Localization evaluation of CAM based explainability techniques for plant disease detection

Authors : Duygu Sinanç Terzi

Pages : 1287-1298

Doi:10.5505/pajes.2025.50955

View : 36 | Download : 138

Publication Date : 2025-12-15

Article Type : Research Paper

Abstract :In recent years, computer vision technologies have played a critical role in precision agriculture, leveraging robotics and artificial intelligence to automate tasks in crop production. While image-based applications hold promise, model interpretability remains a significant challenge. Explainable artificial intelligence aims to address this by providing plant scientists with interpretable, reliable information, improving the understanding of plant diseases. This study focuses on integrating explainability metrics into model evaluation, with a detailed analysis of explainability methods applied to plant disease classification models. Using Class Activation Mapping based visualization methods with architectures such as EfficientNet, MobileNet, ResNet, and ShuffleNet, trained on a public plant disease dataset, the study assessed both classification success and model explainability. Localization results were derived from an energy-based perspective, assessing how well saliency maps aligned with bounding boxes of diseased areas. The findings reveal that feature dimensions and positions in the images significantly influence classification outcomes, highlighting the importance of precise annotations during data labeling. This study uncovers potential biases in disease detection and emphasizes the need for explainability metrics in evaluating deep learning models, paving the way for more accurate and efficient plant disease detection techniques.
Keywords : hassas tarım, bilgisayarlı görme, derin öğrenme, bitki hastalığı sınıflandırması, açıklanabilirlik

ORIGINAL ARTICLE URL