MineC2FNet: Coarse-to-Fine Learning for Mining Footprint Segmentation

Researchers propose MineC2FNet, a coarse-to-fine domain incremental learning framework that leverages abundant coarse boundary data to improve fine-grained segmentation of mining footprints in multispectral imagery. The framework uses a teacher-student architecture with attentive distillation at feature and prediction levels, selectively transferring knowledge from coarse to fine domains while enabling boundary refinement with limited fine-grained data. An expertly validated dataset of 219 images is introduced. The work addresses the scarcity of fine-grained annotated data for monitoring socio-environmental impacts of mining.