Sine tangent search algorithm enabled LeNet for cotton crop classification using satellite image

Abstract

Economic growth of country largely depends on crop production quantity and quality. Among various crops, cotton is one of the major crops in India, where 23 percent of cotton gets exported to various other countries. To classify these cotton crops, farmers consume much time, and this remains inaccurate most probably. Hence, to eradicate this issue, cotton crops are classified using deep learning model, named LeNet in this research paper. Novelty of this paper lies in utilization of hybrid optimization algorithm, named proposed sine tangent search algorithm for training LeNet. Initially, hyperspectral image is pre-processed by anisotropic diffusion, and then allowed for further processing. Also, SegNet is deep learning model that is used for segmenting pre-processed image. For perfect and clear details of pre-processed image, feature extraction is carried out, wherein vegetation index and spectral spatial features of image are found accurately. Finally, cotton crop is classified from segmented image and features extracted, using LeNet that is trained by sine tangent search algorithm. Here, sine tangent search algorithm is formed by hybridization of sine cosine algorithm and tangent search algorithm. Then, performance of sine tangent search algorithm enabled LeNet is assessed with evaluation metrics along with Receiver Operating Characteristic (ROC) curve. These metrics showed that sine tangent search algorithm enabled LeNet is highly effective for cotton crop classification with superior values of accuracy of 91.7%, true negative rate of 92%, and true positive rate of 92%.

Keywords

Segmentation network deep learning anisotropic diffusion vegetation index features hyperspectral image

1. Introduction

Demand on food supply system is increasing due to possibility of human population growth. However, from local to global level, impacts of climate change and catastrophic natural catastrophes, like drought and flood are already impeding agricultural production and endangering food security [19, 20]. Therefore, it is essential to gather accurate data on location, scope, health, as well as kind of crops to provide security on food, fight the poverty, as well as manage water resources. Including effective strategies that support necessity of sustainability and climate change adaptation is also more desirable. To reach certain objectives, it is essential to use effective strategies, like advanced machine learning combined with remote sensing data, to ensure that high-quality information is derived about crops [21]. Hyperspectral imaging has become more significant as a result of improvements in image acquisition related remote sensing processes as well as expanding accessibility of rich spatial along spectral information by sensors. Particularly, the classification of hyperspectral images has grown to be a significant source for practical applications in industries like forestry, agriculture, environmental science, mineral mapping, etc. [3]. Rich spectral fingerprints and intricate spatial information of observed scenes can be captured by hyperspectral imaging equipment [4]. For better examination of Earth’s object, hyperspectral image is typically obtained at high number of consecutive narrow spectral wavelengths. Hyperspectral sensor provides substantial capability in data processing for many humanitarian tasks, like precision agriculture for better agricultural techniques, classification among plant classes for better treatment, etc., because the spectral resolution might be in small dimensions [22].

In addition, crop distribution maps from images taken by remote sensing are constant as well as comparable, so that it is particularly useful for big-term assessment of cropping. Spatial resolution of remotely sensed images has a great impact on the detail and accuracy of crop classification [2]. Large scale crop mapping make sense both in practice as well as science. Acquiring accurate crop mapping is a prerequisite for crop assessment and food security monitoring that provide important information for national decision making and for some other social and economic activities, like crop insurance, and land management [5]. Grain classification is most significant applications in agriculture regarding remote sensing. Knowledge of crops is very useful both locally and globally. This information is valuable in planning and implementing agricultural policy and crops management and also ensuring food security. Classification of crops is also a prerequisite for implementation of other remote sensing-based applications on a field scale like biomass and yield estimation, anomaly detection, etc. [23]. Various crops have phenological differences, and crop classification accuracy can be improved by distinguishing different temporal spectral rules of crops [7]. This contributes to industrial structures and agriculture, and creation of a food policy. This has important consequences for management of production in agriculture. It is an effective method ensuring national food security as well as sustainable improvement on socio-economic side. Remote sensing technology is a widely used method for crop data extraction and area estimation because of its fast data collection, wide monitoring area mapping and strong macroscopic characteristics [24].

Traditional methods such as field survey and statistical analyzes are time dependent. Advanced remote sensing technology, including hyperspectral image, provides a suitable solution and can fill gap with solutions such as crop classification [3]. Crop-type maps are mostly created by supervised classification of remote sensed images. To get accurate classification results of crops from images gathered by remote sensing, several aspects should be properly considered, including the selection of images with optimum temporal and spatial resolution, classification methods, as well as collection of training samples. In terms of classification methods, advanced models from machine learning to deep learning have been applied to the classification of one or more temporal remote sensing images [1]. In recent years, more methods are used to classify hyperspectral images. Early stage classification techniques include Support Vector Machine, Random Forest, multiple logistic regression, as well as decision tree that provide promising results on classification. However, these classification techniques only extract low-level features from hyperspectral images. It shows limited ability to handle highly non-linear hyperspectral image data and limited improvement of its classification accuracy [4]. Using deep learning method to extract spectral and temporal features of remote sensing time series data to improve discrimination of different crops is currently the most important crop classification method [26]. Unlike traditional machine learning methods, deep learning has strong feature learning as well as recognition capabilities and has been utilized in advanced remote sensing image assessment methods so far. Deep learning frameworks currently used for remote sensing are mainly Convolutional Neural Networks (CNNs), recurrent neural networks, and attention models and their variants [25].

This work concentrates on cotton crop classification by satellite image by sine tangent search algorithm enabled LeNet. This research follows various stages, like acquisition of image, pre-processing of image, segmentation of image, extraction of features, and classification of cotton crop. Initially, hyperspectral image input is taken from dataset that is forwarded to pre-processing phase. In pre-processing, noises and artifacts from image are eradicated by utilizing anisotropic diffusion method. Pre-processed image is further allowed towards segmentation stage, at which SegNet is used for segmenting images. In segmentation phase, image is segmented or converted into group of regions with pixels for further better processing. At the same time, pre-processed image is forwarded to feature extraction stage, at which features like spatial spectral features as well as vegetation index features are extracted. Finally, hyperspectral cotton crop classification is done from extracted features and segmented image using LeNet. This LeNet is trained by sine tangent search algorithm, formed by combination of both sine cosine algorithm and tangent search algorithm.

Main algorithmic contribution is: Cotton is the major crop grown mainly for seed as well as fiber, wherein its classification is done by LeNet. The LeNet is trained by sine tangent search algorithm. This sine tangent search algorithm is formed by hybridizing both sine cosine algorithm as well as tangent search algorithm. This combined algorithm is very efficient for solving optimization issues in real world with capacity for accurate classification.

This work is explained with following sections such as, Section 2 shows motivation along with literature part related to classification of cotton crops, Section 3 indicates explanation of sine tangent search algorithm enabled LeNet for cotton crop classification, Section 4 produces results with discussion, and at last Section 5 ends the work.

2. Literature review

Kwak and Park [1] introduced self-training with domain adversarial network for crop classification by remote sensing images. Self-training with domain adversarial network was established to be robust in spectral shift problems because of both spatial and temporal inconsistencies. They suffered from wrong classification of minor crop type and needed vast computational cost. Li et al. [2] used temporal sequence object-based CNN for crop classification from fine resolution remote sensing image time series. temporal sequence object-based CNN was effective in avoiding mislabeling pixels falling near field edges, thus, maintaining boundaries of crop parcels. However, this technique was futile in substantially increasing classification accuracy of crops, such as sunflower and corn. Chen et al. [3] developed imputation-bidirectional long short-term memory network model for incomplete time series Sentinel-2A data imputation as well as classification of crop. This technique was successful in improving interpretability of model and in minimizing complexity. But this method failed in improving efficiency by considering spatial correlation. Wu et al. [4] designed dilation-based multi-layer perceptrons with feature fusion network for precise crop classification. This method was successful in resolving spatial variation problem and heterogeneity within same object. However, numerous fusion layers in dilation-based multi-layer perceptrons with feature fusion network resulted in redundant information that reduced performance.

