Comparison of UAV Image-Based Detection Accuracy of Pine Wilt Disease-Affected Trees Using Different Dataset Compositions and Attention Modules

Abstract

This study evaluated the effects of input-channel composition and the incorporation of attention modules on the detection accuracy of visually suspected pine wilt disease (PWD)-affected trees using unmanned aerial vehicle (UAV) RGB orthomosaic imagery and the YOLO26-Large (YOLO26-L) deep learning model.Four datasets were constructed: Dataset A, consisting of three RGB channels; Dataset B, combining RGB with elevation and aspect (5 channels); Dataset C, combining RGB with three graylevel co-occurrence matrix (GLCM) texture features, namely Angular Second Mement (ASM), Entropy, and Homogeneity (6 channels); and Dataset D, combining RGB with the GLCM texture features and topographic information, namely elevation and aspect (8 channels).The Squeeze-and-Excitation (SE) block and Convolutional Block Attention Module (CBAM) were independently integrated into the YOLO26-L baseline model, and the detection performance of the 12 combinations was analyzed.In the accuracy assessment, the YOLO26-L baseline model trained with Dataset B achieved the highest mean Average Precision (mAP)@50 of 0.63 among the 12 combinations.However, the differences among the 12 combinations were marginal and did not indicate the superiority of a specific combination.Dataset C and Dataset D, which incorporated GLCM texture features, achieved accuracy levels similar to those of Dataset A, which used RGB alone, indicating that GLCM texture features did not substantially improve the detection of PWD-affected trees in 5-cm-resolution imagery.The application of attention modules also did not lead to a consistent improvement in accuracy, and the detection rates of all 12 combinations remained around 60% for objects smaller than 25 m².These results suggest that simply adding texture features or attention modules did not reliably improve detection accuracy.This study extends RGB-based UAV detection of PWD-affected trees by progressively integrating elevation, aspect, and GLCM texture features into four datasets and evaluating their interactions with SE and CBAM attention modules across a 12-combination experimental matrix.The results can inform the selection of input channels and model architectures in future forest disease detection studies.