摘要
Abstract
In the extraction of buildings from high-resolution remote-sensing imagery,the prevailing evaluation framework relies mainly on mean average precision(mAP)and mask intersection over union(Mask IoU),which provides limited insight into boundary accuracy.To address this shortcoming,this paper proposed a boundary intersection over union(BIoU)evaluation protocol tailored to the remote-sensing building extraction domain.The protocol comprised two metrics—BIoU and the 95th-percentile Hausdorff distance(HD95)—and adopted a dual-caliber reporting strategy.Based on the Wuhan Uni-versity(WHU)and Massachusetts building datasets,it systematically compared the performance of eight method configura-tions under the proposed protocol,including variants of the real-time instance-segmentation model(YOLOv8-seg),the U-shaped encoder-decoder network(U-Net),and the deep semantic segmentation network(DeepLabv3+).The results show that,for in-domain scenarios,the intersection over union(IoU)of all methods is comparable(0.843-0.851),whereas the HD95 reveals a paradigm-level gap.In cross-domain scenarios,the you only look once(YOLO)-series detection algorithms retain an IoU of 0.224-0.279,whereas semantic-segmentation models drop to 0.054-0.095,significantly reversing the in-domain ranking.After fine-tuning with only 5%of target-domain samples,the semantic-segmentation models achieve better IoU performance than the zero-shot YOLO.This demonstrates that the proposed BIoU+HD95 protocol effectively exposes the limitations of relying solely on in-domain Mask IoU for evaluation,offering a new tool for method evaluation and selection in remote-sensing building extraction.关键词
建筑物提取/实例分割/边界交并比(BIoU)/评测协议/跨域泛化/95分位豪斯多夫距离(HD95)/实时实例分割模型(YOLOv8-seg)系列变体/U型编解码网络(U-Net)/深度语义分割网络(DeepLabv3+)分类
天文与地球科学