Lei et al. [5] established deep one-class crop classification via positive and unlabeled learning for multi-modal satellite imagery. This method was effective in solving redundant labeling problem in multi-classes and offered high stability with minimal spatial complexity. However, this method failed to carry out real time implementation of approach. Fei et al. [6] used Random Forest with multi-features for cotton classification. Random Forest classifier showed higher stability and accuracy and also required only small amount of remote sensing images. However, it suffered from strong homogeneity as ranking results were more biased towards features with more categories. Hamza et al. [7] designed CNN-transformer hybrid approach for crop classification. This approach successfully minimized wrong classification rate and has fast convergence rate. But, the CNN-transformer hybrid method was futile in substantially minimizing training time. Peng et al. [8] developed deep neural network with conditional random field for crops fine classification in Airborne Hyperspectral Imagery. Although this method achieved high accuracy and reduced mis-classification noises when keeping object boundaries, but this method failed in classifying crops, such as sparse forest, soybean, as well as peach.

The challenges in the existing methods are summarized as follows:

In [4], dilation-based multi-layer perceptrons with feature fusion network was proposed for classifying crops precisely. However, in experimental dataset, if number of classes was not used sufficiently, this will tend to decrement in accuracy of experimental outcomes. Also, the key issue faced by CNN-transformer hybrid method presented in [7] for crop classification was that it failed to consider fusion of output feature sequences of various layers that better utilized relationship among various levels of sequence information. Moreover, the deep neural network with conditional random field model [8] failed in considering other neural network types, like CNN, and incorporation of Unmanned Aerial Vehicle (UAV) as well as aerial images, to meet larger scale needs of fine crops classification. Besides, the deep learning methods required large training data for determining appropriate decision boundaries that limits improvement of their classification accuracy.

3. Sine tangent search algorithm enabled LeNet for hyperspectral cotton crop classification

Accurate cotton maps are very vital to monitor growth of cotton as well as precision management. Accurate statistics for huge areas of cotton are pre-requisites for agricultural production management, and various research types. As multi-temporal monitoring technique, remote sensing method is widely applied in more aspects in present years and is efficiently used for classifying crops. Main intent of this research is developing an effective method to classify cotton crops by proposed STSA. At first, input remote sensing image is acquired from dataset [18], which is subjected to hyperspectral image pre-processing. In the pre-processing step, anisotropic diffusion [27, 9] is used to remove noise from digital images. Later, the noise-free image is applied as input to image segmentation process for segmenting crop area from background, where segmentation is carried out using SegNet [10, 28]. Thereafter, pre-processed image is fed to feature extraction, at which spectral spatial features as well as vegetation index features get extracted. Here, spectral spatial features include Local Binary Pattern (LBP) [11], and Local Gabor XOR pattern (LGXP) [12], wherein vegetation index features include Green Ratio Vegetation Index (GRVI) [29], Wide Dynamic Range Vegetation Index (WDRVI) [14], Normalized Burn Ratio (NBR) [33], Normalized Difference Vegetation Index (NDVI) [29], Enhanced Vegetation Index (EVI) [14], Plant Pigment Ratio (PPR) [32], Green Chlorophyll Index (CI green) [14], Green Leaf Area Index (Green LAI) [31], Difference Vegetation Index (DVI) [30], Ratio Vegetation Index (RVI) [34, 35], and Transformed Vegetation Index (TVI) [13, 35]. Further, cotton crop classification is performed by considering extracted features and segmented image using LeNet [15, 36], which are trained by sine tangent search algorithm. Here, sine tangent search algorithm is devised by combining sine cosine algorithm [16] and tangent search algorithm [17]. Figure 1 indicates diagram of sine tangent search algorithm enabled LeNet for cotton crop classification.

Figure 1.

Block diagram of sine tangent search algorithm enabled LeNet for cotton crop classification.

3.1 Image acquisition

Hyperspectral image of crop is initially acquired from dataset [18], which is further allowed for image pre-processing. Image dataset is indicated as $E$ , which has many images related to agriculture. Dataset with much number of images is denoted by,

$\displaystyle E=\{{E_{1},E_{2},\ldots,E_{a},\ldots E_{b}}\}$ (1)

where $E$ is dataset with number of crop images, $E_{a}$ is image present at $a^{\text{th}}$ position, and $E_{b}$ is considered as image at $b^{\text{th}}$ position, where $b$ is total images in dataset. Here, $E_{a}$ is input image used for pre-processing phase with dimension $(u\times v\times w)$ .

3.2 Image pre-processing by anisotropic diffusion

Next to image acquisition, image pre-processing is done by anisotropic diffusion method [27, 9]. $E_{a}$ is image fed as input for image pre-processing that is carried out for eradicating artifacts and noises from the image. Anisotropic diffusion is an important filtering method that preserves edges in images. This diffusion method smoothens homogeneous areas of images and creates piecewise constant images. Anisotropic diffusion is formulated as below,

$\displaystyle\frac{\partial E_{a}}{\partial r}=\textit{div}(s(\|{\nabla E_{a}}% \|)\nabla E_{a})$ (2)

where $E_{a}$ is original input image, $s$ is stopping function, and $r$ is time. Here, $s\to 1$ when $\|{\nabla E_{a}}\|\to 0$ and $s\to 0$ when $\|{\nabla E_{a}}\|$ is increasing. Moreover, stopping function play important role in anisotropic diffusion behavior that is represented in two ways like,

$\displaystyle s(\|{\nabla E_{a}}\|)=\frac{1}{1+(\|{\nabla E_{a}}\|/A)^{2}}$ (3) $\displaystyle s(\|{\nabla E_{a}}\|)=e^{-(\|{\nabla E_{a}}\|/A)^{2}}$ (4)

where $A$ is threshold value, which acts such that weak edges are smoothed out and strong edges are enhanced. Here, pre-processed image is $P_{a}$ with size $(u\times v\times w)$ that is then taken for image segmenting process.

3.3 Image segmentation by SegNet

Pre-processed image $P_{a}$ is allowed for image segmentation phase using SegNet [10, 28]. Image segmentation is process of converting image into group of regions with pixels that are indicated by labeled image. By classifying image into segments, it is enough to do further processing on important segments rather than on entire image, which reduces time of processing. Moreover, image segmentation enhances clarity of further processing stages.

3.3.1 Architecture of SegNet

SegNet [10, 28] is a deep CNN variant used for segmentation. SegNet has encoder, decoder, as well as pixel-by-pixel classification layer. It was likely created to forecast segmentation masks for unlabeled pictures. Here, the decoder maps the lower-resolution encoder features to the input feature maps to classify. The encoder comprises 13 convolutional layers, which resemble the 13 convolutional layers of the VGG16 network. Also, feature maps are produced by employing the right filters since upsampled maps are sparse. In addition, SegNet has a low parameter count and a large computational capacity, making it simple to train. There are two paths in this model: an upsampling or decoding path and a downsampling or encoding path. In line with this, every encoder has 13-layer decoder. Output from decoder is then passed on to the softmax classifier that creates a probability of class that is independently determined for each pixel in the image.

Thirteen convolutional blocks make up the SegNet encoder path that is then followed using maximum pooling operation with window and a stride 2 for downsampling. Every convolutional block is created in this case using a number of layers of convolution, a Rectified Linear Unit (ReLU), and batch normalization. Similar to the encoder path, the decoder path uses upsampling rather than the encoder path’s max pooling. Here, upsampling uses the outputs from earlier layers and the encoding layers’ maximum pooling indices. Eventually, the softmax classifier is given access to the final decoder output to classify each image pixel. Thus, attained output from SegNet is designated by term $S_{a}$ with dimension $(u\times v)$ . Figure 2 indicates architecture of SegNet with decoders and encoders.

Figure 2.

Model of SegNet.

3.4 Extraction of features

Pre-processed image $P_{a}$ is forwarded for extraction of features, where important features are taken from pre-processed image to gather more information regarding those images. Vegetation index features as well as Spectral spatial features are extracted from pre-processed image. Here, spectral spatial features include Local Binary Pattern (LBP) [11], and Local Gabor XOR pattern (LGXP) [12], as well as vegetation index features include Green Ratio Vegetation Index (GRVI) [29], Wide Dynamic Range Vegetation Index (WDRVI) [14], Normalized Burn Ratio (NBR) [33], Normalized Difference Vegetation Index (NDVI) [29], Enhanced Vegetation Index (EVI) [14], Plant Pigment Ratio (PPR) [32], Green Chlorophyll Index (CI green) [14], Green Leaf Area Index (Green LAI) [31], Difference Vegetation Index (DVI) [30], Ratio Vegetation Index (RVI) [34, 35], and Transformed Vegetation Index (TVI) [13, 35].

