LAND USE AND LAND COVER CLASSIFICATIONUSING SATELLITE IMAGERY AND DEEP LEARNING

Abstract

The precise and effective monitoring of land use and land cover (LULC) is important for urban planning, environmental protection, agriculture, and disaster management. Current approaches for classifying satellite images are dependent on the manual inspection process, which is time-consuming and can cause errors. Hence, a deep learning and computer vision-based solution for efficient LULC classification has been presented in this dissertation. For the purpose of developing our model, we used the EuroSat data set. It is a publicly available data set generated from the images collected by the Sentinel-2 satellite, which is owned and operated by the European Space Agency. Our deep convolutional neural network (CNN), inspired by the LeNet-5 architecture, is intended to classify satellite images of size 64x64 into ten classes such as Annual crop, Forest, Highway, Residential area, and Water bodies. For implementation, we used PyTorch, the popular deep learning framework, in conjunction with a GPU and CUDA. As a part of this dissertation, we have also considered the use of transfer learning. For that, we froze all but the last layer of a ResNet-18 CNN trained on the Imagenet database. It was shown through experiments that the proposed lightweight custom CNN model was able to achieve a validation accuracy of 66.9% in just five training iterations while taking only 2.44 minutes to train on the GPU altogether. The findings show that although deep learning models present themselves as a more reliable option compared to manual classification techniques, the similarity of spectra among vegetation classes cannot be ignored as the source of any possible mistakes. The findings suggest that small CNN architectures are an excellent choice for achieving real-time environmental monitoring in a more effective manner, and transfer learning with ResNet-18 represents a way forward to obtaining better quality. In total, these two concepts present a great platform for making further improvements using the multispectral analysis method and pixel-wise semantic segmentation techniques. (Keywords: Land Use and Land Cover (LULC), Convolutional Neural Network (CNN), EuroSAT, Transfer Learning, ResNet-18, PyTorch, Remote Sensing, Sentinel-2, Deep Learning, Image Classification.)