Abstract:

The advancements in deep learning and computer vision technologies have had a profound impact on the field of air-borne remote sensing， making the analysis of aerial images more efficient. Compared to conventional images， the target boundaries in aerial images are clearer and more distinct， with more regular distributions and stronger spatial structure. However， current state-of-the-art segmentation methods mainly focus on utilizing complex feature extractors to capture stronger contextual relationships， placing more emphasis on single-pixel classification accuracy. This not only demands higher hardware requirements but also overlooks the issue of boundary alignment from a structural perspective. To address this challenge， we propose an innovative boundary-aware loss function， Lossd， designed to enhance the performance of semantic segmentation for aerial remote sensing images， particularly in terms of boundary precision and target segmentation consistency. We innovatively translate structural differences into a loss， unlike traditional methods that focus on single-pixel accuracy. Moreover， we propose an effective solution for the common over-segmentation and under-segmentation problems in semantic segmentation tasks. Extensive experimental validation has been conducted on three widely used large-scale datasets and three benchmark models. Experimental results show that our method significantly improves the semantic segmentation performance without modifying the original network. Specialty， our method achieves 55.8% mIoU （+1.6%） on LoveDA， 70.8% mIoU （+0.8%） on UAVid， and 94.1% mF1 （+0.7%） on Potsdam， reaching or partially surpassing the performance of mainstream approaches.

Key words: aerial images, deep learning, semantic segmentation, loss function, boundary constraints