3.4.1 Spectral spatial features

These are related to spatial resolution as well as spectral resolution. Spatial resolution focuses on measuring quality of image, and spectral resolution characterizes images based on accurate wavelengths. LBP and LGXP are extracted spectral spatial features, which are indicated below,

Local Binary Pattern (LBP): For hyperspectral data, LBP features [11] get extracted for pre-processed image. Basic idea in LBP extraction is comparison of circular neighborhood along with radius $G$ . Central pixels grey value is identified and utilized as central threshold for comparison with nearby pixel values. Thus, from comparison outcomes, various equivalent pattern classes are chosen for obtaining LBP code that represent central pixels grayscale. LBP is thus formulated as,

$\displaystyle\textit{LBP}_{B,C}^{\textit{rin }2}=\left\{{\begin{array}[]{ll}% \sum\nolimits_{B=0}^{B-!}D(c_{b}-c_{e}),&\text{if }F(\textit{LBP}_{B,C})<0\\ B+1,&\text{otherwise}\\ \end{array}}\right.$ (5)

Here,

$\displaystyle F(\textit{LBP}_{B,C})=|{D(c_{b-1}-c_{f})-D(c_{0}-c_{e})}|+\sum% \nolimits_{b=1}^{b-1}{|{D(c_{b}-c_{e})-D(c_{b-1}-c_{e})}|}$ (6)

where $F(\textit{LBP}_{B,C})$ is uniformity function representing conversion numbers created during comparison along neighborhood, $D$ is sign function, $c_{e}$ is central pixels grayscale in every circular neighborhood, $c_{b}(B=0\ldots B-1)$ is circular neighborhoods set with radius $G(G>0)$ . Local spatial image characteristics are obtained from uniform LBP histogram of image. Thus, LBP feature is indicated by term $SF_{1}$ .

Local Gabor XOR pattern (LGXP): This feature is derived from Gabor Phase Response (GPR), which is gained by finding central pixel along neighboring pixels. In LGXP [12], phase response of Gabor filter is quantized to several regions by Local Experience Platform (LXP) operator. This implements logical exclusively-OR (XOR) enabled operator for calculating code based on phase range counts. This LGXP in decimal and binary form is formulated as,

$\displaystyle L_{\textit{GXP},\alpha,\beta,\chi}^{\textit{bin}}(c_{d})=[{L_{% \textit{GXP},\alpha,\beta,\chi}^{F},L_{\textit{GXP},\alpha,\beta,\chi}^{F-1},% \ldots,L_{\textit{GXP},\alpha,\beta,\chi}^{1}}]$ (7) $\displaystyle L_{\textit{GXP},\alpha,\beta,\chi}^{\textit{dec}}(c_{d})=\left[{% \sum\limits_{g=1}^{F}{2^{g-1}}\times L_{\textit{GXP},\alpha,\beta,\chi}^{g}}\right]$ (8)

where $c_{d}$ is position of central pixel in every GPR with orientation $\chi$ and scale $\alpha$ , $F$ is neighborhood numbers, and $L_{\textit{GXP},\alpha,\beta,\chi}^{g}(g=1,2,\ldots,F)$ indicates pattern calculated among $c_{d}$ and it’s neighbor $c_{g}$ . Operator $R_{\textit{LXP}}\{.\}$ calculates XOR outcome for center pixel $c_{d}$ and one of neighborhoods $c_{g}$ , which is indicated as,

$\displaystyle R_{\textit{LXP}}\{{c_{d},c_{g}}\}=\left\{{\begin{array}[]{ll}0;&% c_{d}=c_{g}\\ 1;&\text{otherwise}\\ \end{array}}\right.$ (9)

Hence, LGXP is more resistant to phase alterations because of quantification processing and is indicated by $SF_{2}$ .

Thus, spectral spatial features are indicated in vector form as, $SF=\{{SF_{1},SF_{2}}\}$ .

3.4.2 Vegetative index

Remote sensing is efficient to track phonological alterations like leaf green up as well as autumn coloring from regional to global scale. Some transformation or combination of spectral bands, which accentuates spectral properties of green plants, so that plants look different from some other features of image is involved in vegetative index. Vegetative index is detection of regional scale variation and helped in classification of crop types in an effective way. Vegetative index features extracted are GRVI [29], WDRVI [14], NBR [33], NDVI [29], EVI [14], PPR [32], CI green [14], Green LAI [31], DVI [30], RVI [34, 35], and TVI [13, 35].

Green Ratio Vegetation Index (GRVI): This index is highly sensitive to rate of photosynthesis in forest canopies, as red and green reflectance are very much influenced by alterations in leaf pigments. This index [29] is indicated as below,

$\displaystyle VF_{1}=\frac{\lambda_{g}-\lambda_{r}}{\lambda_{g}+\lambda_{r}}$ (10)

where $\lambda_{g}$ resembles visible green reflectance, $\lambda_{r}$ indicates visible red reflectance, as well as $VF_{1}$ is GRVI index.

Wide Dynamic Range Vegetation Index (WDRVI): WDRVI [14] is index that measures three times greater accuracy of infrared light levels, which is close to red light when associated with pixels. WDRVI enables more robust characterization of crop phenological and physiological characteristics, which is formulated by,

$\displaystyle VF_{2}=\frac{\Im.\lambda_{nr}-\lambda_{r}}{\Im.\lambda_{nr}+% \lambda_{r}}$ (11)

where $VF_{2}$ is WDRVI, weighting coefficient is indicated as $\Im=0.05-1$ , usually, 0.1–0.2 and $\lambda_{nr}$ is near infrared reflectance.

Normalized Burn Ratio (NBR): This index identifies burned areas and provides measure of burn severity. NBR combines both mid infrared and near infrared wavelengths [33]. Here, high NBR represents healthy vegetation and low NBR represents bare ground, which is formulated as below,

$\displaystyle VF_{3}=\frac{\eta_{nr}-\eta_{mr}}{\eta_{nr}+\eta_{mr}}$ (12)

where $VF_{3}$ is NBR index, $\eta_{nr}$ is wavelength of near infrared, and $\eta_{mr}$ is wavelength of mid infrared. This NBR value ranges from $-$ 1 to 1.

Normalized Difference Vegetation Index (NDVI): This is normalized ratio of near infrared reflectance and red reflectance which is used in phonological studies that include timing detection of autumn coloring and green leaf up [29]. This index is formulated as,

$\displaystyle VF_{4}=\frac{\lambda_{nr}-\lambda_{r}}{\lambda_{nr}+\lambda_{r}}$ (13)

where $VF_{4}$ is NDVI index.

Enhanced Vegetation Index (EVI): This index [14] is an optimized vegetation index that quantifies vegetation greenness. This is responsive to the canopy variations in structure, which includes LAI, plant physiognomy, type and architecture of canopy. This enhance vegetation signal with improvised sensitivity in higher regions of biomass as well as enhanced vegetation monitoring via decoupling of canopy background signal that is indicated by,

$\displaystyle VF_{5}=H\times\frac{\lambda_{nr}-\lambda_{r}}{1+\lambda_{nr}+I_{% 1}\lambda_{r}-I_{2}\lambda_{b}}$ (14)

where $VF_{5}$ is EVI index, $H$ is gain factor $=$ 2.5, $I_{1}$ and $I_{2}$ indicates coefficients of aerosol resistance term, where $I_{1}=$ 6, $I_{2}=$ 7.5.

Plant Pigment Ratio (PPR): Plant pigments have fabulous importance in biosphere, and are essential in photosynthesis, and other physiological process, which is necessary for plant growth. Owing to significance of pigments for photosynthesis, pigment content variation provides information on plants physiological condition. This is indicated by reflectance of vegetation in Digital Multispectral Video system (DMSV) band 1 and band 2 [32] that is designated by term $VF_{6}$ . PPR is formulated as below,

$\displaystyle VF_{6}=\frac{V_{550}-V_{450}}{V_{550}+V_{450}}$ (15)

where $VF_{6}$ is PPR index, $V_{550}$ is vegetation in DMSV band 2 reflectance, which is centred at 550 nm, as well as $V_{450}$ is vegetation in DMSV band 1 reflectance, which is centred at 450 nm.

Green Chlorophyll Index (CI green): This index [14] is the ratio of chlorophyll’s reflectance in near infrared band over green band reflectance. This helps to find chlorophyll content of leaves, and is sensitive to little variations in chlorophyll content that is given as below,

$\displaystyle VF_{7}=\frac{\lambda_{nr}}{\lambda_{g}}-1$ (16)

where $VF_{7}$ is CI green index.

