*2.3. Methodology*

2.3.1. Method for the LULC Classification

The SVM is a binary classifier [52]; it may classify many characteristics (over two) using either the one-versus-one (OVO) or one-versus-rest (OVR) techniques. The SVM, a statistical learning approach, provides improved overfitting control and a high convergence rate and is unaffected by the local minima [53]. Many classifications issues have been successfully solved using the SVM classifier [54–56]. In this work, we did LULC classification based on three different bands of Landsat satellite data: red (R), green (G), and blue (B). We divided the pixels in the images into five categories (agricultural land, barren land, vegetation, built-up area, and water body) (Table 2).

**Table 2.** Description of the LULC classes identified in the study area.


As a result, a multiclass SVM classifier based on the RGB band composite was developed to predict the land use classes. To design and test the SVM classifier for classification into five groups, 360 datasets were collected. For training and testing the model, we partitioned the total datasets into two halves in the ratio of 80:20. In other words, 212 datasets were used to train the model, and 56 datasets were used to test it. The training dataset comprised 8619 total pixels. Out of the total pixels, we collected 1306 pixels from water bodies, 2500 from built-up, 1613 from vegetation, 2600 from agricultural land, and 600 from barren land. The SVM model classifies a typical pixel in an image as +1 if it corresponds to a specified class or −1 if it belongs to a different class [56]. This work used the OVO approach with radial basis function kernel to create a multiclass classification [57]. The SVM model was trained using training data from two different classes in the OVO approach.

The training of a new SVM model is performed until each class pair has been completed. In this scenario, picture pixels are classified into five classes, resulting in 10 [=(52 − 5)/2]

combinations. Therefore, 10 SVM models are needed to train for classification into five classes. The class of a particular pixel has been predicted using the voting scheme. A typical pixel is projected as "+1" based on the most votes from all ten classifiers. Any SVM model's principal goal is to find the best separating hyperplane, w.x + b = 0, for categorising pixels into two groups. The optimal hyperplane distinguishes the pixels into +1 and −1 with the maximum margin. We used C-support vector classification to run the model. The SVM was executed with the parameters of: the penalty parameter (C) of 1, Nu of 0.5, *p* of 0.5, radial basis function-based kernel, coefficient 0 of 1, degree of 0.5, and gamma of 1. Based on these optimised parameters, the LULC was classified three times, for 2001, 2011, and 2021.
