Map-Guided Cross-Training for Building Detection

Extraction of building footprints from freely available, high temporal resolution satellite imagery is important for applications ranging from disaster response to monitoring informal settlements. While Sentinel-2 is globally available, limited spatial resolution restricts the reliable detection of small or densely packed structures. We propose a novel framework that re-targets a super-resolution backbone (ESRGAN) to generate cartographic map-style outputs directly and fuses it with a semantic segmentation network through joint training. Unlike pipelines treating super-resolution and detection as sequential tasks, our approach leverages map-centric supervision, unifying enhancement and building delineation. For evaluation, we introduce the S2-BDOT-PL dataset, integrating Sentinel-2 imagery with OpenStreetMap and BDOT10k ground truth. Results demonstrate improved precision and robustness in small-building detection compared to native resolution baselines.