Toward optimized deep learning-based land use/land cover classification using Landsat 8: a systematic multi-condition experiment

High-fidelity land use/land cover (LULC) maps are fundamental for monitoring ecosystem dynamics and supporting sustainable environmental planning. While Landsat imagery supports long-term ecological analysis, its medium resolution presents significant informatics challenges, such as mixed pixels and severe class imbalances among natural habitats. This study proposes a systematic informatics framework for generating large-scale training datasets by integrating public land-cover products with Landsat 8 multispectral imagery. We evaluated 252 experimental combinations of cloud-cover thresholds, spectral bands, and loss functions across U-Net, DeepLabv3+, and Swin-UPerNet architectures. U-Net achieved the highest overall performance, particularly excelling in dominant classes like forest and farmland. Conversely, Swin-UPerNet and Focal Tversky Loss (FTL) provided a critical relative advantage in identifying “hard-to-classify” natural classes—grassland and wetland—which are inherently difficult to distinguish from dominant classes due to spectral overlap and visual ambiguity. The FTL further enhanced minority class F1-scores by up to 10%. While the RGB + NIR + SWIR1/2 combination was optimal, configurations utilizing cloud thresholds ≥20% proved that increased data volume (> 10,000 samples) effectively offsets moderate noise. Ultimately, these findings suggest that a hybrid architecture—integrating Swin-Transformer's global context modeling with U-Net's progressive upsampling—represents the most promising informatics pathway for next-generation, medium-resolution ecological monitoring.