Transfer and Bayesian Learning Approaches for Land Use and Land Cover Classification

Land use and land cover (LULC) classification is essential for understanding the impact of both human activities and natural processes on the Earth’s surface. This classification plays a critical role in deforestation, urban planning, and damage assessment. However, traditional machine learning techniques often face challenges in achieving consistent accuracy in diverse datasets. In this study, we implement transfer learning and Bayesian learning approaches to enhance the accuracy and robustness of LULC classification. Transfer learning leverages knowledge from previous classification problems, while Bayesian learning addresses uncertainties in the classification process. Using the Remote Sensing Image Classification Benchmark (RSI-CB128) dataset, the study evaluates the performance of Bayesian Convolutional Neural Networks (CNN), Mobile Neural Networks (MobileNet), Inception Neural Networks (InceptionNet), Densely Connected Convolutional Networks (DenseNet) and Efficient Neural Networks (EfficientNet). The study also investigates the impact of varying the Bayesian layers on model performance, finding that a medium number of layers optimally captures model uncertainty. The results indicate that the transfer learning models, particularly MobileNet, achieved the highest accuracy of 97. 48% compared to 90. 33% for Bayesian CNN. These findings suggest that transfer learning techniques are highly effective for LULC classification, providing a reliable method for practical applications in remote sensing.