FO-Net: An advanced deep learning network for individual tree identification using UAV high-resolution images

Abstract

The identification of individual trees can reveal the competitive and symbiotic relationships among trees within forest stands, which is fundamental understand biodiversity and forest ecosystems. Highly precise identification of individual trees can significantly improve the efficiency of forest resource inventory, and is valuable for biomass measurement and forest carbon storage assessment. In previous studies through deep learning approaches for identifying individual tree, feature extraction is usually difficult to adapt to the variation of tree crown architecture, and the loss of feature information in the multi-scale fusion process is also a marked challenge for extracting trees by remote sensing images. Based on the one-stage deep learning network structure, this study improves and optimizes the three stages of feature extraction, feature fusion and feature identification in deep learning methods, and constructs a novel feature-oriented individual tree identification network (FO-Net) suitable for UAV high-resolution images. Firstly, an adaptive feature extraction algorithm based on variable position drift convolution was proposed, which improved the feature extraction ability for the individual tree with various crown size and shape in UAV images. Secondly, to enhance the network’s ability to fuse multiscale forest features, a feature fusion algorithm based on the “gather-and-distribute” mechanism is proposed in the feature pyramid network, which realizes the lossless cross-layer transmission of feature map information. Finally, in the stage of individual tree identification, a unified self-attention identification head is introduced to enhanced FO-Net’s perception ability to identify the trees with small crown diameters. FO-Net achieved the best performance in quantitative analysis experiments on self-constructed datasets, with mAP50, F1-score, Precision, and Recall of 90.7%, 0.85, 85.8%, and 82.8%, respectively, realizing a relatively high accuracy for individual tree identification compared to the traditional deep learning methods. The proposed feature extraction and fusion algorithms have improved the accuracy of individual tree identification by 1.1% and 2.7% respectively. The qualitative experiments based on Grad-CAM heat maps also demonstrate that FO-Net can focus more on the contours of an individual tree in high-resolution images, and reduce the influence of background factors during feature extraction and individual tree identification. FO-Net deep learning network improves the accuracy of individual trees identification in UAV high-resolution images without significantly increasing the parameters of the network, which provides a reliable method to support various tasks in fine-scale precision forestry.