Structural-Constrained Corner Point Prediction for Building Footprint Extraction

Accurate detection of building boundaries in remote sensing images is essential for urban development, land-use mapping, infrastructure planning, and property assessment. This study introduces an ensemble loss function designed to improve the performance of building segmentation tasks. The proposed loss function incorporates four distinct terms encapsulating critical information about building orientation, geometry, and surrounding built-up areas. Specifically, the loss components include building corner points detection ($L_{\text{cp}}$), building area ($L_{a}$), building angle relative to the ground ($L_{\vartheta }$), and building position within housing blocks ($L_{p}$). The first three terms enhance the accurate delineation of building polygons, while the latter mitigates oversegmentation. To achieve these objectives, the output tensor of the network is modified to include $2n + 3$ neurons including $2n$ neurons for corner point regression ($x_{11}, x_{12}, x_{21},\text{ and }x_{22}$ in our case), 1 neuron for building area, 1 neuron for building angle ($\vartheta$), and one neuron for building polygon position in image. This architectural design enables the network to incorporate these additional spatial and structural constraints, significantly improving the extraction of the building footprint. The proposed loss function is evaluated across three segmentation networks DeepLabv3, SegNet, RapNet, and DSANet as well as a detection network, YOLO. Thorough experiments conducted on benchmark datasets, such as Inria, WHU, Massachusetts and South Asia, validate the effectiveness of the proposed loss function. Experimental results indicate that the proposed loss function can be seamlessly integrated into existing segmentation networks, outperforming state-of-the-art loss functions both qualitatively and quantitatively. Notably, the DSANet combined with the proposed loss achieves a mean intersection over union (mIoU) of 80.10% with increase of 0.70% from SOTA networks on Inria. The DSANet achieves mIoU of 93.01% with increase of 1.7% on the WHU datasets. Similarly, DSANet with the proposed loss achieves an mIoU of 74.71% with increase of 4.14% from SOTA networks on the Massachusetts dataset.