Multi-Resolution Region Merging (MRRM): A Novel Post-Processing Framework to Mitigate Fragmentation in SAM-Based Geospatial Segmentation of Rural Buildings

Accurate building footprint extraction from high-resolution imagery is essential for rural infrastructure planning, disaster response, and supporting United Nations Sustainable Development Goals. However, applying state-of-the-art foundation models like the Segment Anything Model (SAM) to rural contexts produces severely fragmented outputs, with single buildings often split into multiple segments, rendering them unsuitable for operational mapping. This study introduces the Multi-Resolution Region Merging (MRRM) framework, a novel training-free post-processing approach that transforms fragmented SAM outputs into cartographically accurate building footprints. MRRM uniquely leverages complementary information from two spatial resolutions—detailed 10 cm segmentation for geometric precision and generalized 25 cm boundaries for object coherence—through a systematic region merging strategy. Systematic parameter optimization across Indonesian rural settlements identified optimal configurations (60 cm buffer distance, 0.1 m simplification tolerance) that achieved consistently high spatial accuracy (mean IoU > 0.83). Stratified evaluation across nine morphological categories demonstrated substantial fragmentation reduction (up to 50.2%) while improving segmentation accuracy (ΔIoU up to 0.0478), particularly in dense and semi-regular building configurations. Comparative analysis against four established post-processing methods revealed MRRM's superior performance, achieving the highest IoU (95.9%) and lowest shape distortion (SCD = 1.5%) while maintaining computational efficiency. The framework's training-free nature and compatibility with existing GIS workflows make it immediately deployable for large-scale rural mapping initiatives, particularly in data-scarce regions where traditional supervised methods are infeasible.