RoofMapNet: Utilizing geometric primitives for depicting planar building roof structure from high-resolution remote sensing imagery

Abstract

The accurate extraction of building roof structures from aerial imagery represents a fundamental task for urban digital twin systems, facilitating critical applications such as 3D city modeling and solar potential assessment. Despite recent advancements in geospatial artificial intelligence, existing methods frequently encounter challenges posed by real-world complexities. These include structural heterogeneity caused by diverse architectural styles, discontinuities in roof structures due to occlusions from vegetation and other obstacles, and the limited generalization ability of models stemming from the scarcity of specialized annotated datasets. In this paper, we introduce an end-to-end network called RoofMapNet, specifically designed for extracting roof structures. First, we propose a strategy for roof junction extraction that integrates dynamic Gaussian heatmaps with quadratic coordinate calibration. This strategy enhances the model’s robustness in junction prediction under heterogeneous sample distribution scenarios. To address the loss or blurring of roof lines caused by occlusion and shadow, we propose an adaptive occlusion-aware module. This module employs a bidirectional mapping between geometric and feature spaces to refine candidate lines accurately, thus improving the model’s generalization ability and robustness in roof line detection. Additionally, to comprehensively evaluate the performance of roof structure detection models, we meticulously annotated a diverse, large-scale remote sensing imagery dataset for roof structure extraction, named RoofMapSet. Comprehensive evaluations on the VWB and RoofMapSet datasets demonstrate state-of-the-art performance, with mean $sAP$ improvements of 4.13% and 2.85% over competitors, respectively. Further analyses confirm the resilience to varying spatial resolutions and complex occlusion patterns. Our code and data are available at: https://github.com/CVEO/RoofMapNet.