Feature level fusion for land cover classification with landsat images: A hybrid classification model

Abstract

Classification of land cover using satellite images was a major area for the past few years. A raise in the quantity of data obtained by satellite image systems insists on the requirement for an automated tool for classification. Satellite images demonstrate temporal or/and spatial dependencies, where the traditional artificial intelligence approaches do not succeed to execute well. Hence, the suggested approach utilizes a brand-new framework for classifying land cover Histogram Linearisation is first carried out throughout pre-processing. The features are then retrieved, including spectral and spatial features. Additionally, the generated features are merged throughout the feature fusion process. Finally, at the classification phase, an optimized Long Short-Term Memory (LSTM) and Deep Belief Network (DBN) are introduced that portrays classified results in a precise way. Especially, the Opposition Behavior Learning based Water Wave Optimization (OBL-WWO) model is used for tuning the weights of LSTM and DBN. Atlast, many metrics illustrate the new approach’s effectiveness.

Keywords

Land cover feature fusion Optimized Long Short-Term Memory Optimized Deep Belief Network (DBN)Opposition Behavior Learning based Water Wave Optimization Algorithm

1. Introduction

The Land Cover Classification System (LCCS) is a thorough, standardized a priori classification framework that was developed for mapping activities regardless of scale or mapping technique. Data regarding Land Cover is very important for active management, planning, and monitoring, and for the realistic progression of land [1, 2]. As a result, it has to turn out to be more and more noteworthy to precisely scrutinize LC for the rational growth and exploitation of urban land sources [52, 53, 54]. In recent times, owing to speedy urban growth, Land Cover data has been distorted severely in cities that were captured by satellites [3, 4]. The evolution of remote sensing sensors has magnified the availability of high-resolution images, guiding the necessity for extracting data from satellite images [5, 6, 7]. Satellite images are extensively deployed in analyzing and monitoring variations in LC, environment, and so on [8, 9, 10].

Nowadays, information captured by satellites with shorter revisit intervals is widely deployed for scrutinizing and recognizing variations in Land Cover [11, 12]. Fourier analysis is deployed for extracting constraints, which were essential for classifying Land Cover [13, 14]. Nevertheless, these constraints could not acquire variations in Land Cover with time [55, 56]. Moreover, several Machine Learning approaches like Convolutional Neural Networks (CNN), Support Vector Machine (SVM), Deep Neural Networks (DNN), etc have been widely deployed currently owing to their high classification accuracy and minimal time consumption [15, 16, 17].

For advancing the performances of the Land Cover Classification System (LCCS) with satellite images, effectually utilizing spatiotemporal heterogeneity and correlation, as well as picking the proper spatiotemporal levels, are crucial Geographic Object-Based Image Analysis (GEOBIA) was distributed for neglecting the spectral diversity of geographical objects with adequate resultants, and it performs better than other models [18]. Nevertheless, the deployment of GEOBIA for both temporal and spatial data is said to be a problem.

The legacy of the proposed paper is:

•
As an innovation, this paperwork establishes a new land cover classification model, in which the spatial characteristics like Local Binary Pattern (LBP) and Gray Level Co-Occurrence Matrices (GLCM), and spectral features (Normalized Difference Water Index (MNDWI), Normalized Difference Water Index (NDWI), Modified and Normalized Difference Vegetation Index (NDVI) are derived.
•
Introduces a new feature fusion process, by which the extracted spatial and spectral features are fused.
•
Deploys optimized Long Short-Term Memory (LSTM) and Deep Belief Networks (DBN) for accurate classification outputs.
•
Employs a new Opposition Behavior Learning based Water Wave Optimization (OBL-WWO) model for determining the standard weights in LSTM and DBN.

The review of this paperwork is mentioned in Section 2. In Section 3, the suggested land cover classification framework is described step by step. In Section 4, a novel feature fusion approach is used to extract spatial and spectral characteristics. In Section 5, LSTM and DBN classification framework hybrid classifiers are shown. Section 6 provides an overview of the OBL-WWO optimization technique, while Sections 7 and 8 present the outcomes and recommendations.
2. Literature review

In 2019, Chen et al. [19] suggested a novel land cover classification model that avoided the misclassification that occurred by scale effect by merging Multi-Scale CNN (MCNN). Subsequently, this work analyzed how the scaling constraint of CNN impacted the categorization accurateness and involved spatial data for pre-estimating scale constraint in per-super pixel CNN. The investigational results have shown that MCNN could efficiently evade misclassification.

In 2021, Kareem et al. [20] designed a CNN for classifying the multispectral SAT-4 images into 4 groups: barren land, grassland, trees, and others. The adopted CNN classifier learned the spatial and spectral properties of images from provided samples of ground truth. The concern of this work was 3-fold. A classification model for extraction of features along with normalization. 9 diverse designs of Convolutional Neural Networks (CNN) were constructed, and numerous experimentations were done for classifying the images. A deep perceptive of image resolution and structure was captured by altering diverse optimizers in CNN. Finally, better and more accurately predicted outcomes were obtained using the developed scheme.

In 2019, He et al. [21] suggested the latest fuzzy improbability modeling scheme for presenting the land cover pattern features and presented a modified Fuzzy Clustering technique. The adopted fuzzy improbability modeling scheme was executed in the 2 most important stages. Initially, the input segmentation unit was subjected to object-oriented symbolic modeling. Consequently, features for every type of land cover were presented in the structure of a symbolic vector that described the intra-class uncertainties and improved the separability among diverse groups. Then, type-2 fuzzy was produced for every cluster depending upon the distance of interval-valued vectors.

In 2019, Tabib et al. [22] proposed a feature-level synthesis of Synthetic Aperture Radar (SAR) images for modifying the accuracy of urban Land Cover categorization. In the adopted object-oriented approach, segmented areas were deployed for carrying out knowledge-oriented classification depending on land surface temperature, spectral associations, and SAR quality characteristics calculated in Gray Level Co-Occurrence Matrices (GLCM) space. The estimated resultants had shown the development regarding kappa and accuracy using the developed model.

In 2019, Unnikrishnan et al. [23] utilized the Near Infra Red (NIR) and red band data for classifying satellite datasets. It was done, as ND Vegetation Index (NDVI) calculation required only 2 bands (NIR and red) data, and the classes concerned in the dataset were of vegetation type. This paperwork deployed a novel Deep Learning model for 3 diverse networks by adjusting the network. The customized architecture with a condensed count of filters was tested and trained and thus it was managed to categorize the images into diverse groups. The adopted model was evaluated in opposition to extant models regarding precision, accuracy, and trainable constraints.

In 2019, Mishra et al. [24] examined the textural features of satellite images for enhancing the classification accuracy of Land Use/Land Cover (LULC). The textual characteristics were acquired from sensor data with the support of GLCM. The finest grouping of textural features was identified make use of correlation coefficients and standard deviations thereafter separability study of the LULC group depending upon a training sample. An SVM model was deployed to carry out categorization and the resultants were computed accordingly.

