Supervised terrestrial to airborne laser scanner model calibration for 3D individual-tree attribute mapping using deep neural networks

Abstract

Capturing subcanopy forest information from airborne laser scanning (ALS) is constrained by signal occlusion. This study demonstrates the potential of close-range terrestrial laser scanner (TLS) scanning to mitigate the constraints of ALS in acquiring stem-level forest attributes. A transformer-based neural network was adapted to classify and segment 3D individual trees from ALS data. A deep neural network combined with a gaussian process layer was proposed to estimate tree diameter-at-breast-height (DBH) from ALS data. The performance of these methods was compared to other benchmarked methods using the same dataset, including a total of seven classifiers, five segmentors, and six attribute regressors. The study was conducted across four ALS sample areas and ten combined TLS/ALS plots, primarily in montane forests. Manual delineation of TLS trees provided a precise validation reference. The proposed methods demonstrated high accuracies, with a mean intersection-over-union (mIoU) of 0.92 for ALS tree classification, 0.70 for tree segmentation, and a RMSE of 4.2 cm or 18.9 % for DBH estimation on average of the ten plots. Tree detection accuracy was strongly associated with the final segmentation accuracy. Factors such as tree height, overlapping, inclination, and neighboring conditions impacted segmentation accuracy. Our segmentation method effectively mitigated accuracy loss for short and occluded trees. Overall, this study presents scalable and cost-effective solutions for TLS calibration of ALS scans over two meso-scale montane valleys. Leveraging deep neural networks enables scaling of stem attributes to landscape scales, thereby linking fine-scale forest inventory with sustainable management of expansive forest resources. Our codes are available at https://github.com/truebelief/artemis_treescaling.