Resource-Efficient Adaptation of the Segment Anything Model for Urban Land-Cover Mapping

High-resolution land-cover mapping from remote sensing imagery is essential for urban monitoring, environmental assessment, and sustainable planning. However, adapting foundation models for semantic segmentation typically requires substantial computational resources, limiting their practical adoption. This study presents a parameter-efficient adaptation framework based on the Segment Anything Model (SAM) that enables effective semantic segmentation of remote sensing imagery on resource-constrained hardware. The proposed framework combines selective fine-tuning with Low-Rank Adaptation (LORA) and an imbalance-aware training strategy, reducing the number of trainable parameters to ۳۳ million (۶۵% fewer than full fine-tuning) while maintaining competitive performance. Experiments on the OpenEarthMap benchmark dataset achieve ۵۱.۵۹% mIoU with Test-Time Augmentation, outperforming baseline models under identical training conditions and exceeding the performance reported by other SAM-based approaches on the same benchmark dataset. Resolution ablation experiments further indicate that an input size of ۵۱۲x۵۱۲ provides a balance between segmentation performance and computational cost, making the proposed framework suitable for deployment on limited hardware resources. Beyond semantic segmentation, the resulting land-cover maps enable estimation of Green Space Ratio and Impervious Surface Ratio. These results demonstrate that foundation models can be efficiently adapted for practical remote sensing applications under limited computational resources while supporting reliable urban-scale environmental analysis.

کلیدواژه ها:

Segment Anything Model ، SAM ، OpenEarthMap ، LORA ، Remote Sensing