Green Leaf Area Index (Green LAI): This index is termed as total surface area of green leaves per unit of ground area [31]. Green LAI quantifies amount of foliage in plant canopy and it forms main driving feature for production of nutrient use and production of water, which is indicated as $VF_{8}$ . Green LAI is indicated in formula as,

$\displaystyle VF_{8}=\frac{A_{\textit{leaf}}}{A_{\textit{ground}}}$ (17)

where $VF_{8}$ is Green LAI index, $A_{\textit{leaf}}$ is green leaf area, and $A_{\textit{ground}}$ is area of ground.

Difference Vegetation Index (DVI): DVI index [30] is slope based group, which is obtained by subtracting red reflectance from near infrared reflectance. This is simple than NDVI, however prone to errors on measurement in red and near infrared bands, because it is not normalized by their addition. Range of DVI is infinite and is given as,

$\displaystyle VF_{9}=\lambda_{nr}-\lambda_{r}$ (18)

where $VF_{9}$ is DVI index.

Ratio Vegetation Index (RVI): RVI indicates stress level of crops due to its high correlation with dry biomass, leaf area, as well as chlorophyll content [34, 35]. This index is near infrared band reflectance divided by red band reflectance. Here, larger RVI indicates healthy vegetation and is indicated as,

$\displaystyle VF_{10}=\frac{\lambda_{nr}}{\lambda_{r}}$ (19)

where $VF_{10}$ is RVI index.

Transformed Vegetation Index (TVI): TVI [13, 35] is modified NDVI by adding constant of 0.5 to all its values and taking square root of results. This constant 0.5 is introduced in avoiding operating with the negative values of NDVI. Square root calculation is used to correct NDVI values, which approximate Poisson distribution and establish normal distribution. Also, there is no technical variation among TVI and NDVI in relation with active vegetation detection or image output, which is formulated as,

$\displaystyle VF_{11}=(VF_{4}+0.5)^{1/2}$ (20)

where $VF_{11}$ is TVI index, and $VF_{4}$ is NDVI.

Thus, vegetative index features are indicated in vector form as, $VF=\{{VF_{1},VF_{2},\ldots,VF_{11}}\}$ .

Overall features acquired from pre-processed image is,

$\displaystyle F_{a}=\{{SF,VF}\}$ (21)

3.5 Hyperspectral image cotton crop classification by LeNet

Extracted features $F_{a}$ and segmented image $S_{a}$ are allowed for hyperspectral image cotton crop classification. This classification is carried out by LeNet that is trained by sine tangent search algorithm.

3.5.1 Architecture of LeNet

CNN structure LeNet [15, 36] encourages deep learning development. This neural network is a typical example influenced by the visual system of humans. Given that the topology is CNN, spatial linkages are leveraged to cut down on the number of parameters. LeNet treats smaller image proportions as input to the low layer of its hierarchical structure. Basic structural components include convolutional layer, fully connected layer, as well as pooling layer.

Convolutional layer

This is significant part in CNN having autonomous filter group. Each filter is convoluted autonomously with feature maps. If convoluted image is of dimension of ${\rm Z}_{ts}\times\Upsilon_{tq}$ with $x\times y$ as filter size, output feature map is attained with size of ${\rm P}_{x}\times{\rm P}_{y}$ , which is formulated as,

$\displaystyle{\rm P}_{x}=\frac{{\rm Z}_{ts}-x+2{\rm E}_{x}}{{\rm K}_{x}}+1$ (22) $\displaystyle{\rm P}_{y}=\frac{{\rm Y}_{tq}-y+2{\rm E}_{y}}{{\rm K}_{y}}+1$ (23)

whereas, padding of zeros in height as well as width is ${\rm E}_{y}$ , ${\rm E}_{x}$ , stride in directions of horizontal as well as vertical is ${\rm K}_{x}$ , ${\rm K}_{y}$ .

All feature map output is found by input maps convolution with inclusion of bias, linear filters, and applies non-linear function as,

$\displaystyle XY_{st}^{yz}={\rm T}_{t}\left({\sum\limits_{xw\in TU_{st}}{XY_{% xw}^{zy-1}*Se_{\textit{xw.st}}^{zy}+\textit{bias}_{st}^{zy}}}\right)$ (24)

where layer number is $z y$ , convolutional kernel is $Se_{\textit{xw.st}}$ , bias is $\textit{bias}_{st}$ , and activation function is ${\rm T}_{t}(.)$ .

The ability to learn and complete difficult tasks depends on this function. These functions are utilised here as input but are not linear transformations. This makes it clear whether a neuron is receiving the information it needs or not.

ReLU is a nonlinear function that causes errors to spread backward. It uses ReLU to activate many of its neurons. The main benefit of ReLU is that it uses sparse, simple, and efficient computation to activate a small number of neurons. Equation of ReLU is given by,

$\displaystyle{\rm T}_{t}({XY})=\left\{{\begin{array}[]{ll}XY,&XY\geqslant 0\\ 0,&XY<0\\ \end{array}}\right.$ (25)

Initial convolutional layer takes more low level features, such as lines, edges, and corners. Stacking more convolution layers tends network for learning global features.

Pooling layer

This decreases spatial dimension of the depiction as well as parameter counts used in network computation, which in turn controls amount of fitting. The pooling function is used to function each layer of input. Additionally, spatial pooling is of numerous categories, such as min, max, total, and average. While using spatial pooling, if the feature map is max pooling, the high element is taken up. Average term is taken into account while pooling averages. Maximum pooling produces beneficial results for dual reasoning, like reducing computation in the top layers by eliminating non-maximal values and providing translational variance generation. This is the best way to reduce the size of intermediary representations because it increases location robustness.

Fully connected layers

This shows that each neuron in current layer is connected with neuron in next layer. In this case, bias offset is used after matrix multiplication to activate. The final connected layer’s neurons are the same as the classes that are required to be discovered. Convolutional and subsampling layers’ features are excellent for classification, but feature integration is superior. As a result, the fully connected layer includes characteristics from the subsequent convolutional and subsampling layers. The fully linked layer, which is utilized for multilevel classification, uses softmax activation, a fully global logistic activation function. Thus, outcome from LeNet is $\textit{LeN}_{a}$ and its architecture is given in Fig. 3.

Figure 3.

Architecture of LeNet.

3.5.2 Training of LeNet by sine tangent search algorithm

LeNet that is used for cotton crop classification is trained by sine tangent search algorithm, which is formed by combining sine cosine algorithm [16] and tangent search algorithm [17]. Tangent search algorithm is population related algorithm for resolving optimization issues. Tangent search algorithm utilizes tangent function as mathematical model for bringing out a better solution. This offers great capacity in exploring search space and also periodicity of function maintains good balance among intensification and exploration. Moreover, tangent search algorithm is very supportive in resolving all optimization problems. Sine cosine algorithm is population enabled optimization algorithm, which used many random and adaptive variables for facilitating convergence and divergence of search agents. Here, multimodal, unimodal, and composite functions were used for testing exploitation, exploration, convergence, and local optima avoidance. Tangent search algorithm helps to solve problems in different fields and solve various optimization problems. Combination of both tangent search algorithm and sine cosine algorithm to form sine tangent search algorithm is highly efficient in solving real time issues with unknown as well as constrained search spaces. Also, they are applied well on solving classification problems with high accuracy and effectiveness.

Tangent position encoding

In the search space $\Omega$ , fine solution is gained where learning parameter of LeNet is given as $\hbar$ indicated by $\Omega=1\times\hbar$ .

Finding fitness

Fitness is found to attain maximum solution for solving optimization problems and this function utilizes results obtained from LeNet and includes targeted output.

$\displaystyle Y=\frac{1}{b}\sum\limits_{a=1}^{b}{[{\textit{Tar}_{a}^{\ast}-% \textit{LeN}_{a}}]}^{2}$ (26)

where $b$ is full count of samples taken for processing, $\textit{LeN}_{a}$ is generated LeNet output, and $\textit{Tar}_{a}^{\ast}$ is targeted output.

Step 1: Initialization

Initialization is started by generation of population within the boundaries of solution space in a random way. Initial solution is distributed over the search space uniformly, which is formulated as below,

$\displaystyle J_{0}=B_{l}+(B_{u}-B_{l})\ast\textit{rand}(K)$ (27)

where $B_{l}$ is lower bound of problem, $B_{u}$ is upper bound of problem, rand is random value that ranges [0, 1], as well as dimension of problem is indicated as $K$ .

