An Explainable Hybrid Deep Learning–Fuzzy Decision Framework for Human-Centered Plant Stress Severity Assessment

Abstract

Abstract Mild stress is often difficult to distinguish from non-stress signals that may even mask the detection of stress; therefore, early diagnosis and precision grading of plant/microbial stress severity are essential for sustainable precision agriculture toward achieving optimized yields. We propose an interpretable hybrid deep learning–fuzzy decision framework combining EfficientNet B7 and Inception-ResNet-v2 with multiscale feature aggregation integrating Sparse Pyramid Pool (SPP) and Atrous Spatial Pyramid Pooling (ASPP). A Gaussian-based fuzzy inference system is incorporated to derive severity reasoning in a linguistically interpretable manner to address uncertainty and overlapping stress stages. Unlike conventional approaches evaluated only on controlled datasets, the proposed framework is validated through stringent cross-dataset generalization between the PlantVillage and PlantDoc datasets. The model demonstrates robustness under environmental disturbances and passes statistical significance tests. On the PlantVillage benchmark, the framework achieves an exact-match accuracy of $97.82\%$, a macro F1-score of $97.60\%$, and an AUC of $0.979$. When evaluated across a different domain, the performance decreases by only $4.8\%$, indicating strong generalization capability. The integration of fuzzy logic reduces adjacent-class error by $3.4\%$ and improves probability calibration with an Expected Calibration Error (ECE) of $0.021$. Grad-CAM visualizations and saliency analyses further confirm that the model focuses on biologically relevant diseased regions. These results demonstrate that combining multiscale deep feature learning with structured fuzzy reasoning enhances robustness, interpretability, and decision stability, thereby supporting human-centered agricultural monitoring systems.

Plant phenotyping relevance

植物のストレス重症度を画像から推定する深層学習・ファジー推論手法を開発し、異なるデータセット間で検証しているため、植物フェノタイピング手法が中心である。

Code and data availability

The paper's plant-phenotyping measurements rely on two public leaf-image datasets, PlantVillage and PlantDoc, both cited with explicit public URLs in the supplied text. No author analysis code, trained models, or code deposit is mentioned.