Using deep learning to create an enhanced basemap from RGB satellite images for urban planning

Precise segmentation of high-resolution satellite imagery is critical for a wide range of geospatial applications, including urban planning, environmental monitoring, and renewable energy infrastructure optimization. This study explores the use of the Segformer deep learning model, a Transformer-based architecture, to create a detailed basemap from high-resolution RGB satellite images. The basemap includes 13 land cover classes, such as roads, buildings, vegetation, trees, bare land, and other key features essential for geospatial analysis. To evaluate the performance of Segformer, we compare it with three well-established segmentation models: U-Net, DeepLabV3, and PSPNet. Experimental results show that Segformer significantly outperforms the other models in terms of accuracy. Specifically, SegFormer achieves an overall accuracy of 97%, with a mean IoU of 89.8% and an F1 score of 93.8%, clearly outperforming the other models. In comparison, U-Net reached 88% accuracy (71.7% mIoU, 83.5% F1), DeepLabV3 achieved 86% accuracy (54.2% mIoU, 70.3% F1), and PSPNet obtained 81% accuracy (64.9% mIoU, 78.7% F1). These results suggest that Segformer’s ability to capture complex spatial relationships at multiple scales makes it particularly well-suited for high-resolution satellite image segmentation. This study demonstrates the effectiveness of Segformer in generating precise and detailed basemaps, with potential applications in urban infrastructure management, natural ecosystem monitoring, and solar energy planning. These findings open up promising opportunities for using deep learning in large-scale geospatial analysis and mapping. Beyond technical performance, the ability to generate detailed and accurate basemaps from accessible RGB imagery provides a practical pathway to support sustainable urban planning, resource-efficient governance, and the development of smart cities.