soybean crop.
More information about the data and subclass regions in both datasets is presented in Figure 5.
On the whole, it can be concluded that the selected data is more suitable for crop classification applications. Indian Pines data consists of 64% pixel regions as different types of crops, and for Salinas data, crop regions are found to be 56%.   More information about the data and subclass regions in both datasets is presented in Figure 5. On the whole, it can be concluded that the selected data is more suitable for crop classification applications. Indian Pines data consists of 64% pixel regions as different types of crops, and for Salinas data, crop regions are found to be 56%.

Experimental Results and Analysis
This section consists of parameter tuning, complete experimental results, and anal-   More information about the data and subclass regions in both datasets is presen in Figure 5. On the whole, it can be concluded that the selected data is more suitable for c classification applications. Indian Pines data consists of 64% pixel regions as diffe types of crops, and for Salinas data, crop regions are found to be 56%.   More information about the data and subclass regions in both datasets is presented in Figure 5. On the whole, it can be concluded that the selected data is more suitable for crop classification applications. Indian Pines data consists of 64% pixel regions as different types of crops, and for Salinas data, crop regions are found to be 56%.

Experimental Results and Analysis
This section consists of parameter tuning, complete experimental results, and analysis. Crop classification results are evaluated against standard metrics.

Experimental Results and Analysis
This section consists of parameter tuning, complete experimental results, and analysis. Crop classification results are evaluated against standard metrics.

Subsection Parameter Tuning in Band Selection
AVIRIS Sensor data is observed to be spread over three partitions. The first partition is the visible region (VIS) in the wavelength region of 400-700 nm and is equipped with Appl. Sci. 2022, 12, 1670 9 of 20 32 bands (d 1 ). The second partition is the near-infrared region (NIR) in the wavelength region of 700-1000 nm and is equipped with 40 bands (d 2 ). The third partition is the short wave infrared region (SWIR) in the wavelength region of 1000-2500 nm and is equipped with 148 bands (d 3 ).
The threshold value to select bands from the VIS region is set as 4.6 (δ1), and the number of bands selected is 15 (m 1 ) for Indian Pines data, and for Salinas data, these values are set as 4 and 12, respectively. The threshold value to select bands from the NIR region is set as 0.2 (δ2), and the number of bands selected is 15 (m 2 ) for Indian Pines data, and for Salinas data, these values are set as 0.11 and 16, respectively. The threshold value to select bands from the SWIR region is set as 0.62 (δ3), and the number of bands selected is 15 (m 3 ) for Indian Pines data, and for Salinas data, these values are set as 0.99 and 15, respectively.
The red band and green band information is essential to calculate NDVI and MNDWI. These are shown in Figure 6 for both datasets [48,49].
AVIRIS Sensor data is observed to be spread over three partitions. The first partition is the visible region (VIS) in the wavelength region of 400-700 nm and is equipped with 32 bands (d1). The second partition is the near-infrared region (NIR) in the wavelength region of 700-1000 nm and is equipped with 40 bands (d2). The third partition is the short wave infrared region (SWIR) in the wavelength region of 1000-2500 nm and is equipped with 148 bands (d3).
The threshold value to select bands from the VIS region is set as 4.6 (δ1), and the number of bands selected is 15 (m1) for Indian Pines data, and for Salinas data, these values are set as 4 and 12, respectively. The threshold value to select bands from the NIR region is set as 0.2 (δ2), and the number of bands selected is 15 (m2) for Indian Pines data, and for Salinas data, these values are set as 0.11 and 16, respectively. The threshold value to select bands from the SWIR region is set as 0.62 (δ3), and the number of bands selected is 15 (m3) for Indian Pines data, and for Salinas data, these values are set as 0.99 and 15, respectively.
The red band and green band information is essential to calculate NDVI and MNDWI. These are shown in Figure 6 for both datasets [48,49].
The δ1, δ2, and δ3 values were set as shown in Figure 7. Using the number of available bands (224) from the AVIRIS Sensor, we considered 1/5th of the number of the bands. Accordingly, we set m1, m2, and m3 values. From Salinas data, a total of 42 (m) bands were chosen with following band numbers: The graphical representation, which shows the entropy of visible region bands, NDVI of NIR region bands, and MNDWI of SWIR region bands, is in Figure 7 for both datasets.

Parameter Tuning in Classification Using CNN
After the selection of representative bands, it is necessary to perform classification. The detail of the split between training and testing samples of both datasets is presented in Table 2. Parameter tuning for common parameters of the CNN model is presented in Table 3. Batch size 2**3 6 Convolution Layer: filters_1 2**6 The graphical representation, which shows the entropy of visible region bands, NDVI of NIR region bands, and MNDWI of SWIR region bands, is in Figure 7 for both datasets.

