Enhanced LULC Classification Using CNNs with Transfer Learning and Fine-Tuning: A Regional Study

Abstract

In recent days, due to the high population and rapid urbanization, we have challenged several problems related to environmental degradation and climate change. Therefore, Land Use Land Cover (LULC) classification is important in providing accurate and timely information about natural and land resources. Traditional methods for the classification of satellite imagery face several challenges due to the complexities and variability of the data. In this paper, we proposed a novel approach to enhance LULC classification using deep learning-based convolutional neural networks with extraction of features, transfer learning, and fine-tuning. The proposed work first designs convolutional neural networks from scratch to capture spatial features from multispectral resolution satellite imagery covering the study area of Mysuru taluk, Karnataka State, India. Transfer learning is then applied to adapt the pre-trained CNN model to the LULC classification. Furthermore, fine-tuning is employed to fine-tune the adapted CNN model on the target dataset, enabling the model to learn domain-specific features and improve classification performance. The proposed deep learning model performance is demonstrated through experiments on multispectral datasets, where convolutional neural networks, transfer learning, and fine-tuning models provide classification accuracy of 90.41%, 92.50%, and 94.37%, respectively.