In 2021, He et al. [25] proposed a multi-spectral LULC technique depending on the DL model. By deploying the estimable detailed capture capability of contour let transform for obtaining the possible data, a spectral-texture classification scheme was constructed. Thus, the feature spaces merged with DL for feature extraction were enhanced. In the end, the investigation was done that substantiated the modification of the used strategy concerning accuracy.

In 2019, Jayanth et al. [26] deployed Elephant Herding Optimization (EHO) model for analyzing multispectral pixels and for determining the data of classes. While the spectral resolution of satellite images was amplified, the statistical separability among LULC groups in spectral spaces was reduced and it tends to maintain the accuracy of classification. These were mainly for each pixel and parametrical in nature. Investigational results have exposed that EHO has shown a development over Support Vector Machine (SVM).

In 2020, Mukherjee et al. [50] investigated the Indian cities of Surat and Bharuch over 20 years to know the alternate in land use, land cover, and surface temperature using data from Landsat 5 Thematic Mapper and Landsat 8 OLI/TIRS. The explorations show that both cities have gently experienced an enormous amount of expansion in the built-up area and a loss of green space in LST.

In 2021, Hong et al. [51] suggested the shared and specific feature learning (S2FL) model for land cover classification. Especially for heterogeneous data sources, S2FL can decompose multimodal RS data into modality-specific components and modality-shared, making it possible to blend information from many modalities more successfully. Extensive testing on the three datasets shows that by classifying land cover, our S2FL model performs better than the techniques in use.

The overview of traditional methods for classifying land cover is shown in Table 1. Initially, the CNN model used in [19] provides accurate classification with minimal error values. However, it needs exploration on a larger scale area. CNN approach [20] improves the accuracy with reduced vegetation loss. However, larger remote sensing datasets are not analyzed in this work. Type-2 fuzzy [21] has enhanced accuracy with lesser noise; however, local and global relations are not attained. GLCM was exploited in [22] that increased accuracy with better kappa values; however, it needs to focus on contextual information. The convNet model adopted in [23] is very much accurate and offers high precision. Nevertheless, the count of training constraints needs to be minimized. Moreover, the GLCM method was deployed in [24] that offer high spatial resolution with high accuracy. Nevertheless, it needs new schemes for the automated election of textural features. DBN method [25] ensures high accuracy and perfect mining of spatial distribution law; however, reliability needs more exploration. Besides, the EHO approach was established in [26], which provides high Producer Accuracy with high User Accuracy. Although, it should concentrate on change detection.

Table 1
Review of conventional land cover classification schemes

Adopted method	Features	Challenges	Author
CNN	Accurate classification Minimal error values	A larger scale area needs to be considered.	Chen et al. [19]
CNN method	Reduced vegetation loss High accuracy	A larger remote-sensing dataset was not considered	Kareem et al. [20]
Type-2 fuzzy	Enhanced accuracy Less noise	No consideration of local and global relations.	He et al. [21]
GLCM	High accuracy Better kappa values	Contextual information must be given additional consideration.	Tabib et al. [22]
ConvNet model	Higher precision High accurate outcomes	The count of training constraints needs to be minimized.	Unni krishnan et al. [23]
GLCM	High spatial resolution High accuracy	Needs new schemes for the automated election of textural features.	Mishra et al. [24]
DBN	High accuracy Perfect mining of spatial distribution law	Reliability needs more exploration.	He et al. [25]
EHO	High PA High UA	Change detection should be concerned more.	Jayanth et al. [26]

Numerous models have been focused on under land cover classification models. However, there exist frequent problems such as no consideration in large remote sensing datasets, no consideration of local and global relations, and reliability, there is a need for a new scheme for the automated selection of textural features, and training constraints ought to be minimized. Therefore, to rectify the abovementioned issues, this study develops a novel land cover classification model. The proposed model is described in the following section.

3. A stepwise portrayal of the adopted land cover classification approach

The proposed Land Cover classification method includes 4 vital phases. Initially, in pre-processing Histogram Equalization process is employed. After that, the spatial features such as LBP features $({Fe_{\textit{lbp}}})$ , GLCM features $({Fe_{GL}})$ , and spectral features including NDWI $({Fe_{\textit{NDWI}}})$ , MNDWI, $({Fe_{\textit{MNDWI}}})$ and NDVI features $({Fe_{\textit{NDVI}}})$ , are derived. Subsequently, the feature-level fusion technique is applied to all features and thus fused features are obtained.

Next, the merged features are fed as input to optimized DBN and optimized LSTM that provides the final classified output. Specifically, the OBL-WWO framework assists in achieving improved classification results in a precise manner and is utilized to optimize both the weights of LSTM and DBN.

The demonstration of the created OBL-WWO-based system is shown in Fig. 1.

Figure 1.

Demonstration for developed Land-cover classification model.

The satellite image (Im) is pre-processed initially using the HE approach, which alters the magnitude of the image to enhance variation. This method typically increases the overall contrast of many photos, mainly while the utilizable information of the image is symbolized by closer contrast values. This permits lower contrast areas to achieve high contrast.

The extracted features are then applied to the obtained image, represented by $({\text{Im}_{HE}})$ .

4. Extraction of spatial and spectral features with novel feature fusion process

4.1 Spatial features

Local Binary Pattern

The LBP descriptors are demonstrated with increased differentiative capacity and clarity in a variety of relative analyses [27, 28]. Moreover, the basic LBP is frequently engaged to extract the distinguished features with a specified reference pixel $K_{c}$ and its neighbors with radius $\hat{R}$ as in Eq. (1). The $K_{c}$ neighbor is signified as $K_{G,\hat{R}}$ and neighborhood index in a certain local sub-region $Q$ , enclosed by $K_{c}$ is signified as $\hat{g}$ . The $s_{1}({{}^{\prime},^{\prime}})$ coding term is determined with $K_{c}$ as well as neighborhood pixel for selecting the entire neighborhoods label. The threshold operation is shown in Eq. (2), wherein $s_{1}({{}^{\prime},^{\prime}})$ points out the orientation of the binary gradient [28].

$\displaystyle Fe_{\textit{lbp}_{G,\hat{R}}}({K_{c}})=s_{1}\left({\begin{array}% []{l}Q({K_{1,\hat{R}}}),\\ ,Q({K_{c}})\\ \end{array}}\right),s_{2}\left({\begin{array}[]{l}Q({K_{2,\hat{R}}}),\\ ,Q({K_{c}})\\ \end{array}}\right),\ldots,s_{1}\left({\begin{array}[]{l}Q({K_{\hat{g},\hat{R}% }}),\\ Q({K_{c}})\\ \end{array}}\right)$ (1) $\displaystyle s_{1}\left({\begin{array}[]{l}Q(K_{\hat{g},\hat{R}}),\\ ,Q(K_{c})\\ \end{array}}\right)=\left\{{{\begin{array}[]{ll}1&{\text{if }({1-({K_{\hat{g},% \hat{R}}})-Q({K_{c}})})\geqslant 0}\\ 0&\text{else}\\ \end{array}}}\right.$ (2)

Gray Level Co-Occurrence Matrices

It is established for calculating the spatial relationship amongst the pixel [29]. The succinct simplification of GLCM is mentioned in Table 2.