Step 2: Determining fitness measure

Fitness is determined based on targeted output and output from LeNet as per Eq. (26).

Step 3: Intensification search

Tangent search algorithm makes initially random local walk, which is guided as below and variables of attained solution is replaced by values of variable in present optimum solution. Variables are changed in proportion of 50 percent having dimension less or equal to 4 variables, and proportion of 20 percent having dimension higher than 4 and is denoted as,

$\displaystyle J_{h}^{i+1}=J_{h}^{i}+\textit{step}\ast\tan(\delta)\ast(J_{h}^{i% }-\textit{optL}_{h}^{i})$ (28) $\displaystyle J_{h}^{i+1}=J_{h}^{i}+\textit{step}\ast\tan(\delta)\ast J_{h}^{i% }-\textit{step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i}$ (29) $\displaystyle J_{h}^{i+1}=J_{h}^{i}[1+\textit{step}\ast\tan(\delta)]-\textit{% step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i}$ (30)

where $J_{h}^{i+1}$ is location of $h^{\text{th}}$ solution at $(i+1)$ iteration, step indicates step size function, which decreases as iteration $i$ decreases, $\delta$ is angle, as well asoptL is best current solution for guiding search process towards best solution.

Position update equation of sine cosine algorithm is given as,

$\displaystyle J_{h}^{i+1}=J_{h}^{i}+m_{1}\ast\cos(m_{2})\ast|{m_{3}Z_{h}^{i}-J% _{h}^{i}}|$ (31)

where $m_{1}$ is next position region, $m_{2}$ is outwards the destination, $m_{3}$ is random number that ranges [0, 1], $Z_{h}^{i}$ is destination point position in $h^{\text{th}}$ dimension, and $|\,|$ indicates absolute value. Assume, $Z_{h}^{i}>J_{h}^{i}$ , then Eq. (31) becomes,

$\displaystyle J_{h}^{i+1}=J_{h}^{i}+m_{1}\ast\cos(m_{2})\ast(m_{3}Z_{h}^{i}-J_% {h}^{i})$ (32) $\displaystyle J_{h}^{i+1}=J_{h}^{i}+m_{1}\ast\cos(m_{2})m_{3}Z_{h}^{i}-m_{1}% \ast\cos(m_{2})J_{h}^{i}$ (33) $\displaystyle J_{h}^{i+1}=J_{h}^{i}[1-m_{1}\ast\cos(m_{2})]+m_{1}m_{3}\cos(m_{% 2})Z_{h}^{i}$ (34) $\displaystyle J_{h}^{i}[1-m_{1}\ast\cos(m_{2})]=J_{h}^{i+1}-m_{1}m_{3}\cos(m_{% 2})Z_{h}^{i}$ (35) $\displaystyle J_{h}^{i}=\frac{J_{h}^{i+1}-m_{1}m_{3}\cos(m_{2})Z_{h}^{i}}{1-m_% {1}\ast\cos(m_{2})}$ (36)

Substituting Eq. (36) in Eq. (30) enhances hybridization of sine cosine algorithm with tangent search algorithm,

$\displaystyle J_{h}^{i+1}=\frac{J_{h}^{i+1}-m_{1}m_{3}\cos(m_{2})Z_{h}^{i}}{1-% m_{1}\ast\cos(m_{2})}[1+\textit{step}\ast\tan(\delta)]-\textit{step}\ast\tan(% \delta)\ast\textit{optL}_{h}^{i}$ (37) $\displaystyle\frac{J_{h}^{i+1}(1-m_{1}\ast\cos(m_{2}))-J_{h}^{i+1}}{1-m_{1}% \ast\cos(m_{2})}=\frac{\begin{array}[]{c}-m_{1}m_{3}\cos(m_{2})Z_{h}^{i}[1+% \textit{step}\ast\tan(\delta)]\\ -(\textit{step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i})(1-m_{1}\ast\cos(m_{2% }))\end{array}}{1-m_{1}\ast\cos(m_{2})}$ (38) $\displaystyle J_{h}^{i+1}(1-m_{1}\ast\cos(m_{2}))-J_{h}^{i+1}=-m_{1}m_{3}\cos(% m_{2})Z_{h}^{i}[1+\textit{step}\ast\tan(\delta)]$ (39) $\displaystyle{}-(\textit{step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i})(1-m_{% 1}\ast\cos(m_{2}))$ $\displaystyle J_{h}^{i+1}[1-m_{1}\ast\cos(m_{2})-1]=-m_{1}m_{3}\cos(m_{2})Z_{h% }^{i}[1+\textit{step}\ast\tan(\delta)]$ (40) $\displaystyle{}-(\textit{step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i})(1-m_{% 1}\ast\cos(m_{2}))$ $\displaystyle J_{h}^{i+1}[-m_{1}\ast\cos(m_{2})]=-m_{1}m_{3}\cos(m_{2})Z_{h}^{% i}[1+\textit{step}\ast\tan(\delta)]$ (41) $\displaystyle{}-(\textit{step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i})(1-m_{% 1}\ast\cos(m_{2}))$ $\displaystyle J_{h}^{i+1}=\frac{1}{m_{1}\ast\cos(m_{2})}[(m_{1}m_{3}\cos(m_{2}% )Z_{h}^{i})[1+\textit{step}\ast\tan(\delta)]$ (42) $\displaystyle{}+(\textit{step}\ast\tan(\delta)\ast\textit{optL}_{h}^{i})(1-m_{% 1}\ast\cos(m_{2}))]$

The above Eq. (3.5) is update equation of sine tangent search algorithm for training LeNet.

Where $J_{h}^{i+1}$ is position of $h^{\text{th}}$ solution at $(i+1)$ iteration, $m_{1}$ is next position region, $m_{2}$ is outwards the destination, and $m_{3}$ is random number ranging [0, 1].

Step 4: Exploration search

Tangent search algorithm has great capacity towards exploration and uses product of variable step size as well as tangent flight for making global random walk. Tangent function supports exploration efficiently in search space. Merging of local and global random walk is given below indicating exploration search as,

$\displaystyle J_{h}^{i+1}=J_{h}^{i}+\textit{step}\ast\tan(\delta)$ (43)

where while $\delta$ is near to $lh/2$ , then tangent value is made huge as well as attained solution is far from current solution, and while $\delta$ is near to 0, then tangent value is made smaller as well as attained solution is near to current solution. Moreover, $\delta=lh/2$ diverges tangent search algorithm, and in this exploration search $\delta$ ranges $[0,lh/3]$ . The above intensification search as well as exploration search depends on parameter Mswitch.

Step 5: Procedure of escape local minima

Tangent search algorithm uses method dealing with specific procedure for escaping from local minima stagnation problem. This includes Mesc, and $N$ is selected and followed in any one of the below equations,

$\displaystyle J=J+N\ast(\textit{optL}-\textit{rand}\ast(\textit{optL}-J))$ (44) $\displaystyle J=J+\tan(\delta)\ast(B_{u}-B_{l})$ (45)

where $N$ is agent search, optL is best current solution in guiding search process towards best solution, $\delta$ is angle, and $J$ is current solution.

Step 6: Termination

Sine tangent search algorithm is thus terminated when optimal maximal solution is gained to train LeNet. This solution is gained based on fitness, as per Eq. (26) and pseudocode resembling sine tangent search algorithm is depicted in Algorithm 1.

Algorithm 1: Pseudo code of sine tangent search algorithm
1 Input: Population as well as maximum iteration
2 Output: Maximum solution $J_{h}^{i+1}$
3 Start sine tangent search algorithm
4 Start population in a random manner by Eq. (27)
5 Find fitness by Eq. (26)
6 If $\textit{rand}<\textit{Mswitch}$
7 Apply intensification search using Eq. (28)
8 Hybridization of sine cosine algorithm along tangent search algorithm;
9 Find update equation of sine tangent search algorithm by Eq. (3.5)
10 Else
11 Apply exploration search using Eq. (43)
12 End
13 If $\textit{rand}<\textit{Mesc}$
14 Select agent search $(N)$
15 Apply escape local minima by Eqs (44) and (45)
16 End
17 Re-evaluate fitness function using Eq. (26).
18 End sine tangent search algorithm

Figure 4 shows the flowchart of the sine tangent search algorithm enabled LeNet.

