Context-Aware Explanation Drift Detection (CA-EDD): For Plant Disease Severity Estimation

Abstract

Deep learning models have shown high accuracy in automated plant disease classification; however, their black-box nature limits adoption in precision agriculture, where biological validity and interpretability are critical. Conventional Explainable Artificial Intelligence (XAI) methods, such as Grad-CAM, generate visual saliency maps that often lack alignment with true pathological symptoms, leading to predictions that are accurate yet biologically inconsistent. This paper proposes SymptomConsistency Guided Explainable AI (SCG-XAI), a novel ContextAware Explanation Drift Detection (CA-EDD) framework that validates model reasoning against established plant pathology principles. The framework integrates an explanation generator, an explanation embedding module that encodes attribution maps into structured symptom descriptors capturing lesion color, texture, and spatial distribution, a temporal drift analyzer to detect shifts in model reasoning across disease severity stages, and a context integration layer that constrains explanation validity using agronomic criteria, specifically the Relative Lesion Height (RLH) and the Standard Evaluation System (SES) for rice sheath blight caused by Rhizoctonia solani. By evaluating the semantic consistency between model explanations and physiological disease symptoms, SCG-XAI enables the detection of logic drift that is not reflected in conventional performance measures. Experimental results on a multi-severity rice sheath blight dataset demonstrate that SCG-XAI maintains competitive classification accuracy while ensuring that model explanations are biologically consistent and trustworthy for real-world field deployment.

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.