Table 2

Features of GLCM

S. no.	Features	Arithmetical term
1.	“Entropy	$\textit{Entropy}=-\sum\limits_{a}\sum\limits_{b}{s_{ab}}\log_{2}s_{ab}$
2.	Energy	$E=\sum\limits_{a}\sum\limits_{b}s_{ab}^{2}$ here $s_{ab}$ is the $({a,b})^{\text{th}}$ entry in GLCM
3.	Variance	$V=\sum\limits_{a}\sum\limits_{b}({a-\mu})^{2}s_{ab}$ , where $\mu$ specifies the mean of $s_{ab}$
4.	Contrast	$\textit{Con}=\sum\limits_{a}\sum\limits_{b}({a-b})^{2}s_{ab}$
5.	Correlation	$\textit{Cor}=\frac{\sum\limits_{a}\sum\limits_{b}({ab})s_{ab}-\mu_{x}\mu_{y}}{% \sigma_{x}\sigma_{y}}$ , where $\sigma_{x}$ , $\sigma_{y}$ , $\mu_{x}$ , $\mu_{y}$ are the std deviations
		and mean of $s_{x}$ , $s_{y}$
6.	Sum Average	$SA=\sum\limits_{a=2}^{2N_{s}}a.s_{x+y}(a)$ , where $N_{s}$ indicates the varied gray levels in the
		image.
7.	Homogeneity	$H=\sum\limits_{a}\sum\limits_{a}\frac{1}{1+({a-b})^{2}}s_{ab}$
8.	Sum Variance	$SV=\sum\limits_{a=2}^{2N_{s}}{({a-SE})^{2}s_{x+y}(a)}$
9.	Sum Entropy	$SE=\sum\limits_{a=2}^{2N_{s}}s_{x+y}(a)\log\{{s_{x+y}(a)}\}$
10.	Difference Variance	$DV=$ variance of $s_{a_{x-y}}$
11.	MCC (2 ${}^{\text{nd}}$ higher Eigenvalue of $Q$ ) ${}^{0.5}$	$\textit{MCC}=\sum\limits_{k}\frac{g({a,k})g({b,k})}{g_{x}(a)g_{y}(k)}$
12.	Difference Entropy	$DE=\sum\limits_{a=0}^{N_{s-1}}s_{x-y}(a)\log\{{s_{x-y}(a)}\}$
13.	Information Measures of Correlation 1	$\textit{IMC}1=\frac{\textit{HXY}-\textit{HXY}1}{\max\{{HX,HY}\}}$
14.	Information Measures of Correlation 2	$\textit{IMC}2=\sqrt{({1-\exp[{-2.0[{\textit{HXY}2-\textit{HXY}}]}]})}$ , where
		$\textit{HXY}=-\sum\limits_{a}{\sum\limits_{b}{s_{ab}}}\log_{2}s_{ab}\textit{% HXY}1=-\sum\limits_{a}\sum\limits_{b}{s_{ab}}\log_{2}\{{s_{x}(a)s_{y}(b)}\}$ ”

$Fe_{\textit{GLCM}}$ represents the retrieved GLCM features. The derived spatial features (LBP $({Fe_{\textit{lbp}}})$ as well as GLCM features $({Fe_{\textit{GLCM}}})$ ) are together pointed out $Fe_{ST}$ .

4.2 Spectral features

ND Vegetation Index (NDVI)

It [28] is “a simple graphical indicator that can be used to analyze remote sensing measurements, typically, but not necessarily, from a space platform, and assess whether the target being observed contains live green vegetation or not”. It is computed as specified in Eq. (3), in which, NIR denoted near-infrared band.

$\displaystyle\textit{NDVI}=\frac{\textit{NIR}}{\textit{NIR}+\textit{Red}}-% \frac{\textit{Red}}{\textit{NIR}+\textit{Red}}$ (3)

Modified Normalized Difference Water Index (MNDWI)

MNDWI uses SWIR and green bands for enhancing the open water feature and it is evaluated as in Eq. (4). It moreover lessens urbanized area features, which are often related to open water in another index.

$\displaystyle\textit{MNDVI}=\frac{({\textit{Green}-\textit{SWIR}})}{({\textit{% Green}+\textit{SWIR}})}$ (4)

Normalized Difference Water Index (NDWI)

NDWI is utilized for observing variations associated with water content, using NIR and green wavelength. It is evaluated as in Eq. (5).

$\displaystyle\textit{NDWI}=\frac{({\textit{Green}-\textit{NIR}})}{({\textit{% Green}+\textit{NIR}})}$ (5)

The derived spectral features (NDVI, MNDWI as well as NDWI features) are together pointed out $Fe_{SP}$ . Accordingly, spatial features $Fe_{ST}$ and spectral features $Fe_{SP}$ are further given as input to optimized hybrid classifiers.

4.3 Proposed feature fusion process

The extracted spatial features and spectral features are fused before being provided to the classification stage. Let the extracted spatial features be denoted by $Fe_{ST1},Fe_{ST2}\ldots Fe_{\textit{STN}}$ and let the extracted spectral features be denoted by $Fe_{SP1},Fe_{SP2}\ldots Fe_{\textit{SPN}}$ . The fused features are depicted in Eq. (6), wherein, $Fe_{SP1}$ and $Fe_{SP2}$ are calculated as displayed in Eqs (7) and (8). In Eq. (6), $w_{1}$ , $w_{2}$ refers to weights that are generated randomly among 0 and 1.

$\displaystyle FF=w_{1}Fe_{SP1}+w_{2}Fe_{ST2}$ (6) $\displaystyle Fe_{SP1}=\frac{1}{N}\sum\limits_{i=1}^{N}{\frac{1}{Fe_{{SP}_{i}}% ^{\prime}}}$ (7) $\displaystyle Fe_{ST2}=\frac{1}{N}\sum\limits_{i=1}^{N}{\frac{1}{Fe_{\textit{% STi}}^{\prime}}}$ (8)

The optimized LSTM and DBN are then provided with the merged features as input for categorization.

5. Hybrid classifiers: Long short-term memory and deep belief network frameworks for classification

LSTMs are frequently used to learn, process, and classify sequential data, as these networks are capable of learning long-term connections between time steps of data. Also, DBN classifiers have the benefit of being less computationally costly. Hence, in this work, to land cover classification the LSTM and DBN classifiers are hybridized to make the classification result more accurate.