Figure 4.

Flowchart of sine tangent search algorithm enabled LeNet. SCA – sine cosine algorithm, TSA – tangent search algorithm, STSA – sine tangent search algorithm.

4. Results and discussion

Sine tangent search algorithm enabled LeNet is hybrid optimization algorithm that is used for classification problem. Sine tangent search algorithm enabled LeNet is compared with many methods and its performance is assessed and elaborated in this section. The implementation of sine tangent search algorithm enabled LeNet is carried out in python tool with various performance metrics utilizing Bhuvan $|$ NRSC Open EO Dataset [18]. The experimental parameters of all methods are provided in Table 1.

Table 1
Parameter details

	Parameter	Value
STDAN
	Epoch	25
	Batch size	32
	Learning rate	0.001
TS-OCNN
	Epoch	20
	Batch size	64
	Learning rate	0.001
Im-BiLSTM
	Epoch	20
	Batch size	64
	Learning rate	0.001
DMLPFFN
	Epoch	25
	Batch size	64
	Learning rate	0.001
STSA_LeNet
	Epoch	30
	Batch size	100
	Learning rate	0.001
	Loss function	Categorical_crossentropy

STDAN – Self-Training with Domain Adversarial Network. TS-OCNN – Temporal Sequence Object-based CNN. Im-BiLSTM – Imputation-Bidirectional Long Short-Term Memory. DMLPFFN – Dilation-based Multi-Layer Perceptrons with Feature Fusion Network. STSA_LeNet – Sine Tangent Search Algorithm enabled LeNet.

4.1 Experimental results

Figure 5 indicates experimental results of image regarding classification of cotton crop by sine tangent search algorithm enabled LeNet. Figure 5a indicates input image of cotton crop. Figure 5b represents pre-processed image. Figure 5c is representation of segmented image of cotton crop by SegNet.

Figure 5.

Experimental outcomes of image.

4.2 Dataset description

Bhuvan $|$ National Remote Sensing Centre (NRSC) Open EO Dataset [18] is utilized in this paper. NRSC is primary centres of Indian Space Research Organization (ISRO). NRSC is mandate for ground station establishment to receive satellite data, data product generation, development of techniques for remote sensing applications that include geospatial services for good governance, dissemination to users, as well as capacity building for professionals, and disaster management support. This dataset has many satellite products, terrestrial sciences products, ocean sciences products, atmospheric and climate sciences products, as well as cryospheric products with data availability, resolution and frequency. Moreover, land use with scale, availability, and features are also indicated in this dataset. Here, one time satellite data or spatial details of Indian Remote Sensing (IRS) sensors, and satellite derived products pertaining to many domains and themes are available. The hyperspectral images from this dataset are received from HySI and IMS1 that operates in 64 spectral bands in 400 nm to 900 nm with 500 meter spatial resolution and swath of 128 kms.

4.3 Evaluation metrics

Sine tangent search algorithm enabled LeNet is assessed with many metrics like accuracy, True Negative Rate (TNR), True Positive Rate (TPR), and Receiver Operating Characteristic (ROC) curve, which are explained below.

4.3.1 Accuracy

This metric provide exactness of classification for presenting best output. This metric indicates fraction of terms indicating correct classification like true positive and true negative to overall classification of all positives and negatives. This is indicated by below formula,

$\displaystyle\Omega=\frac{{\rm T}r(+)+{\rm T}r(-)}{{\rm T}r(+)+{\rm T}r(-)+Fa(% +)+Fa(-)}$ (46)

where $\Omega$ is accuracy, $Fa(+)$ is false positive, false negative is $Fa(-)$ , ${\rm T}r(+)$ is true positive, and ${\rm T}r(-)$ is true negative.

4.3.2 True Positive Rate (TPR)

TPR indicates true classification rate like true positive from overall true predictions such as true positive and false negative. This predicts accurate positive class from overall true classes. This proportion of accurate predictions in positive class prediction is formulated as,

$\displaystyle T_{p}=\frac{{\rm T}r(+)}{{\rm T}r(+)+Fa(-)}$ (47)

where $T_{p}$ is term indicating TPR.

4.3.3 True Negative Rate (TNR)

TNR indicates fraction of true negative from overall false predictions like true negative and false positive. This predicts accurate negative cases from overall false classes. This proportion of accurate negative class prediction is formulated as,

$\displaystyle T_{n}=\frac{{\rm T}r(-)}{{\rm T}r(-)+Fa(+)}$ (48)

where $T_{n}$ is term indicating TNR.

4.3.4 Receiver Operating Characteristic (ROC) curve

There, This curve is plotted among two parameters like False Positive Rate (FPR) and True Positive Rate (TPR). This Receiver Operating Characteristic (ROC) curve is used for classifying problems at different threshold settings. This is probability curve indicating models capability among various distinguishable classes.

Figure 6.

Comparative analysis by altering training set. DMLPFFN – Dilation-based Multi-Layer Perceptrons with Feature Fusion Network. Im-BiLSTM – Imputation-Bidirectional Long Short-Term Memory. STDAN – Self-Training with Domain Adversarial Network. STSA_LeNet – Sine Tangent Search Algorithm enabled LeNet. TS-OCNN – Temporal Sequence Object-based CNN.

Figure 7.

Comparative analysis by changing k-value. DMLPFFN – Dilation-based Multi-Layer Perceptrons with Feature Fusion Network. Im-BiLSTM – Imputation-Bidirectional Long Short-Term Memory. STDAN – Self-Training with Domain Adversarial Network. STSA_LeNet – Sine Tangent Search Algorithm enabled LeNet. TS-OCNN – Temporal Sequence Object-based CNN.

4.4 Comparative methods

Various methods used for comparative analysis with sine tangent search algorithm enabled LeNet (STSA_LeNet) include Self-Training with Domain Adversarial Network (STDAN) [1], Temporal Sequence Object-based CNN (TS-OCNN) [2], Imputation-Bidirectional Long Short-Term Memory model (Im-BiLSTM) [3], and Dilation-based Multi-Layer Perceptrons with Feature Fusion Network (DMLPFFN) [4]. Comparative assessment of STSA_LeNet is done in two ways like varying training set and varying k-value in terms of three performance metrics.

Self-Training with Domain Adversarial Network (STDAN): It combined the adversarial training for alleviating the issues in spectral discrepancy. It generates the new training data automatically in the target domain with the help of ground truth or existing thematic map details.

Temporal Sequence Object-based CNN (TS-OCNN): It did the crop classification from fine spatial resolution images. The boundary information are maintained clearly by using this model.

Imputation-Bidirectional Long Short-Term Memory model (Im-BiLSTM): This model was used to perform the crop classification and missing data imputation with less errors and uncertainty.

Dilation-based Multi-Layer Perceptrons with Feature Fusion Network (DMLPFFN): The dilation-based multilayer perceptron model was developed for keeping the relative spatial pixels locations unchanged. The multi-branch residual blocks used the multi-level feature details for efficient crop classification.

4.4.1 Comparative assessment by changing training set

Figure 6 is comparative analysis of STSA_LeNet by changing training set. Figure 6a indicates accuracy-based comparative analysis by changing training set. While training data is 50%, accuracy is 0.685, 0.750, 0.780, 0.816, and 0.850 for STDAN, TS-OCNN, Im-BiLSTM, DMLPFFN, and STSA_LeNet, with performance enhancement of 19.32%, 11.78%, 8.24%, and 3.96%. Figure 6b indicates TPR enabled comparative analysis by varying training set. While training percentage $=$ 60, then TPR is high for STSA_LeNet with value of 0.877, wherein other methods exhibit low values of TPR of 0.750, 0.726, 0.796, and 0.850, showing improvement in performance of 14.46%, 17.19%, 9.21%, and 3.08%. Figure 6c indicates TNR enabled comparative assessment by varying training set. When 70 is taken as training percentage, then TNR value is higher of 0.879 for STSA_LeNet, wherein other techniques show low values of 0.746, 0.800, 0.815, and 0.826 for other models. Here, performance improvement is 15.10%, 8.95%, 7.18%, and 5.99%.

Figure 8.

