Rooftop Segmentation using Geospatial Artificial Intelligence for Rural India

Abstract. Rooftop identification using satellite imagery is having a wide range of practical use cases, which includes habitation identification, disaster management, infrastructure monitoring and solar panel deployment. Recent advances in GeoAI have improved rooftop extraction by integrating spatial technologies and deep learning models. However, it remains a challenge in rural India due to the irregular shapes of rooftops, variations in surface texture, low resolution open-source satellite data, canopy cover, and diverse geographic conditions. This paper focuses on rooftop extraction from satellite imagery utilizing the YOLOv11l segmentation framework trained on CDAC PARAM Siddhi-AI HPC platform. Seamless labels were created over satellite images using QGIS open-source software. We applied the super-resolution model to improve overall quality, thereby up-scale the resolution from 320×320 to 640×640 pixels. The super-resolution images were tiled into 320x320 for training. Additionally, image processing techniques (contour detection, erosion and dilation) were used to reduce noise and maintain consistency of images. The model trained on the pre–super-resolution dataset achieved a mean Intersection over Union (mIoU) of 36.20% and a precision of 76.42%, whereas the model trained on the post–super-resolution dataset achieved a higher mIoU of 61.15% and a precision of 80.92% on the test dataset. It indicates that the super-resolution model performed better prediction. The model is a deployable GeoAI solution for national schemes like PMGSY National GIS and PM Surya Ghar.