5.1 Optimized long short-term memory classifier

It [30] includes a series of recurring LSTM cells. Three units, including “the input gate, output gate as well as forget gate,” make up every LSTM cell. Let the variables $M$ $C$ indicate hidden and cell state. $({X_{t},C_{t-1},M_{t-1}})$ and $({M_{t},C_{t}})$ signifies input and output layers.

The $O_{t},I_{t},F_{t}$ specifies the output, input, and forget gate at the time $t$ . LSTM primarily deploys $F_{t}$ for arranging data to ignore. The arranged data denote specific partial features associated with prior gaze direction; $F_{t}$ is formulated as specified in Eq. (9).

$\displaystyle F_{t}=\sigma({J_{IF}X_{t}+B_{IF}+J_{MF}M_{t-1}+B_{MF}})$ (9)

In Eq. (9), $({J_{MF},B_{MF}})$ and $({J_{IF},B_{IF}})$ imply weight and bias parameter to map hidden and input layers to forget the gate and activation function is implied $\sigma$ .

The LSTM exploits the input gate as shown in Eqs (10)–(12), which, $({J_{MG},B_{MG}})$ as well $({B_{IG},J_{IG}})$ signifies bias and weight parameters to map hidden as well as input layers to the cell gate. $({B_{MI},J_{MI}})$ and $({B_{II},J_{II}})$ imply bias and weight parameters to map hidden and input layers $I_{t}$ .

$\displaystyle G_{t}=\tanh({B_{IG}+J_{IG}X_{t}+B_{MG}+J_{MG}M_{t-1}})$ (10) $\displaystyle I_{t}=\sigma({J_{II}X_{t}+B_{MI}+J_{MI}M_{t-1}+B_{II}})$ (11) $\displaystyle C_{t}=I_{t}G_{t}+F_{t}C_{t-1}$ (12) $\displaystyle O_{t}=\sigma({J_{IO}X_{t}+B_{IO}+J_{MO}M_{t-1}+B_{MO}})$ (13) $\displaystyle M_{t}=O_{t}\tanh({C_{t}})$ (14)

According to Eqs (13) and (14), the LSTM cell receives the output hidden layer from the output unit, wherein $({B_{MO},J_{MO}})$ and $({B_{IO},J_{IO}})$ implies weight and bias to map the hidden and input layer $O_{t}$ . Accordingly, the weights $J_{IF}+J_{MF}+J_{MG}+J_{MI}+J_{IG}+J_{II}=J$ are optimally chosen by the developed OBL-WWO model.

5.2 Optimized deep belief network classifier

DBN [31] comprises different layers requiring apparent together with hidden neurons. The output denoted by $\overline{PO}$ is modeled in Eq. (16) and the probability function $\bar{P}_{q}(\zeta)$ is modeled in Eq. (15), where $t^{P}$ denotes pseudo-temperature. The DBN model is mentioned in Eq. (17).

$\displaystyle\bar{P}_{q}(\zeta)=\frac{1}{1+e^{\frac{-\zeta}{t^{P}}}}$ (15) $\displaystyle\overline{PO}=\left\{{{\begin{array}[]{ll}1&{\text{with }1-\bar{P% }_{q}(\zeta)}\\ 0&{\text{with }\bar{P}_{q}(\zeta)}\\ \end{array}}}\right\}$ (16) $\displaystyle\mathop{\lim}\limits_{t^{P}\to 0^{+}}\bar{P}_{q}(\zeta)=\mathop{% \lim}\limits_{t^{P}\to 0^{+}}\frac{1}{1+e^{\frac{-\zeta}{t^{P}}}}=\left\{{{% \begin{array}[]{lll}0&\text{for}&{\zeta<0}\\ {\frac{1}{2}}&\text{for}&{\zeta=0}\\ 1&\text{for}&{\zeta>0}\\ \end{array}}}\right.$ (17)

The mathematical model to arrange the binary state is shown in Eqs (18) and (19), which $L_{a,l}$ points out weights amid neurons as well as $\theta_{a}$ points out biases.

$\displaystyle EN({bi})=-\sum\limits_{a}{bi_{a}\theta_{a}}-\sum\limits_{a<l}{bi% _{a}}L_{a,l}bi_{l}$ (18) $\displaystyle\Delta EN({bi_{a}})=\sum\limits_{l}{\theta_{a}+L_{a,l}bi_{l}}$ (19)

The illustration of energy for hidden neuron and visible $({x,y})$ is shown by Eqs (20)–(22), which $C_{l}$ indicates biases $y_{l}$ $x_{a}$ mentioned the binary state of the hidden unit $l$ and visible unit $ak_{a}$ .

$\displaystyle EN({x,y})=-\sum\limits_{({a,l})}{L_{a,l}}x_{a}y_{l}-\left[{\sum% \limits_{a}{k_{a}}x_{a}+\sum\limits_{l}{C_{l}}y_{l}}\right]$ (20) $\displaystyle\Delta EN({x_{a},\bar{y}})=\left({k_{a}+\sum\limits_{l}{L_{al}}y_% {l}}\right)$ (21) $\displaystyle\Delta EN({\vec{x},y_{l}})=\left({C_{l}+\sum\limits_{a}{L_{al}}x_% {a}}\right)$ (22)

Equation (23) shows the final weight distribution after RBM training.

$\displaystyle W_{(\hat{m})}=\mathop{\max}\limits_{W}\mathop{\Pi}\limits_{\vec{% x}\in N}c(\vec{x})$ (23)

The RBM uses the energy function described in Eq. (24) to calculate a chance, wherein $PR^{F}$ implies a splitting expression as modeled in Eq. (25).

$\displaystyle c({\vec{x},\vec{y}})=\frac{1}{PR^{F}}e^{-EN({\vec{x},\vec{y}})}$ (24) $\displaystyle PR^{F}=\sum\limits_{\vec{x},\vec{y}}{e^{-EN({\vec{x},\bar{y}})}}$ (25)

The procedures of the DBN model are as follows.

Step 1: Select $x$ samples.

Step 2: Computation has hidden neuron probability $c_{y}$ as $c_{y}=\sigma({x.W})$ based upon Eq. (26), which $\sigma$ denotes activation function. Here, the weight factor $W$ is optimally selected through the OBL-WWO model.

$\displaystyle c({\vec{y}_{l}\to 1|\bar{x}})=\sigma\left({C_{l}+\sum\limits_{a}% {x_{a}L_{a,l}}}\right)$ (26)

Step 3: Analyse hidden condition $y$ from $c_{y}$ probability.

Step 4: Calculate $x$ and $c_{y}$ that is calculated as a positive gradient $\phi^{+}=x.c_{y}^{t^{P}}$ .

Step 5: Examine the transformation of ${x}^{\prime}$ visible states $y$ as revealed in Eq. (27).

$\displaystyle c({\vec{x}_{l}\to 1|\vec{y}})=\sigma\left({k_{a}+\sum\limits_{a}% {x_{l}L_{a,l}}}\right)$ (27)

Step 6: Calculate external product by negative gradients $\phi^{-}={x}^{\prime}.{y}^{\prime t^{P}}$ .

