Accuracy of Deep Learning-Based Satellite Image Analysis in Early Detection of Insect Infestation-Induced Tree Mortality: A Comparative Analysis with Conventional Remote Sensing Methods

Abstract

Kaan Alper

Bark beetles, exacerbated by drought and temperature increases intensified by climate change, have caused unprecedented levels of tree mortality in coniferous forests of the Northern Hemisphere over the past decade. Early-stage detection of infestation foci — particularly during the green-attack phase, when no visual symptoms are yet apparent in the foliage — is critically important for preventing outbreak spread and reducing economic losses. This study aims to systematically and comparatively examine the accuracy of deep learning-based satellite image analysis in the early detection of insect infestation-induced tree mortality against commonly employed conventional remote sensing methods. The research was conducted in the Bohemian Forest on’ndaki Ips typographus NDVI/NDMI thresholding, Random Forest, Support Vector Machines, and Maximum Likelihood classification within a five-fold spatial cross-validation framework. Results demonstrated that the U-Net model achieved the highest overall performance with 91.4% overall accuracy and a Kappa coefficient of 0.88. (Picea abies) multitemporal Sentinel-2 imagery from the 2019–2022 period. A U-Net model with a ResNet-50 backbone was compared with NDVI/NDMI thresholding, Random Forest, Support Vector Machines, and Maximum Likelihood classification within a five-fold spatial cross-validation framework. Results demonstrated that the U-Net model achieved the highest overall performance with 91.4% overall accuracy and a Kappa coefficient of 0.88. The performance gap among methods varied systematically according to infestation stage: while the F1-score difference between U-Net and Random Forest was only 3 percentage points during the grey-attack stage, this gap increased to 21 percentage points (U-Net: 0,72; RF: 0,51) during the green-attack stage. Grad-CAM and SHAP explainability analyses revealed that the model primarily focused on SWIR bands (B11, B12) and red-edge bands (B05, B06) for green-attack detection, whereas the contribution of NDVI’remained negligible. These findings demonstrate that deep learning provides a statistically significant advantage over conventional methods in early bark beetle detection, and that this advantage predominantly emerges during the green-attack stage, where spectral change is least pronounced. The study recommends updating operational forest health monitoring systems with SWIR-based indices and deep learning modules. This article was prepared using the purpose-built MUTEFFERIQA software and the Claude Opus 4.6 Large Language Model (LLM) to contribute to scientific research.

PDF