Research on eucalyptus individual tree segmentation and age estimation utilizing improved Mask R-CNN algorithm based on UAV stereo images

Eucalyptus, as one of the world's three major fast-growing tree species, not only provides timber for various industries but also makes significant contributions to carbon sequestration and mitigating climate change. However, due to the rapid growth and diverse variations in eucalyptus, accurate individual tree segmentation poses certain challenges. Additionally, achieving individual tree segmentation while extracting the age of eucalyptus plantations presents greater challenges to segmentation algorithms. Based on this, this study utilizes the Mask R-CNN algorithm framework. Firstly, Transnext is employed to enhance the backbone's performance in extracting image texture and spatial information. Secondly, CARAFEFPN replaces the original FPN structure to increase the receptive field, thereby improving the model's detection performance in dense eucalyptus plantations. Subsequently, Seesawloss is used as the loss function, and SABLHead serves as the detection head to address sample imbalance and enhance model robustness. Finally, the improved Mask R-CNN algorithm is utilized for individual tree segmentation and stand age extraction in eucalyptus plantations. The results indicate that the improved Mask R-CNN model achieves precise individual tree segmentation in different stages of eucalyptus plantations, achieving MAP50 of 80.6 %, MAP75 of 54.2 %, and MAP0.5:0.95 of 48.9 %, all surpassing the Mask R-CNN model. Based on the improved Mask R-CNN model, the SHAI algorithm enables accurate estimation of stand age in eucalyptus plantations. The results of stand uniformity calculated based on individual tree crown width and tree height show relatively significant differences, with stand uniformity calculated based on individual tree height outperforming that calculated based on individual tree crown width. The research findings demonstrate that the improved Mask R-CNN algorithm can achieve both precise individual tree segmentation in different stages of eucalyptus plantations and accurate estimation of stand age in eucalyptus plantations, providing data support for the evaluation of stand uniformity and productivity in eucalyptus plantations. • Mask R-CNN model was improved using Transnext, CARAFEFPN, SABLHead and Seesawloss. • Achieved 80.6 % MAP50 for eucalyptus individual tree segmentation with improved model. • Stand age of regional eucalyptus plantations was accurately estimated. • Stand uniformity estimated from tree height was better than that of tree crown width.