Step 7: The reformed weight as specified by Eq. (28), which $\eta$ denotes the learning rate.

$\displaystyle\Delta\hat{L}=\eta({\phi^{+}-\phi^{-}})$ (28)

Step 8: Equation (29) delineates the upgrade of weights with new values.

$\displaystyle L^{\prime}_{a,l}=L_{a,l}+\Delta L_{a,l}$ (29)

Let the training pattern be $(K^{\hat{H}},U^{\hat{H}})$ , wherein $K^{\hat{H}}$ and $U^{\hat{H}}$ denotes input and output vector, and $1\leqslant\hat{H}\leqslant V$ , $V$ indicate training pattern count. The neuron miscalculation at the output is defined by Eq. (30. Eventually, the square error of $\hat{H}$ the pattern is mentioned in Eq. (31). $SE_{\textit{avg}}$ is given in Eq. (32). The aim of this paperwork is to reduce the faulty as shown in Eq. (33), thereby, upgrading the exactness.

$\displaystyle e_{l}^{\hat{H}}=K^{\hat{H}}-U^{\hat{H}}$ (30) $\displaystyle SE_{\hat{H}}=\frac{1}{\tilde{o}_{y}}\sum\limits_{l=1}^{\tilde{o}% _{y}}{({e_{l}^{\hat{H}}})}^{2}=\frac{1}{\tilde{o}_{y}}\sum\limits_{l=1}^{% \tilde{o}_{y}}{({K^{\hat{H}}-U^{\hat{H}}})}^{2}$ (31) $\displaystyle SE_{\textit{avg}}=\frac{1}{V}\sum\limits_{\hat{H}=1}^{V}{SE_{% \hat{H}}}$ (32) $\displaystyle\textit{Obj}=\textit{Min}(e)$ (33)

6. Introduction of opposition behavior learning-based WWO (OBL-WWO) algorithm for optimization

Solution Encoding

As previously stated, the LSTM weights represented by $(J)$ and DBN weights indicated by $(W)$ are selected using the OBL-WWO technique. The interpretation for solutions is represented in Fig. 2, which $N l$ indicates the full quantity of LSTM weights and $N r$ denotes the full quantity of DBN weights.

Figure 2.

Solution encoding.

The traditional WWO model [32] shows excellent solutions; although, it undergoes low accuracy. To label the disadvantages of classical WWO, upgrading is made to the suggested algorithm. Own-suggestion is accepted to be encouraging in conventional optimization algorithms [33, 34, 35, 36, 37, 38, 39, 40]. OBL-WWO [32] is deployed for resolving optimization issues. The OBL-WWO approach has many benefits, like being easy to implement, requiring fewer control parameters and population vectors, and being particularly effective at finding the best solution in a high-amplitude search space. When compared with local search optimization like hill climbing, tabu search, and global optimization algorithms like evolutionary algorithms, and memetic algorithms, the proposed OBL-WWO approach offers huge benefits and helps in fine-tuning the mass of DBN and LSTM. Therefore, influenced by the advantages of the OBL-WWO method this paper utilized OBL-WWO for land cover classification.

The fitness $r\in S$ is evaluated contrariwise through the seabed’s depth. The fitness is raised based on the space to still water. The model holds a population with height $H$ and wavelength $\lambda$ . For all waves, $H$ is constant $H^{\max}$ and $\lambda$ be 0.5. The method involves 3 functions: “(i) Propagation (ii) refraction (iii) Breaking”.

The suggested method made use of the OBL paradigm, which is designed to be used with real people and its contradictions. To remain with the greatest alternative, the values, and their opposing points are thus estimated together.

Propagation

During this phase, a novel wave ${r}^{\prime}$ is generated, which is acquired by operating the measurements $\hat{d}$ of $r$ the wave as exposed in Eq. (34), wherein, $\textit{rand}[{-1,1}]$ pointing out arbitrary count among ( $-$ 1, 1), $LE(\hat{d})$ symbolizes the dimension length $\hat{d}$ .

$\displaystyle{r}^{\prime}(\hat{d})=r(\hat{d})+\textit{rand}[{-1,1}]\cdot% \lambda LE(\hat{d})$ (34)

The observations of all $r$ are upgraded following all generations as in Eq. (35), wherein $Q_{\max}$ and $Q_{\min}$ designates the maximal and minimal fitness values amongst the current population, in that order, $\alpha$ designates the wavelength reduction coefficients and $\varepsilon$ designates an integer to evade division-by-0.

$\displaystyle\lambda=\lambda\cdot\alpha^{-({Q(r)-Q_{\min}+\varepsilon})/({Q_{% \max}-Q_{\min}+\varepsilon})}$ (35)

Refraction

When determining the place following refraction, as in Eq. (36), it is done on waves involving altitudes that have been reduced to zero, wherein, $r^{*}$ points out the best solution and $GN({\mu,\sigma})$ points out Gaussian arbitrary integer that includes $\sigma$ standard deviation and $\mu$ means.

$\displaystyle{r}^{\prime}(\hat{d})=GN\left({\frac{r^{*}(\hat{d})+r(d)}{2},% \frac{|r^{*}(\hat{d})-r(\hat{d})|}{2}}\right)$ (36)

After refraction, the height ${r}^{\prime}$ is returned $H_{\max}$ , and the wavelength is set as per Eq. (37).

$\displaystyle{\lambda}^{\prime}=\lambda\frac{Q(r)}{Q({r}^{\prime})}$ (37)

Breaking

It is done on a wave $r$ that acknowledges the new greatest results $r^{*}$ , and consequently, it performs a local search near $r^{*}$ for starting wave breaking. As per the advanced OBL-WWO model, the new update is designed as shown in Eq. (37), wherein, $r^{*}$ denotes the current best individual and $\eta$ refers to controlling constraints for the magnitude of learning.

$\displaystyle{r}^{\prime}(\hat{d})=r_{\textit{best}}(\hat{d})+({r^{*}(\hat{d})% -r(\hat{d})}).\eta$ (38)

If there are no solitary waves, then it is superior to $r^{*}$ and $r^{*}$ exist, if not $r^{*}$ is replaced with a fitter one.

