High-Performance Lightweight Vision Models for Land Cover Classification with Coresets and Compression

Land cover classification from satellite imagery is critical for environmental monitoring, agriculture, and urban planning. However, deploying deep learning models in real-world remote sensing platforms often faces stringent computational and memory constraints. We present a unified framework that integrates lightweight vision backbones with coreset selection and adaptive model compression to address these challenges. Evaluated on the EuroSAT and UC Merced Land Use datasets, our approach leverages four compact architectures: ConvNeXt-Tiny, Swin-Tiny, EfficientNetV2-S, and RegNetY-3.2GF, combined with three coreset strategies (random, forgetting-based, and margin-based) and both fixed and adaptive pruning and quantization. Experiments show that using just 10% of the training data and applying compression can reduce model size by up to 6x while retaining over 92% of baseline accuracy. These results highlight the potential of our method for enabling efficient, accurate land cover classification in edge-deployable remote sensing applications.