Building Extraction Based on Dual-Stream Detail-Concerned Network

Objectives The distribution of buildings is an important indicator to measure regional development. Automatic extraction of building information from remote sensing images is of great significance for urban and rural planning. Most existing methods underestimate building details such as boundaries and tiny buildings.

Methods In order to increase attention to building details, a dual-stream detail-concerned network (DSDCNet) is proposed in an encoder-decoder manner. First, a dual-stream feature extraction module is used to extract semantic features and detail-concerned features. They are fed into the decoder consisting of a series of detail refinement modules where detail-concerned features make up for the missing details of semantic features and the semantic features enhance semantic continuity of detail-concerned features. Then, a semantic-detail fusion module is used to fuse and squeeze two refined features. Furthermore, deep supervision is conducted and the multi-level outputs are used in detail-concerned loss function so as to strengthen the supervision of building details.

Results Five mainstream networks are selected for comparison in WHU dataset, ISPRS Vaihingen dataset and a domestic high-resolution dataset. The evaluation results show that DSDCNet has better performance than other networks, especially in F1-score and intersection over union without introducing too much network complexity.

Conclusions DSDCNet not only manages to improve the overall performance of building extraction results, but also effectively maintains the integrity of building boundaries and reduces the missed detection of small buildings. It has better extraction effect on the buildings with small size and complex context.