Comparative Evaluation of Deep Learning Semantic Segmentation Models for Land Use Land Cover Mapping in Tamil Nadu

Accurate mapping of Land Use and Land Cover (LULC) is important for environmental monitoring, urban planning, agricultural management and sustainable utilisation of resources. Recent breakthroughs in deep learning have shown considerable improvements in the accuracy of semantic segmentation of remote sensing images, allowing the automatic extraction of spatial and contextual elements. The present study is a comparative evaluation of five state-of-the-art deep learning semantic segmentation models namely U-Net, UNet++, DeepLabV3+, SegFormer and Swin-UNet for LULC mapping in Tamil Nadu, India using Sentinel-2 multispectral satellite images. The preprocessing workflow consists of selecting cloud-free images, normalising the images, generating patches, augmenting data, and preparing pixel-wise ground truth masks. To guarantee a fair comparison, all models are trained under the same experimental settings using the same training, validation and testing datasets. Model performance is measured in terms of Overall Accuracy, Precision, Recall, F1-score, Intersection over Union (IoU). The comparative analysis studies the classification accuracy, boundary delineation, segmentation quality, training time, inference time and robustness over diverse land cover classes. The results show that Transformer-based architectures usually yield higher segmentation accuracy by effectively capturing global contextual information, whereas CNN-based models give competitive performance with lower computing costs. The findings shed light on the advantages and disadvantages of both architectures and offer practical advice on selecting acceptable deep learning models for regional LULC mapping.