Enhancing land use and land cover classification with deep learning-based satellite imagery segmentation

Abstract

Semantic segmentation of satellite imagery plays a vital role in applications such as sustainable development, agriculture, forestry, urban planning, and climate change monitoring. Despite its importance, optimizing deep learning models for this task remains challenging. This study evaluates several advanced deep learning architectures UNet, LinkNet, DeepLabV3+, and a modified version, AE-DeepLabV3+ in combination with different backbone networks, including ResNet101, ResNet152, Xception, MobileNetV2 and, EfficientNetV2. The objective is to identify the most effective model for classifying satellite images into eight land cover categories: built-up areas, roads, water bodies, agricultural land, shrubland, forest, grassland, and others. A high-resolution dataset with corresponding segmentation masks was developed to support this analysis and serve as a resource for future research. Preprocessing steps included normalization and data augmentation techniques such as vertical and horizontal flipping and random brightness adjustments. Experimental results indicate that UNet with the Xception backbone, LinkNet with ResNet152, DeepLabV3+ with the Xception backbone, and AE-DeepLabV3+ with the Xception backbone achieved Dice coefficients of 85.7%, 86.7%, 90.4%, and 91.3%, respectively. Among these, AE-DeepLabV3+ with Xception demonstrated the highest segmentation accuracy. The findings are contextualized through comparison with recent studies, highlighting the model’s ability to generalize across diverse geographic regions. To enhance model transparency and interpretability, explainable AI (XAI) techniques Seg-Grad-CAM++ and Seg-Score-CAM are employed to visualize class-specific feature attributions and better understand the model’s decision-making process.