Efficient Application of the Deep Learning-Driven SGS-3DRecon Framework in Large-Scale Forest Surveys Under Complex Scenarios

While LiDAR is widely used in tree measurement, its high cost and complexity remain limiting factors. With advances in photogrammetry and deep learning, efficient and accurate alternatives are gaining importance in forest resource surveys. Therefore, this study proposes a fully automated framework for tree stem volume measurement based on virtual measurement from a single photograph. This framework integrates the Segformer architecture, generative adversarial network inverse mapping technology, and Sobel operator (SGS-3DRecon), enabling multilevel analysis from image segmentation to 3-D reconstruction and measurement. To validate its potential application, we collected images of 5192 individual trees across multiple scenarios (natural forests, afforested area, and urban parks) in Beijing. Additionally, point cloud data of 250 trees were acquired using handheld mobile laser scanning (HMLS) technology for comparative evaluation. Results indicate that SGS-3DRecon achieved intersection over union (IoU) and recall accuracies of 86.40% and 92.61%, respectively, in the combined scenarios. The average accuracy fluctuation between different scenarios was 3.41%, with the highest accuracy observed in the park scenario. Compared to the complex processing required by HMLS, the SGS-3DRecon method achieved an 85.5% improvement in efficiency for volume measurement. While ensuring high efficiency, the average relative root-mean-square error (rRMSE) was 17.91%, and the average rBias was −0.40%. In summary, this framework provides an efficient and low-cost solution for forest resource surveys. It shows significant potential, especially in monitoring tasks with limited resources and large-scale requirements. This innovative approach has the potential to further advance forest resource survey technologies.