Algorithm 1: Opposition Behavior Learning-based WWO Algorithm
1. Initialize the population
2. Apply OBL and generate opposite solutions
3. while (no termination condition) do
4. For each $r$ does
5. Propagate $r$ to ${r}^{\prime}$ as revealed in Eq. (34)
6. if $F({r}^{\prime})>F(r)$ then
7. if $F({r}^{\prime})>F({r^{*}})$ then
8. Break ${r}^{\prime}$ based on the proposed formulation as exposed in Eq. (38)
9. Update $r^{*}$ with ${r}^{\prime}$
10. Substitute $r$ with ${r}^{\prime}$
11. else
12. Decrease $r\cdot H$ by one
13. if $r\cdot H=0$ then
14. Refract $r$ to ${r}^{\prime}$ as in Eqs (36) and (37)
15. Determine $\alpha$
16. Update the wavelength as in Eq. (35)
17. Return $r^{*}$

7. Results and discussion

7.1 Simulation set up

A projected HC $+$ OBL-WWO-related LC classification plan was accomplished in Matlab. Now, the review was over using a dataset that was downloaded from [41]. Appropriately, the execution of the approved approach was calculated over remaining models such as WWO [32], PRO [42], SMO [43], SLO [44], SSOA [45], CNN [46], SVM [47], RNN [48], BI-LSTM [30] and MPU-WWO [49] regarding varied metrics. However, the execution study was accomplished by replacing the LR with values between 60 and 80. Selected strategies, statistics, and convergence studies were also executed to demonstrate efficacy. In this research, the training, as well as testing rates, are set at 70% and 30%. The sample image showing the segmented outputs (land, vegetation, and water) is shown in Fig. 3.

Figure 3.

Sample image of (a) water (b) vegetation (c) land for original image 1 (d) water (e) vegetation (f) land for original image 2.

7.2 Performance analysis

Comparing created HC $+$ OBL-WWO performance to the traditional system, various measures are used as shown in Figs 4–6. The evaluation of the proposed HC $+$ OBL-WWO framework over HC $+$ WWO, HC $+$ PRO, HC $+$ SMO, HC $+$ SLO, HC $+$ SSOA, CNN, SVM, RNN, BI-LSTM, and MPU-WWO models by altering LR among 60, 70, and 80. Consequently, the earth explorer dataset was used for the investigation, and the analogous solution is presented in Fig. 4. The given HC $+$ OBL-WWO model outperformed the evaluated methods after examining all the graphs. Excellent positive values and small negative values primarily make certain the model’s greater categorization rate. Looking at Fig. 4a, we can see that the developed technique has a precision of 0.857 at the 60th LR, while the proposed methodology has a greater accuracy of 0.833333 and 0.833333 at the 70th LR and 80th LR. The outputs for all positive measures fall as LR rises, whereas the outputs for negative measures increase as LR increases. Particularly, both suggested and conventional techniques provide enormous yields at the 60 ${}^{\text{th}}$ LR. However, the presented system has produced more useful results than distinct strategies for all LRs. Specifically from Fig. 4b, the established technique at the 60th LR achieves the highest higher accuracy (0.92) in comparison to the outcomes achieved during other LRs. For illustration, at 60 ${}^{\text{th}}$ LR, the suggested approach’s accuracy is 10%, 6.45%, 17.86%, 17.86%, 10%, 22.22%, 57.14%, 17.86%, 22.22%, and 6.45% superior to the values attained for an existing strategy like HC $+$ WWO, HC $+$ PRO, HC $+$ SMO, HC $+$ SLO, HC $+$ SSOA, CNN, SVM, RNN, BI-LSTM, and MPU-WWO accordingly. Similarly, the accepted model’s negative measures show marginal improvements over other methods. As a consequence, the analysis proved that the addition of optimization theory to the HC $+$ OBL-WWO approach increased its effectiveness.

Figure 4.

Analysis of the proposed method over the existing method concerning positive measures.

Figure 5.

Analysis of the proposed method over the existing method concerning neutral measures.

Figure 6.

Analysis of the proposed method over the existing method concerning negative measures.

7.3 Convergence analysis

The assumption of the HC $+$ OBL-WWO method over conventional methods like HC $+$ WWO, HC $+$ PRO, HC $+$ SMO, HC $+$ SLO, and HC $+$ SSOA methods for convergence (cost) evaluation are compared in Fig. 7, for different iterations. The analysis is accomplished by altering the repetitions from 0, 5, 10, 15, 20, and 25. The prescribed HC $+$ OBL-WWO have acquired the least value for all iterations than the comparative ones, as can be seen from the OUTCOMES. The cost numbers for the designed and estimated systems begin with larger from repetition 0 to 2, but greater (minimum cost)results arrive as the iteration count increases. Specifically, from iteration 12 to iteration 24, both the suggested and evaluated models’ cost values continue to decline, but the implemented HC $+$ OBL-WWO strategy exhibits values that are smallest compared to the current ones. In particular, the combined use of the introduced optimization technique has allowed the proposed strategy to achieve the lowest cost value (almost 1.46). The total evaluation thus supports the enlargement of the current proposal. This determines how the hybrid algorithm converges to the desired fitness.

Figure 7.

Convergence analysis.

7.4 Statistical analysis

The statistical analysis of the approved HC $+$ OBL-WWO method over traditional approaches is shown in Table 3 in terms of exactness. The met investigative approaches are theoretical, and to achieve Eq. (33), every model is repeatedly verified to make sure a correct evaluation. Examining the experimental findings, it was established that the HC $+$ OBL-WWO strategy produced the lowest outcomes for each situation. Following the development of the model, the HC $+$ WWO strategy has achieved lower cost ratios than HC $+$ PRO, HC $+$ SMO, HC $+$ SLO, and HC $+$ SSOA methods. The proposed framework achieves the lowest median case situation results as it is strengthened using the optimization concept. The HC $+$ OBL-WWO model has also obtained minimal results for each case, demonstrating the superiority of the principles used.

Table 3
Statistical analysis

Methods	Standard deviation	Mean	Median	Maximum	Minimum
HC $+$ WWO	0.042082	1.513296	1.531915	1.542857	1.465116
HC $+$ PRO	0.080307	1.627308	1.636364	1.702703	1.542857
HC $+$ SMO	0.037638	1.535638	1.531915	1.575	1.5
HC $+$ SLO	0.072999	1.542543	1.542857	1.615385	1.469388
HC $+$ SSOA	0.06407	1.533226	1.565217	1.575	1.459459
HC $+$ OBL-WWO	0.036313	1.497125	1.5	1.531915	1.459459

7.5 Region of curve analysis

“When a binary classifier system’s discrimination threshold is changed, a Receiver Operating Characteristic (ROC) curve, a graphical representation, shows how well the algorithm can diagnose the problem”. The Region of the Curve (FPR) is yielded by the representation of the True Positive Rate (TPR) against the False Positive Rate (FPR). Figure 8 shows the ROC curve attained by the proposed HC $+$ OBL-WWO method and traditional HC $+$ WWO, HC $+$ PRO, HC $+$ SMO, HC $+$ SLO, HC $+$ SSOA, CNN, SVM, RNN, BI-LSTM, and MPU-WWO respectively.

Figure 8.

Analysis of the proposed model and existing models for ROC.

