VrsNet - density map prediction network for individual tree detection and counting from UAV images

Abstract

Individual tree detection and counting in unmanned aerial vehicle (UAV) imagery constitute a vital and practical research field. Vegetation remote sensing captures large-scale trees characterized by complex textures, significant growth variations, and high species similarity within the vegetation, which presents significant challenges for annotation and detection. Existing methods based on bounding boxes have struggled to convey semantics information about tree crowns. This paper proposes a novel deep learning network called VrsNet based on the density map information. The proposed work pioneers the segmentation and counting application by utilizing the semantic information of Gaussian contour. Besides, we sample and create the UAV vegetation remote sensing density dataset TreeFsc for experiments. In quantitative comparison across multiple datasets, the proposed method demonstrates high performance, with a 3.45 increase in MAE and a 4.75 increase in RMSE. Experiments demonstrate superior cross-region, cross-scale, and cross-species target detection capabilities of the proposed approach compared with the existing object detection methods. Our code and dataset are available at: https://github.com/luotiger123/VrsNet/tree/main/VrsNet.

Keywords

• Vegetation mapping
• Remote sensing
• Semi-supervised learning
• Object detection
• Tree segmentation
• Tree counting