Land Cover Classification using Hyperspectral Imagery: A Deep Learning Approach

Abstract

Objectives: This research aimed to perform proper land cover classification of the Krishnagiri and Dharmapuri districts using hyperspectral imagery. This research also attempted to determine the usefulness of deep learning models in detecting large land cover classes, such as quarry, barren, forest, built-up, and agricultural land, using hyperspectral data. Method: The analysis of the land cover patterns in the study area used hyperspectral images (HSI) taken between 2015 and 2022, which had 230 spectral bands. A total of 3000 hyperspectral images were processed, 2000 of which were utilized for training and 1000 for testing. The images were then preprocessed, segmented into superpixels using superpixel segmentation, and classified. Three deep learning models, a Convolutional Neural Network (CNN), Residual Neural Network 18 (ResNet-18), and Visual Geometry Group-16 (VGG-16), were used as classification models. Standard measures of accuracy, precision, recall, and F-score were used to evaluate the performance and effectiveness of the models. Findings: The results indicate that ResNet-18 outperformed the other models in determining land cover classes using hyperspectral images. The model has shown high classification levels among various types of land, with built-up land having an accuracy level of approximately 95.09, indicating a better ability to extract and classify features than CNN and VGG-16. Novelty: The novelty of this study lies in the combination of hyperspectral image processing with deep learning frameworks to classify regions based on land cover. The relative analysis of CNN, ResNet-18, and VGG-16 demonstrated the usefulness of residual learning networks in enhancing the classification accuracy in hyperspectral remote sensing scenarios. Keywords: Hyperspectral Image (HSI), Land Cover Classification, Deep Learning, Convolutional Neural Network (CNN), ResNet-18, VGG-16, Remote Sensing, Spectral–Spatial Analysis