8. Conclusion

The new Land Cover classification method with enhanced hybrid classifiers has been developed in this article. After performing pre-processing, spatial and spectral characteristics were extracted. Through the application of upgraded LSTM and upgraded DBN, these features were then combined and further classified. Specifically, a new OBL-WWO algorithm was employed to optimize the weights of the LSTM and DBN. Finally. it was established that the advanced method was suitable to the conventional method in many ways. At 60 ${}^{\text{th}}$ LR, the accuracy of the planned strategy is 10%, 6.45%, 17.86%, 17.86%, 10%, 22.22%, 57.14%, 17.86%, 22.22%, and 6.45% superior to the values attained for existing methods like HC $+$ WWO, HC $+$ PRO, HC $+$ SMO, HC $+$ SLO, HC $+$ SSOA, CNN, SVM, RNN, BI-LSTM, and MPU-WWO accordingly. Therefore, the presented approach’s superiority was firmly developed. It is intended to conduct an analysis of local and global connections in the future.

Footnotes

Abbreviations

Author’s Bio

Malige Gangappa working as Associate Professor in the Department of Computer Science, VNR VJ IET. He has 20 years of teaching, and 10 years of Research experience. He has published many papers in various reputed National and International journals. He was awarded the Elite

+

Gold medal for his superb score in NPTEL. He also worked in various administrative positions in VNR VJIET. He is a distinct division holder throughout his academics. His areas of Interest are Data mining and Soft computing.

References

Elagouz

M.H.

Abou-Shleel

S.M.

Belal

A.A.

and El-Mohandes

M.A.O.

, Detection of land use/cover change in Egyptian Nile Delta using remote sensing, The Egyptian Journal of Remote Sensing and Space Science 23 (2020), 57–62.

Mishra

P.K.

Rai

and Rai

S.C.

, Land use and land cover change detection using geospatial techniques in the Sikkim Himalaya, India, The Egyptian Journal of Remote Sensing and Space Science, 2019.

Xing

Sieber

and Caelli

, A scale-invariant change detection method for land use/cover change research, ISPRS Journal of Photogrammetry and Remote Sensing 141 (2018), 252–264.

Xing

Chen

Zhang

and Liu

, A service relation model for web-based land cover change detection, ISPRS Journal of Photogrammetry and Remote Sensing 132 (2017), 20–32.

Jin

Yang

Zhu

and Homer

, A land cover change detection and classification protocol for updating Alaska NLCD 2001 to 2011, Remote Sensing of Environment 195 (2017), 44–55.

Chen

Tang

Rao

and Wu

, Land cover change detection by integrating object-based data blending model of Landsat and MODIS, Remote Sensing of Environment 184 (2016), 374–386.

Gil-Yepes

J.L.

Ruiz

L.A.

Recio

J.A.

Balaguer-Beser

Á.

and Hermosilla

, Description and validation of a new set of object-based temporal geostatistical features for land-use/land-cover change detection, ISPRS Journal of Photogrammetry and Remote Sensing 121 (2016), 77–91.

Amici

Marcantonio

Porta

N.L.

and Rocchini

, A multi-temporal approach in MaxEnt modeling: A new frontier for land use/land cover change detection, Ecological Informatics 40 (2017), 40–49.

DeWitt

J.D.

Chirico

P.G.

Bergstresser

S.E.

and Warner

T.A.

, Multi-scale 46-year remote sensing change detection of diamond mining and land cover in a conflict and post-conflict setting, Remote Sensing Applications: Society and Environment 8 (2017), 126–139.

10.

Haque

M.I.

and Basak

, Land cover change detection using GIS and remote sensing techniques: A spatiotemporal study on Tanguar Haor, Sunamganj, Bangladesh, The Egyptian Journal of Remote Sensing and Space Science 20 (2017), 251–263.

11.

Deng

Zhu

and Tang

, Land use/land cover classification using time series Landsat 8 images in a heavily urbanized area, Advances in Space Research 63 (2019), 2144–2154.

12.

Tong

X.Y.

Xia

G.S.

Shen

and Zhang

, Land-cover classification with high-resolution remote sensing images using transferable deep models, Remote Sensing of Environment 237 (2020).

13.

Renza

Martinez

Molina

and Ballesteros L

D.M.

, Unsupervised change detection in a particular vegetation land cover type using spectral angle mapper, Advances in Space Research 59 (2017), 2019–2031.

14.

Yeh

A.G.

and Zhang

, A three-component method for timely detection of land cover changes using polarimetric SAR images, ISPRS Journal of Photogrammetry and Remote Sensing 107 (2015), 3–21.

15.

Pan

Guan

Chen

and Li

, Land-cover classification of multispectral LiDAR data using CNN with optimized hyper-parameters, ISPRS Journal of Photogrammetry and Remote Sensing 166 (2020), 241–254.

16.

Lossou

Prempeh

N.O.

and Agyemang

, Monitoring Land Cover changes in the tropical high forests using multi-temporal remote sensing and spatial analysis techniques, Remote Sensing Applications: Society and Environment 16 (2019).

17.

Fan

K.C.

and Hung

T.Y.

, A novel local pattern descriptor – local vector pattern in high-order derivative space for face recognition, Ieee Transactions On Image Processing, 23 (2014), 2877–2889.

18.

Suresh

and Lal

, A metaheuristic framework based automated Spatial-Spectral graph for land cover classification from multispectral and hyperspectral satellite images, Infrared Physics & Technology 105 (2020).

19.

Chen

Ming

and Lv

, Superpixelbased land cover classification of VHR satellite image combining multi-scale CNN and scale parameter estimation, Earth Sci Inform 12 (2019), 341–363.

20.

Kareem

R.S.A.

Ramanjineyulu

A.G.

Rajan

Setiawan

Sharma

D.K.

Gupta

M.K.

and Sengan

, Multilabel land cover aerial image classification using convolutional neural networks, Arab J Geosci 14 (2021), 1681.

21.

Xing

and Yu

, Novel fuzzy uncertainty modeling for land cover classification based on clustering analysis, Sci. China Earth Sci. 62 (2019), 438–450.

22.

Mahmoudi

F.T.

Arabsaeedi

and Alavipanah

S.K.

, Feature-Level Fusion of Landsat 8 Data and SAR Texture Images for Urban Land Cover Classification, J Indian Soc Remote Sens 47 (2019), 479–485.

23.

Unnikrishnan

Sowmya

and Soman

K.P.

, Deep learning architectures for land cover classification using red and near-infrared satellite images, Multimed Tools Appl 78 (2019), 18379–18394.

24.

Mishra

V.N.

Prasad

Rai

P.K.

Vishwakarma

A.K.

and Arora

, Performance evaluation of textural features in improving land use/land cover classification accuracy of heterogeneous landscape using multi-sensor remote sensing data, Earth Sci Inform 12 (2019), 71–86.

25.

and Wang

, Multi-spectral remote sensing land-cover classification based on deep learning methods, J Supercomput 77 (2021), 2829–2843.

26.

Jayanth

Shalini

V.S.

Kumar

T.A.

and Koliwad

, Land-use/land-cover classification using elephant herding algorithm, J Indian Soc Remote Sens 47 (2019), 223–232.

27.

Ojala

Pietikäinen

and Mäenpää

, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Trans. Pattern Anal. Mach. Intell. 24 (2002), 971–987.

