Woodland Extraction of SPOT7 Image Based on Multi-scale Attention Mechanism and Edge Constraint

Objectives As woodland is an important natural and economic resource of China, it is important to understand the distribution of woodland for the investigation and management of woodland resources.We design a woodland extraction method combining multi-scale attention mechanism and edge constraint to tackle the issue of low accuracy and irregular boundaries in traditional forest extraction methods.

Methods First, an end-to-end multi-scale attentional neural network model is constructed to fully extract the context features of woodland in remote sensing images, and semantically describe woodland at different scales to achieve high-precision pixel-level expression of woodland. Second, the edge constraint rules are constructed to optimize the boundary of the extraction results, to improve the readability of the extraction results. To prove the effectiveness of the proposed method, Santai County, Mianyang City, Sichuan Province, China is taken as the experimental area to establish datasets and carry out woodland extraction experiments.

Results The results show that the extraction accuracy of this method is 81.9%, the recall rate is 75.6%, F1 score is 78.1%, intersection of union is 64.2%.

Conclusions The propsed method has a better effect in the application of woodland extraction with remote sensing image.