Receiver Operating Characteristic (ROC) analysis. DMLPFFN – Dilation-based Multi-Layer Perceptrons with Feature Fusion Network. Im-BiLSTM – Imputation-Bidirectional Long Short-Term Memory. STDAN – Self-Training with Domain Adversarial Network. STSA_LeNet – Sine Tangent Search Algorithm enabled LeNet. TS-OCNN – Temporal Sequence Object-based CNN.

Figure 9.

Algorithmic analysis. Jaya – Jaya optimization algorithm, CSO – Competitive Swam Optimizer, SCA – Sine Cosine Algorithm, STSA – Sine Tangent Search Algorithm, TSA – Tangent Search Algorithm.

4.4.2 Comparative assessment by changing k-value

Figure 7 shows comparative analysis of STSA_LeNet by altering k-value. Figure 7a indicates accuracy based comparative analysis by altering k-value. While k-fold is 7, accuracy is 0.727, 0.777, 0.806, 0.837, and 0.897 for STDAN, TS-OCNN, Im-BiLSTM, DMLPFFN, and STSA_LeNet. This shows performance improvement of 18.96%, 13.39%, 10.06%, and 6.69%. Figure 7b indicates TPR enabled comparative assessment by varying k-value. While k-value is 8, then TPR is 0.906 for STSA_LeNet, wherein other methods show values of 0.806, 0.836, 0.855, and 0.855, with performance enhancement of 11.04%, 7.64%, 5.54%, and 5.54%. Figure 7c indicates TNR enabled comparative assessment by varying k-value. While k-fold $=$ 9, then TNR is 0.920 for STSA_LeNet, whereas other models show low values of TNR of 0.800, 0.817, 0.835, and 0.876, with improvement in performance of 13.05%, 11.23%, 9.18%, as well as 4.81%.

4.5 Analysis with Receiver Operating Characteristic (ROC) curve

Figure 8 indicates ROC curve analysis in terms of FPR and TPR. Here, when FPR is 0.4, then TPR values are 0.333, 0.663, 0.776, 0.884, and 1 for STDAN, TS-OCNN, Im-BiLSTM, DMLPFFN, and STSA_LeNet. Moreover, when FPR is 1, then TPR is 1 for STSA_LeNet, wherein other models show lower TPR of 0.535, 0.739, 0.83, as well as 0.915.

4.6 Algorithmic assessment

STSA_LeNet is analyzed with various algorithms like, Jaya optimization algorithm (Jaya) [37], Competitive Swam Optimizer (CSO) [38], Sine Cosine Algorithm (SCA) [16], and Tangent Search Algorithm (TSA) [17]. Figure 9 indicates assessment on algorithm of STSA_LeNet by altering iteration. Figure 9a indicates accuracy enabled algorithmic assessment. Here, while iteration is 50, then accuracy is 0.726, 0.756, 0.807, 0.826, and 0.866 for Jaya, CSO, SCA, TSA, and STSA_LeNet. Here, performance improvement is 16.20%, 12.74%, 6.86%, and 4.66%. Figure 9b indicates TPR enabled algorithmic analysis. For iteration of 90, TPR is higher of 0.927 for STSA_LeNet, wherein other models show low TPR values of 0.770, 0.797, 0.827, and 0.915, with performance improvement of 16.91%, 14.03%, 10.75%, and 2.0%. Figure 9c indicates algorithmic analysis on basis of TNR. When iteration is 60, then TNR is 0.874 for STSA_LeNet, whereas TNR values are 0.727, 0.703, 0.777, and 0.860 for other techniques. This shows performance enhancement of 16.79%, 19.50%, 11.12%, and 1.58%.

Table 2
Comparative discussion

Classification	Metrics/methods	STDAN	TS-OCNN	Im-BiLSTM	DMLPFFN	Proposed STSA_LeNet
Varying training set (%)	TPR	75.7%	80.7%	85.6%	89.5%	91.0%
	Accuracy	77.7%	80.4%	81.7%	89.0%	91.3%
	TNR	80.0%	77.7%	83.5%	88.9%	92.0%
Varying k-value	TPR	84.5%	81.4%	88.6%	85.5%	92.0%
	Accuracy	75.7%	78.6%	81.7%	86.7%	91.7%
	TNR	80.0%	81.7%	83.5%	87.6%	92.0%

DMLPFFN – Dilation-based Multi-Layer Perceptrons with Feature Fusion Network. Im-BiLSTM – Imputation-Bidirectional Long Short-Term Memory. STDAN – Self-Training with Domain Adversarial Network. STSA_LeNet – Sine Tangent Search Algorithm enabled LeNet. TS-OCNN – Temporal Sequence Object-based CNN.

Figure 10.

Confusion matrix.

4.7 Comparative discussion

Table 2 indicates comparative discussion of STSA_LeNet with many other methods such as STDAN, TS-OCNN, Im-BiLSTM, and DMLPFFN. From Table 2, accuracy is more of 91.7%, while compared with other techniques and this is due to utilization of anisotropic diffusion for image pre-processing that eradicated noises and artifacts in image. Similarly, 92% is attained TNR value for STSA_LeNet and this high value is because of SegNet utilized for segmentation. Also, TPR is high of 92% for STSA_LeNet while compared with other methods and this is due to LeNet used for cotton crop classification that is trained by hybrid optimization algorithm, sine tangent search algorithm. These higher accuracy, TPR, and TNR values are attained for STSA_LeNet by altering k-value.

4.8 Analysis on confusion matrix

Figure 10 depicts the confusion matrix of the sine tangent search algorithm enabled LeNet. This analysis is done based on the false and true values of the predicted label and true label. It summarizes the efficiency of the proposed classification approach.

5. Conclusion

Cotton is cultivated in most of the world, which is one of the economically important agricultural crops. It is significant to classify cotton crop to reduce issues caused during export and import scenarios. In this paper, novelty lies in utilization of hybrid optimization algorithm, named sine tangent search algorithm for training LeNet. Here, hyperspectral image is taken from dataset and then pre-processed by anisotropic diffusion, at which unwanted noises in image is removed. Further, segmentation is carried out using SegNet for identifying accurate segments and simultaneously, pre-processed image is fed for extraction of feature, where vegetation index and spectral spatial features of image are accurately extracted. Finally, cotton crop is classified from image that is segmented as well as from extracted features using LeNet, trained by sine tangent search algorithm. Here, sine tangent search algorithm is formed by uniting sine cosine algorithm and tangent search algorithm. Furthermore, sine tangent search algorithm enabled LeNet is analyzed with evaluation metrics like accuracy that shows superior value of 91.7%, True Negative Rate (TNR) with superior value of 92%, and True Positive Rate (TPR) with superior value of 92%. Further, this model can be analyzed with some other hybrid optimization algorithms to train deep learning model that accurately classify cotton crops. Moreover, this model can be applied to other classifications of crop like rice, maize, etc.

Footnotes

Author’s Bios

	Devyani Jadhav is Assistant Professor and Associate Dean, IQAC in Sanjivani COE Kopargaon who is expertise is problem solving using Programming Language and Operating System, Data Structure, DBMS and machine learning etc. She is highly passionate, self-motivated, dedicated, hard-working person. She is extremely organized and focused on continuously learning, upgrading best teaching learning practices, developed processes and procedures over the years for managing life of thousands of students and their families. In last 17 years she has guided 30 $+$ UG projects in IT and Computer Engg. Dr. Devyani received the Lady Engineer Award by IEI, Ahmednagar Local Chapter. ORCID: 0009-0003-6197-2575.
	P. Ramesh, working as Assistant Professor in Aditya College of Engineering since 2009 onwards. I completed my schooling in Saint Judes from my native place., Yellamanchili (Vizag (D.t). Being a Passionate towards Teaching, I am interested and joined as a Assistant professor in 2005. I taught different subjects and satisfied with the teaching till now. Since my childhood i am interested in satellite communications, I want to publish papers in this area.
	Garigipati Rama Krishna, working as Assistant Professor in the Computer science and Engineering Department at Koneru Lakshmaiah Education Foundation (KLEF) FCA.P, India. I have six years of Experience in Teaching Field. Research interests are data mining, big-data, data-science, machine learning and deep learning.

References

Kwak

G.H.

and Park

N.W.

, Unsupervised domain adaptation with adversarial self-training for crop classification using remote sensing images, Remote Sensing 14(8) (2022), 4639.

Tian

Zhang

and Atkinson

P.M.

, Temporal sequence Object-based CNN (TS-OCNN) for crop classification from fine resolution remote sensing image time-series, The Crop Journal 10(5) (2022), 1507–1516.