28.

Ahonen

Hadid

and Pietikäinen

, Face description with local binary patterns: Application to face recognition, IEEE Trans. Pattern Anal. Mach. Intell. 28 (2006), 2037–2041.

29.

Arabi

P.M.

Joshi

and Deepa

N.V.

, Performance evaluation of GLCM and pixel intensity matrix for skin texture analysis, Perspectives in Science 8 (2016), 203–206.

30.

Zhou

Lin

Zhang

Shao

and Liu

, Improved tracker combined with bidirectional long short-term memory for 3D gaze estimation using appearance cues, Neurocomputing, 2019.

31.

Wang

H.Z.

Wang

G.B.

G.Q.

Peng

J.C.

and Liu

Y.T.

, Deep belief network based deterministic and probabilistic wind speed forecasting approach, Applied Energy 182 (2016), 80–93.

32.

Zheng

Y.J.

, Waterwave optimization: A new nature-inspired metaheuristic, Computers & Operations Research 55 (2015), 1–11.

33.

Rajakumar

B.R.

, Impact of static and adaptive mutation techniques on genetic algorithm, International Journal of Hybrid Intelligent Systems 10 (2013), 11–22.

34.

Rajakumar

B.R.

, Static and adaptive mutation techniques for genetic algorithm: A systematic comparative analysis, International Journal of Computational Science and Engineering 8 (2013), 180–193.

35.

Swamy

S.M.

Rajakumar

B.R.

and Valarmathi

I.R.

, Design of Hybrid Wind and Photovoltaic Power System using Opposition-based Genetic Algorithm with Cauchy Mutation, in: IET Chennai Fourth International Conference on Sustainable Energy and Intelligent Systems, 2013.

36.

George

and Rajakumar

B.R.

, APOGA: An Adaptive Population Pool Size based Genetic Algorithm, in: AASRI Procedia – 2013 AASRI Conference on Intelligent Systems and Control (ISC 2013), Vol. 4, 2013, pp. 288–296.

37.

Rajakumar

B.R.

and George

, A New Adaptive Mutation Technique for Genetic Algorithm, in: Proceedings of IEEE International Conference on Computational Intelligence and Computing Research (ICCIC), 2012, pp. 1–7.

38.

Wagh

M.B.

and Gomathi

, Improved GWO-CS Algorithm-Based Optimal Routing Strategy in VANET, Journal of Networking and Communication Systems 2 (2019), 34–42.

39.

Halbhavi

B.S.

Kodad

S.F.

Ambekar

S.K.

and Manjunath

, Enhanced invasive weed optimization algorithm with chaos theory for weightage based combined economic emission dispatch, Journal of Computational Mechanics, Power System and Control 2 (2019), 19–27.

40.

Jadhav

A.N.

and Gomathi

, DIGWO: Hybridization of dragonfly algorithm with improved grey wolf optimization algorithm for data clustering, Multimedia Research 2 (2019), 1–11.

41.

https://www.google.com/url?q=https://earthexplorer.usgs.gov/&sa=D&source=hangouts&ust=1636716891428000&usg=AOvVaw0t0mpOsrzRl0brKRT58Pqk.

42.

Moosavi

S.H.S.

and Bardsiri

V.K.

, Poor and rich optimization algorithm: A new human-based and multi populations algorithm, Engineering Applications of Artificial Intelligence 86 (2019), 165–181.

43.

Sharma

Hazrati

and Bansal

J.C.

, Spider Monkey Optimization Algorithm, 2019.

44.

Masadeh

Mahafzah

B.A.

and Sharieh

, Sea lion optimization algorithm, International Journal of Advanced Computer Science and Applications 10 (2019), 388–395.

45.

Azari

S.M.

Haddad

O.B.

and Chu

, Shark Smell Optimization (SSO) Algorithm. In: Bozorg-Haddad O. (eds) Advanced Optimization by Nature-Inspired Algorithms, Studies in Computational Intelligence 720 (2018).

46.

Wang

Kuen

Shahroudy

Shuai

Liu

Wang

Cai

and Chen

, Recent advances in convolutional neural networks, Pattern Recognition 77 (2018), 354–377.

47.

Serrano

F.J.G.

Vázquez

Á.N.

and Martín

A.A.

, Training support vector machines with privacy-protected data, Pattern Recognition 72 (2017), 93–107.

48.

Kao

L.J.

and Chiu

C.C.

, Application of integrated recurrent neural network with multivariate adaptive regression splines on SPC-EPC process, Journal of Manufacturing Systems 57 (2020), 109–118.

49.

Gangappa

Mai

C.K.

and Sammulal

, Classification of land cover images using modified water wave Optimization-based hybrid classifier, International Journal of Computers and Applications 43(10) (2021), 1054–1064.

50.

Mukherjee

and Singh

, Assessing land use-land cover change and its impact on land surface temperature using LANDSAT data: A comparison of two urban areas in India, Earth Systems and Environment 4 (2020), 385–407.

51.

Hong

Yao

Chanussot

and Zhu

X.X.

, Multimodal remote sensing benchmark datasets for land cover classification with a shared and specific feature learning model, ISPRS Journal of Photogrammetry and Remote Sensing 178 (2021), 68–80.

52.

Wolek

Suchanek

and Czuba

, Factors influencing walking trips, Evidence from Gdynia, Poland. Plos One 16 (2021), e0254949.

53.

Sahaaya Arul Mary

S.A.

Kanthan Narayanan

Mohana

Senthamil Selvi

Karthik

and Ramya

, Geographical Information System-Aided Landmark Recognition System Using Machine Learning, In Computer Networks and Inventive Communication Technologies, 2023, 257–268.

54.

Gangappa

Kiran Mai

and Sammulal

, Enhanced Crow Search Optimization Algorithm and Hybrid NN-CNN Classifiers for Classification of Land Cover Images, Multimedia Research 2 (2019), 12–22.

55.

Jegatheeswari

and Angelin Deepa

, Fuzzy weighted least square filter for pansharpening in satellite images, Multimedia Research 2 (2019), 17–22.

56.

Thomas

, Adaptive filter using improved pigeon inspired optimization algorithm for satellite image denoising, Multimedia Research 3 (2020).

Feature level fusion for land cover classification with landsat images: A hybrid classification model

Abstract

Keywords

1. Introduction

Table 1 Review of conventional land cover classification schemes

4.1 Spatial features

Local Binary Pattern

Gray Level Co-Occurrence Matrices

ND Vegetation Index (NDVI)

Modified Normalized Difference Water Index (MNDWI)

Normalized Difference Water Index (NDWI)

5.1 Optimized long short-term memory classifier

Solution Encoding

Propagation

Refraction

Breaking

7.1 Simulation set up

Table 3 Statistical analysis

Footnotes

Abbreviations

Author’s Bio

References

Table 1
Review of conventional land cover classification schemes

Table 3
Statistical analysis