Deep Learning-Based Detection and Severity Classification of Individual Trees Damaged by Dendrolimus sibiricus Using UAV Multispectral Imagery and LiDAR Data

Outbreaks of the Siberian silk moth (Dendrolimus sibiricus) pose an increasingly serious threat to larch forests in Mongolia, particularly as rising air temperatures and increasing drought frequency create favorable conditions for pest proliferation, degrading forest ecosystem stability. This study presents a UAV-based deep learning (DL) framework that integrates multispectral (MS) and LiDAR sensor data to detect individual trees and classify their health condition in a Dendrolimus sibiricus-affected larch forest. Field surveys were conducted over a 59-hectare study area using UAV flights during the pre-pesticide application period. Vegetation indices (NDVI, GNDVI, NDRE and CIRE) combined with tree structural point cloud data were employed to train a Mask R-CNN DL model to classify tree health into 3 classes: (i) healthy, (ii) moderate and (iii) severely damaged. The model was evaluated on 10 held-out test plots, containing 573 trees. Detection precision, recall and F1-score were 0.679, 0.684 and 0.682, while Total (detection plus correct health class) precision, recall and F1-score were 0.627, 0.632 and 0.630, respectively. Compared with spectral-only input (RGB plus vegetation indices), the full 10-channel input improved the Detection F1 from 0.675 to 0.682 and the Total F1 from 0.619 to 0.630. The findings contribute to forest pest early warning systems, hotspot identification and forest health monitoring in Mongolia.