Chen

Zheng

Wang

Hellwich

Chen

Yang

Liu

Luo

Bao

and Chen

, A joint learning Im-BiLSTM model for incomplete time-series Sentinel-2A data imputation and crop classification, International Journal of Applied Earth Observation and Geoinformation 108 (2022), 102762.

Zhou

Wang

and Iwahori

, Precise crop classification of hyperspectral images using multi-branch feature fusion and dilation-based MLP, Remote Sensing 14(11) (2022), 2713.

Lei

Wang

Zhong

Zhao

and Luo

, DOCC: Deep one-class crop classification via positive and unlabeled learning for multi-modal satellite imagery, International Journal of Applied Earth Observation and Geoinformation 105 (2021), 102598.

Fei

Fan

Wang

Zhang

Wang

Chen

and Bai

, Cotton classification method at the county scale based on multi-features and Random Forest feature selection algorithm and classifier, Remote Sensing 14(4) (2022), 829.

Chen

and Zhang

, A CNN-transformer hybrid approach for crop classification using multitemporal multisensor images, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 13 (2020), 847–858.

Wei

Wang

Liang

Wang

and Cao

, Crops fine classification in airborne hyperspectral imagery based on multi-feature fusion and deep learning, Remote Sensing 13(5) (2021), 2917.

Shakya

Biswas

and Pal

, SAR and Optical data fusion based on anisotropic Diffusion with PCA and Classification using Patch-based with LBP, in: The Proceeding of IEEE India Geoscience and Remote Sensing Symposium (InGARSS), IEEE, Ahmedabad, India, 2021.

10.

Badrinarayanan

Kendall

and Cipolla

, SegNet: A deep convolutional encoder-decoder architecture for image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence 39(12) (2017), 2481–2495.

11.

Cheng

and Kong

, Hyperspectral imaging classification based on LBP feature extraction and multimodel ensemble learning, Computers & Electrical Engineering 92 (2021), 107199.

12.

Xin

and Xiaojun

, Palm vein recognition method based on fusion of local Gabor histograms, The Journal of China Universities of Posts and Telecommunications 24(6) (2017), 55–66.

13.

, The generalized difference vegetation index (GDVI) for dryland characterization, Remote Sensing 6(2) (2014), 1211–1233.

14.

Viña

Gitelson

A.A.

Nguy-Robertson

A.L.

and Peng

, Comparison of different vegetation indices for the remote assessment of green leaf area index of crops, Remote Sensing of Environment 115(2) (2011), 3468–3478.

15.

Sarraf

and Tofighi

, Classification of alzheimer’s disease using fmri data and deep learning convolutional neural networks, arXiv preprint arXiv:1603.08631, 2016.

16.

Mirjalili

, SCA: A sine cosine algorithm for solving optimization problems, Knowledge-Based Systems 96 (2016), 120–133.

17.

Layeb

, Tangent search algorithm for solving optimization problems, Neural Computing and Applications 34(11) (2022), 8853–8884.

18.

Bhuvan

|

NRSC Open EO Dataset will be taken from, https://bhuvan-app3.nrsc.gov.in/data/download/index.php?c=p&s=NI&p=acd&g=TS#, accessed on March 2023.

19.

Khan

M.A.

Tahir

Khurshid

Husnain

M.I.U.

Ahmed

and Boughanmi

, Economic effects of climate change-induced loss of agricultural production by 2050: A case study of Pakistan, Sustainability 12(3) (2020), 1216.

20.

Shi

Wang

and Liu

, Crop yield and production responses to climate disasters in China, Science of the Total Environment 750 (2021), 141147.

21.

Seydi

S.T.

Amani

and Ghorbanian

, A dual attention convolutional neural network for crop classification using time-series sentinel-2 imagery, Remote Sensing 14(3) (2022), 498.

22.

Hamza

M.A.

Alrowais

Alzahrani

J.S.

Mahgoub

Salem

N.M.

and Marzouk

, Squirrel search optimization with deep transfer learning-enabled crop classification model on hyperspectral remote sensing imagery, Applied Sciences 12(11) (2022), 5650.

23.

Arias

Campo-Bescós

M.Á.

and Álvarez-Mozos

, Crop classification based on temporal signatures of Sentinel-1 observations over Navarre province, Spain, Remote Sensing 12(2) (2020), 278.

24.

Jin

and Guo

, A high-precision crop classification method based on time-series UAV images, Agriculture 13(1) (2023), 97.

25.

Wang

and Yang

, A remote sensing method for crop mapping based on multiscale neighborhood feature extraction, Remote Sensing 15(1) (2023), 47.

26.

Zhang

Wei

and Duan

, Learning discriminative spatiotemporal features for precise crop classification from multi-temporal satellite images, International Journal of Remote Sensing 41(8) (2020), 3162–3174.

27.

Yuan

and He

, Application of an anisotropic diffusion based preprocessing filtering algorithm for high resolution remote sensing image segmentation, in: Proceedings of 2008 Congress on Image and Signal Processing, IEEE, Vol. 3, 2008, pp. 629–633.

28.

Alqazzaz

Sun

Yang

and Nokes

, Automated brain tumor segmentation on multi-modal MR image using SegNet, Computational Visual Media 5 (2019), 209–219.

29.

Motohka

Nasahara

K.N.

Oguma

and Tsuchida

, Applicability of green-red vegetation index for remote sensing of vegetation phenology, Remote Sensing 2(10) (2010), 2369–2387.

30.

Naji

T.A.

, Study of vegetation cover distribution using DVI, PVI, WDVI indices with 2D-space plot, in: Journal of Physics: Conference Series, IOP Publishing, Vol. 1003(1), 2018, p. 012083.

31.

Parker

G.G.

, Tamm review: Leaf Area Index (LAI) is both a determinant and a consequence of important processes in vegetation canopies, Forest Ecology and Managemen 477 (2020), 118496.

32.

Metternicht

, Vegetation indices derived from high-resolution airborne videography for precision crop management, International Journal of Remote Sensing 24(14) (2003), 2855–2877.

33.

Roy

D.P.

Boschetti

and Trigg

S.N.

, Remote sensing of fire severity: Assessing the performance of the normalized burn ratio, IEEE Geoscience and Remote Sensing Letters 3(1) (2006), 112–116.

34.

Tan

Sun

J.Y.

Zhang

Chen

Liu

and Liu

X.D.

, Sensitivity of a ratio vegetation index derived from hyperspectral remote sensing to the brown planthopper stress on rice plants, Sensors 19(2) (2019), 375.

35.

Silleos

N.G.

Alexandridis

T.K.

Gitas

I.Z.

and Perakis

, Vegetation indices: Advances made in biomass estimation and vegetation monitoring in the last 30 years, Geocarto International 21(4) (2006), 21–28.

36.

AhilaPriyadharshini

Arivazhagan

Arun

and Mirnalini

, Maize leaf disease classification using deep convolutional neural networks, Neural Computing and Applications 31(12) (2019), 8887–8895.

37.

Rao

, Jaya: A simple and new optimization algorithm for solving constrained and unconstrained optimization problems, International Journal of Industrial Engineering Computations 7(1) (2016), 19–34.

38.

Cheng

and Jin

, A competitive swarm optimizer for large scale optimization, IEEE Transactions on Cybernetics 45(2) (2014), 191–204.

Sine tangent search algorithm enabled LeNet for cotton crop classification using satellite image

Abstract

Keywords

1. Introduction

2. Literature review

3. Sine tangent search algorithm enabled LeNet for hyperspectral cotton crop classification

3.3.1 Architecture of SegNet

3.4.1 Spectral spatial features

3.5.1 Architecture of LeNet

Convolutional layer

Pooling layer

Fully connected layers

Tangent position encoding

Finding fitness

Step 1: Initialization

Step 2: Determining fitness measure

Step 3: Intensification search

Step 4: Exploration search

Step 5: Procedure of escape local minima

Step 6: Termination

Table 1 Parameter details

4.3 Evaluation metrics

4.3.1 Accuracy

4.4.1 Comparative assessment by changing training set

4.5 Analysis with Receiver Operating Characteristic (ROC) curve

4.6 Algorithmic assessment

Table 2 Comparative discussion

4.8 Analysis on confusion matrix

5. Conclusion

Footnotes

Author’s Bios

References

Table 1
Parameter details

Table 2
Comparative discussion