Building Footprint Segmentation Using the Modified YOLOv8 Model

Abstract This paper proposes an enhanced YOLOv8 model specifically designed for precise building footprint segmentation. The model incorporates several key modifica- tions to achieve superior performance and efficiency. Firstly, a novel fusion layer integrates RGB image information with a Digital Elevation Model (DEM), enrich- ing feature representation and facilitating the distinction of building structures. Secondly, Depthwise Separable Convolution (DSConv) replaces standard convo- lutions throughout the backbone and head, leading to a more compact model with faster inference speed. Thirdly, Varifocal Loss (VFL) is employed as the clas- sification loss function, effectively addressing class imbalance issues prevalent in segmentation tasks. Our proposed model demonstrates significant improvements over three DeepLabv3+, SAM, and the original YOLOv8 state-of-the-art models. We achieve a precision of 91.11%, a recall rate of 89.71%, and a mAP (mean Average Precision) of 87.42%, surpassing all compared models in accuracy. Fur- thermore, the proposed model boasts a remarkably fast inference time of only 45.1 milliseconds per image, making it suitable for real-time applications.