Building extraction method for aerial images based on DeepLabv3+ semantic segmentation

Abstract

Aerial imagery can provide rich geographic information. As an important ground object information, quickly and accurately extracting buildings from aerial images can achieve target monitoring, location positioning, and further enrich specific geographic information in a given area. To address the issues of segmentation result merging and irregular contour lines in semantic segmentation algorithms for building extraction, an improved model based on DeepLabv3+ for aerial building extraction is proposed by improving the feature fusion structure, constructing a comprehensive loss function, and incorporating an improved Douglas Peucker algorithm. Experimental results show that the improved model achieves an IoU of 0.794 on the test set, a 14.7% improvement compared to the original model. It effectively avoids the problem of merged segmentation between neighboring buildings and results in more regular segmentation boundaries, enabling more accurate extraction of the building contours.