Parameter Tuning in Classification Using CNN
After the selection of representative bands, it is necessary to perform classification. The detail of the split between training and testing samples of both datasets is presented in Table 2. Parameter tuning for common parameters of the CNN model is presented in Table 3.  Table 4, and the CNN network model for Salinas data is presented in Table 5.

Results with Accuracy Measures
In this work, we used both quantitative and qualitative result analysis to show the performance of the proposed approach for crop classification. Three standard evaluation quantitative metrics were used [42] and are defined as:  Table 6. The qualitative metric used in this work is the classification map. Figure 7 shows the classification map of the Indian Pines data, and Figure 8 shows the classification map of the Salinas data [49][50][51].
For Indian Pines data, the highest classification result was obtained for two crop classes: (Corn-mintill: class number "3") with an accuracy of 97.12%, and (Soybean-mintill: class number "11") with an accuracy of 99.35%. These two crop class regions are circled in Figure 8d of the classification map result.
For Salinas data, the highest classification result was obtained for two crop classes: (Fal-low_rough_plow: class number "4") with an accuracy of 100%, and (Lettuce_romaine_6wk: class number "13") with an accuracy of 100%. These two crop class regions are pointed out in Figure 9d of the classification map result.

Discussion
This subsection discusses the classification accuracies of classes from the two datasets and also elaborates on the comparison of the proposed framework with the state-of-the-art methods.
For Salinas data, the highest classification result was obtained for two crop classes: (Fallow_rough_plow: class number "4") with an accuracy of 100%, and (Let-tuce_romaine_6wk: class number "13") with an accuracy of 100%. These two crop class regions are pointed out in Figure 9d of the classification map result.

Discussion
This subsection discusses the classification accuracies of classes from the two datasets and also elaborates on the comparison of the proposed framework with the From the Salinas data, it was observed that there are 6 types of crops, Fallow, Fal-low_rough_plow, Fallow_smooth, Lettuce_romaine_4wk, Lettuce_romaine_5wk, and Let-tuce_romaine_6wk, with class numbers 3, 4, 5, 11, 12, and 13. The robust implementation of the classification method resulted in accuracies of 94.53%, 100%, 99.25%, 96.52%, 97.51%, and 100%, respectively. The accuracy range was found to be 94.53-100%.
The first method used for the comparison is the integration of 3-dimensional discrete wavelet transform and Markov random field for hyperspectral image classification called 3DGSVM [52]. In this work, more importance was given to spatial information. 3DDWT is used to extract spatial features. Probabilistic SVM coupled with MRF-based post-processing was used for HSI classification.
The second method used for the comparison is Hyperspectral Image Classification Using Convolutional Neural Networks and Multiple Feature Learning called CNN-MFL [53]. In this work, multiple features were extracted first, followed by several CNN blocks for each set of features. Here, geometric features were incorporated using attribute profiles. This is a novel technique that takes advantage of multiple feature learning and CNN to perform accurate HSI classification.
The third method used for the comparison is Spectral-Spatial Exploration for Hyperspectral Image Classification via the Fusion of Fully Convolutional Networks called SS3FC [54]. This method used spectral, spatial, and semantic information along with Fusion of Fully Convolutional Networks for HSI classification. A novel technique for the balanced splitting of the training/test dataset was introduced to solve the insufficient training samples problem.
The fourth method used for the comparison is unsupervised band selection based on weighted information entropy and 3D discrete cosine transform for hyperspectral image classification called WEDCT-MI [30]. In this work, original HSI data was first converted in discrete cosine transform-based coefficient matrices. The weighted entropy was calculated to quantify each band. Then, top-ranked bands were selected. Finally, SVM was used for classification. Figure 10 shows the crop classification from Indian Pines and Salinas datasets. It can be seen from Figure 10 that the proposed band selection approach is effective in extracting the bands which contain much information about the crops. This is due to the inclusion of the physical properties of light in partitioning the bands and the biological properties of plants in band quantization.

Crop-Wise Analysis
The crop-wise analysis [55] on the two datasets is shown in this subsection.

Corn Crops
The first crop used for the comparison is "Corn-notill," class number 2 from the Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, SS3FC, and WEDCT-MI with higher accuracy of 96.64%. The CNN-MFL method had 94.82% accuracy, which is on par with the proposed method. The second crop used for the comparison is "Corn-mintill", class number 3 from the Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, SS3FC, and WEDCT-MI with higher accuracy of 97.12%. The CNN-MFL method had 96% accuracy, which is on par with the proposed method. The third crop used for the comparison is "Corn", class number 4 from the Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods SS3FC and WEDCT-MI with higher accuracy of 96.67%. The methods 3DGSVM and CNN-MFL had 96.64% and 96% accuracy, respectively, which are on par with the proposed method. Figure 10 shows the crop classification from Indian Pines and Salinas datasets. It can be seen from Figure 10 that the proposed band selection approach is effective in extracting the bands which contain much information about the crops. This is due to the inclusion of the physical properties of light in partitioning the bands and the biological properties of plants in band quantization.

