Optimized Deep Learning Model for LULC Classification in the Terai Region: Comparing U-Net and Resunet with Spectral Index Fusion and Cross-Area Fine-Tuning

Abstract

Accurate and up-to-date Land Use/Land Cover (LULC) mapping is crucial for environmental management and regional planning, but its development is often limited by insufficient labeled data, spectral similarity between land-cover classes, and difficulties in transferring models across regions. This study developed an optimized deep learning framework for LULC classification in Dhanusha District using cross-area fine-tuning with U-Net and a ResUNet-style architecture based on a ResNet-50 encoder backbone. The approach aimed to improve model transferability by adapting models trained in a data-rich source area (Mahottari District) to a target area (Dhanusha District) with limited labeled samples. Sentinel-2 Level-2A imagery from March 2021 was preprocessed and five LULC classes - Water Bodies, Cropland, Built-Up, Tree Cover, and Others were manually digitized from Google Earth imagery and converted into raster masks. Along with RGB imagery, spectral-index composites including NDVI, SAVI, EVI, MSAVI, GNDVI, NDWI, and NDBI were generated to improve class separability. Image and mask datasets were tiled into sizes of 128×128, 256×256, and 512×512 pixels and divided into training and testing sets. Hyperparameters such as batch size, dropout, L2 regularization, learning-rate scheduling, and early stopping were systematically optimized. Results demonstrated strong segmentation performance after cross-area adaptation. ResNet-50 consistently outperformed U-Net, achieving overall accuracies of up to 0.96 in Mahottari and 0.95 in Dhanusha, while U-Net achieved up to 0.95 and 0.94, respectively. Spectral-index composites reduced confusion among challenging classes, especially water bodies and built-up areas. Intermediate tile sizes (256×256) provided the best balance between spatial context, computational efficiency, and stable model convergence, resulting in improved generalization across districts.