Early detection of Dendrolimus species infestations: integrating UAV hyperspectral and LiDAR data

TL;DR

This study improves early detection of Dendrolimus pest infestations in pine forests by combining hyperspectral and LiDAR data from drones, achieving higher accuracy than using either method alone.

Contribution

The novel integration of hyperspectral imaging and LiDAR data with optimized band selection algorithms enhances early detection of Dendrolimus infestations.

Findings

Combining hyperspectral and LiDAR data achieved the highest detection accuracy (83.33%) for Dendrolimus infestations.

LiDAR effectively captures spatial changes in needle distribution due to defoliation.

Optimized band selection using ISIC-SPA-RF reduced data redundancy and improved model performance.

Abstract

Dendrolimus species are the major defoliating forest pests in China, causing severe damage to pine forests. Establishing an effective early monitoring system was crucial for timely implementation of control measures to prevent further infestation, significantly reducing economic losses and ecological damage. While previous studies have demonstrated the limited effectiveness of spectral data alone in early detection of Dendrolimus spp. infestations, our research reveals that needle loss is the primary damage symptom, whereas canopy structural characteristics remain underexplored in early monitoring. To address this knowledge gap, this study innovatively integrates unmanned aerial vehicle-based hyperspectral imaging (HSI) with Light Detection and Ranging (LiDAR) data. This study employed SPA, ISIC, and ISIC-SPA algorithms in combination with Random Forest (RF) to select sensitive hyperspectral imaging (HSI) bands. Subsequently, vegetation indices (VIs) were developed from these optimal wavelengths and integrated with LiDAR metrics. Finally, the performance of RF models trained on individual data sources (HSI VIs or LiDAR metrics) and on the combined data (HSI+LiDAR) was evaluated for detecting Dendrolimus spp. damage at the individual tree level. For HSI band selection, compared to the 10 bands selected by SPA-RF (OA = 71.05, Kappa=0.57) and the 21 bands selected by ISIC-RF (OA = 75.44, Kappa=0.63), ISIC-SPA-RF (OA = 70.18, Kappa=0.55) selected only 3 bands and achieved good classification results on the validation set, which substantially reduced data redundancy and improved VI construction. For individual tree-level detection of Dendrolimus spp. damage, four VIS and seven LiDAR-derived metrics were utilized. The results showed that the HSI method (OA = 72.81%, Kappa=0.59) outperformed the LiDAR method (OA = 71.05%, Kappa=0.56). The combined data approach achieved the highest overall accuracy (OA = 83.33%, Kappa=0.75), with an early detection accuracy of 82.93%, which was significantly better than using HSI or LiDAR data alone. Our study demonstrates that LiDAR can effectively capture the spatial distribution changes of needles caused by defoliation, while also revealing spectral reflectance characteristics in the near-infrared (NIR) band. The integration of HSI and LiDAR data significantly enhances the early detection accuracy for Dendrolimus spp. infestations. This approach not only provides critical technical support for Dendrolimus spp. control, but also establishes a novel remote sensing methodology for monitoring other defoliation pests.

Figures