Crop-Wise Analysis
The crop-wise analysis [55] on the two datasets is shown in this subsection.

Corn Crops
The first crop used for the comparison is "Corn-notill," class number 2 from the Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, SS3FC, and WEDCT-MI with higher accuracy of 96.64%. The CNN-MFL method had 94.82% accuracy, which is on par with the proposed method. The second crop used for the comparison is "Corn-mintill", class number 3 from the Indian

Soya Crops
The first crop used for the comparison is "Soybean-no till", class number 10 from the Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, CNN-MFL, SS3FC, and WEDCT-MI with higher accuracy of 97.94%. The second crop used for the comparison is "Soybean-min till", class number 11 from Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, SS3FC, and WEDCT-MI with higher accuracy of 99.35%. The third crop used for the comparison is "Soybean-clean", class number 12 from the Indian Pines data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, SS3FC, and WEDCT-MI with higher accuracy of 95.97%. The method CNN-MFL had 95% accuracy, which is on par with the proposed method.

Fallow Crops
The first crop used for the comparison is "Fallow", class number 3 from the Salinas data. For this crop, the proposed method outperformed the state-of-the-art methods SS3FC and WEDCT-MI with higher accuracy of 94.53%. The methods 3DGSVM and CNN-MFL were slightly more accurate than the proposed method. The second crop used for the comparison "Fallow-rough-plough", class number 4 from the Salinas data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, CNN-MFL, SS3FC, and WEDCT-MI with higher accuracy of 100%. The third crop used for the comparison was "Fallow-smooth", class number 5 from the Salinas data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, CNN-MFL, and SS3FC with higher accuracy of 99.25%. The method WEDCT-MI had 99.03% accuracy, which is on par with the proposed method.

Lettuce Romaine Crops
The first crop used for the comparison is "Lettuce-romaine-4wk", class number 11 from the Salinas data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM and SS3FC with higher accuracy of 96.52%. Methods CNN-MFL and WEDCT-MI were slightly more accurate than the proposed method. The second crop used for the comparison is "Lettuce-romaine-5wk", class number 12 from the Salinas data. For this crop, the proposed method outperformed the state-of-the-art methods 3DGSVM, CNN-MFL, and WEDCT-MI with higher accuracy of 97.51%. Method SS3FC was slightly more accurate than the proposed method. The third crop used for the comparison is "Lettuce-romaine-6wk", class number 13 from the Salinas data. For this crop, the proposed method and CNN-MFL had 100% accuracy and outperformed the remaining state-of-the-art methods.

Ablation Study
We conducted an ablation study to verify the effectiveness of the proposed band selection. There are two modules in the proposed methodology known as Partition and Ranking. We conducted experiments by varying the modules (replace Proposed partition based on VIS, NIR, and SWIR regions into Coarser partition into 3 regions and replace Proposed ranking based on Entropy, NDVI, and NDWI into random selection). The ablation analysis for Indian Pines and Salinas data is shown in Table 7. The accuracy values show that the proposed band selection method has a positive impact on the hyperspectral classification.

Application of Proposed Methodology with Other Satellite Data
In order to test the adaptability and the effectiveness as in [56], we applied the proposed framework on the WHU-Hi-HongHu dataset [57][58][59]. This data consists of a complex agricultural area with a variety of crops. The data were acquired in Honghu City, China. The number of pixels is 940 × 475, with 270 bands acquired from 400 to 1000 nm wavelength. The data were acquired using an unmanned aerial vehicle (UAV)-borne hyperspectral system with high spatial resolution of 0.043 m. Out of 22 class regions, 15 classes were found with class numbers 4, 6,7,8,9,10,11,12,13,14,16,17,18,19,20. Table 8 shows class-wise accuracies after the application of the proposed framework on the WHU-Hi-Hong Hu dataset. We used all the parameters similar to the Indian Pines and Salinas datasets. It can be concluded that the proposed framework was successful in classifying, with an average accuracy of 98.56% for 15 crop classes.

Conclusions
In this paper, an approach for crop classification for HSI images is proposed. Firstly, the bands are selected based on agricultural phenology, where biophysical properties of plants are also taken into consideration. Then, a two-dimensional CNN is trained with the extracted bands. The proposed method is tested and validated on two benchmark datasets. The average accuracy of the crops corn and soya from Indian Pines is 96.81% and 97.75%. For the Salinas dataset, the average accuracy of the crops fallow and lettuce romaine is 97.93% and 98.01%, respectively. The results clearly show the effectiveness of the proposed band selection in selecting the required features necessary for crop classification. In the future, we can incorporate spatial information along with spectral information and extend it to more crops from the datasets.