Dual-Path Tree Crown Detection on Remote Sensing Images

Thanks to the rapid development of remote sensing technology, neural networks have been proven efficient and effective in providing decision-making support statistics for forestry and grassland departments regarding large geographical areas of interest. Tree crown detection is a typical and challenging task, due to the complexity of forest environments, including tree crown overlapping, growth stages and height difference, and lighting conditions and shadows. This paper proposes a dual-path feature fusion architecture for DOM (Digital Orthophoto Map) and DSM (Digital Surface Model) multimodal data, towards an enhanced tree crown detection. Specifically, we have designed a bimodal feature fusion module B-HOIM (Bimodal High-Order Interaction Module) for DOM and DSM, and developed a corresponding framework for practical applications. We first constructed a custom dataset for the tree crown detection dataset, including registered and labeled DOM and DSM. The proposed method achieved metric values of 0.6962, 0.6444, and 0.6626 on precision, recall, and AP, respectively. Compared with the single-modal detection method, the proposed method achieves an improvement of 3.49%, 0.46% and 1.89%.