A Novel Architecture for building rooftop extraction using remote sensing and deep learning

Abstract

Enhancing the accuracy of building rooftop extraction from UAV and remote sensing imagery is crucial for urban planning, disaster management, 3D modeling, and solar resource assessment. In response to this need, a high-quality, open-source dataset focused on rooftop segmentation in Wuhan's Hongshan District has been developed, presenting over 14,000 annotated images. These images are accurately labeled using the Efficient Interactive Segmentation tool (EISeg) for precise superpixel identification. Moreover, to address challenges related to high acquisition costs and reliance on single data sources, a novel framework is proposed that utilizes open-source, high-resolution Google Earth imagery and UAV data. Furthermore, this framework employs tile segmentation techniques for efficient large-scale data management and leverages an advanced EISeg tool for high-precision annotations. Also, four deep learning models were evaluated for semantic segmentation, including the Asymmetric Neural Network (ANN), DeepLabv3, PP-LiteSeg, and Dual Attention Network (DANet). Consequently, the ANN model achieved the highest accuracy at 96 percent, outperforming DANet at 95.09 percent, PP-LiteSeg at 94.54 percent, and DeepLabv3 at 81.61 percent. Furthermore, an intelligent mosaicking algorithm based on GDAL, combined with post-processing optimization, improved processing efficiency by 3.2 times while preserving image accuracy. This research provides a precise and cost-effective solution for building rooftop detection in complex urban environments, significantly improving the scalability and reliability of remote sensing data processing. These improvements enable more efficient large-scale urban analysis, ultimately supporting critical applications in smart city development, disaster response, and solar energy assessment.