The project “Development of an effective method for detecting small buildings in satellite imagery” aimed to develop a reliable approach for identifying small building structures (below 10×10 m) in Sentinel-2 data. Traditional segmentation and classification methods, such as U-Net or ResNet, show limited accuracy when building footprints are smaller than a pixel or surrounded by complex background elements like vegetation or shadows. Initial research revealed that a significant improvement could be achieved by enhancing the spatial resolution and task-specific representation of the imagery. Consequently, a novel framework was developed, combining Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) with semantic segmentation (DeepLabV3) through joint cross-training. This architecture converts Sentinel-2 imagery directly into cartographic, map-style representations optimized for building delineation and therefore becomes inherently task-oriented. To effectively train and evaluate the framework, a dedicated dataset, S2-BDOT-PL, was composed by integrating Sentinel-2 images with OpenStreetMap and Poland’s official BDOT10k topographic database. The proposed method demonstrated consistent improvement over baseline models across various metrics. The results have been presented in the paper “Map-Guided Cross-Training for Building Detection” which has been already pre-accepted (pending revision) for publication in IEEE Geoscience and Remote Sensing Letters, a journal with Impact Factor of 4 (2024) and score of 140 points on the Ministerial list.