Deep Learning-Guided High-Completeness Building Segmentation Sample Selection via Otsu Thresholding

Deep learning (DL) has significantly accelerated remote sensing (RS) imagery interpretation. However, annotation errors, such as omission, misalignment, and geometric deformation, can severely degrade model performance. Among these, omissions—where objects in the imagery are partially or entirely missing from their corresponding labels—are particularly detrimental. To address this issue, this letter proposes a method for selecting high-completeness training samples using a model trained on a building segmentation dataset with omission errors. Utilizing Otsu’s thresholding characteristic for foreground–background discrimination, the method estimates the reference completeness of each sample and identifies those with high completeness based on a given threshold. Extensive experiments are conducted on both a simulated dataset with omission errors and an OpenStreetMap (OSM)-derived building segmentation dataset. The results demonstrate that the proposed method effectively identifies high-completeness samples. Furthermore, models trained on these selected samples exhibit significantly improved performance compared with those trained on the original, unfiltered datasets.