Individual tree crown extraction of natural elm in UAV RGB imagery via an efficient two-stage instance segmentation model

The advancement of near-ground remote sensing and artificial intelligence techniques has revolutionized field surveys, replacing traditional manual methods. Nevertheless, understanding and exploring the growth patterns and intricate morphology of natural elm tree crowns present significant challenges, especially when attempting to extract their features, which are often susceptible to interference from surrounding grass and vegetation. In addition, existing detection and segmentation models based on convolutional neural networks exhibit redundancies in their network architectures and employ less efficient algorithms, such as mask region-based convolutional neural networks. As a result, these models may not be the most suitable options for analyzing extensive and highly detailed remote-sensing image data. We focus on detecting trees in semi-arid regions and extracting their canopy parameters, such as canopy width and area. A training set is established by outlining a total of 20,594 tree canopies on high-spatial resolution unmanned aerial vehicle images. A two-stage instance segmentation model is proposed to develop a method for individual tree detection and efficient extraction of canopy parameters in complex natural environments. The results demonstrate the method's capability to accurately detect the location, number, and canopy parameters (e.g., crown width and area) of individual trees in diverse natural scenes. The model achieves a detection speed of 13.3 fps@1024, with the model weight parameters totaling 8.08 M and computation requiring 8.96 Giga floating point operations per seconds (GFLOPs). Moreover, the detection accuracy and segmentation accuracy of individual trees on the validation set are reported as 0.463 and 0.465, respectively. Compared with Mack RCNN and Mask Scoring RCNN, the proposed method reduces the weight parameters and computational complexity of the model by 82.4%, 83.5% and 96.8%, 92.8%, respectively, while increasing the inference speed by 47.4% and 26.3%. This method offers an efficient and accurate solution for obtaining the structural parameters of individual trees.<br />(c) 2023 Society of Photo-Optical Instrumentation Engineers (SPIE)