Assessing Gonipterus defoliation levels using multispectral unmanned aerial vehicle (UAV) data in Eucalyptus plantations

Abstract

Invasive insect pest Gonipterus sp. n. 2 (Coleoptera: Curculionidae) threatens Eucalyptus species, causing defoliation and yield loss through adult and larval feeding. Early detection is important for early intervention to prevent pest outbreaks. As conventional insect pest monitoring methods are time-consuming and spatially restrictive, this study assessed the potential of UAV monitoring. Multispectral imagery was obtained with Unmanned Aerial Vehicles (UAVs) across six different stands of young Eucalyptus dunnii with varying levels of Gonipterus sp. n. 2 infestations. Some stands were revisited, a total of 9 datasets were covered. Reference damage levels were obtained through visual assessments of (n = 89–100) trees at each site. Across sites, a decrease in canopy reflectance in both the visual and the near-infrared domains with increasing damage levels was consistently observed. Several vegetation indices showed consistent patterns, but none showed site independence. XGBoost, Support Vector Machine and Random Forest (RF) were used to predict damage levels using five input spectral data types. XGBoost performed best, closely followed by RF. Both models consistently selected very similar features. The best-performing models included reflectance, vegetation indices and grey-level co-occurrence matrix data. When data from 10 different wavelengths were used, the highest classification accuracy was 92 % across all sites in classifying defoliation levels. With a classical 5-band multispectral camera, accuracy was 88 %, but distinguishing medium damage from low remained challenging. However, the method was less reliable when trained and validated on separate fields. This study highlights the potential of multi-site datasets in increasing the model's generalization, using UAV based multispectral imagery to assess Gonipterus sp. n. 2 damage and demonstrating reliable upscaling from individual tree assessments to stand scale. However, it also recognises the difficulty of generating a robust model that performs well on untrained sites.