AgroXAI: a multi-XAI deep learning system for explainable multi-species plant disease classification

Automated plant disease classification can benefit from models that provide both predictive performance and interpretable evidence for their decisions. This study presents AgroXAI, a Multi-XAI deep learning framework for explainable multi-species plant disease classification under a controlled, leakage-aware evaluation protocol. A publicly available dataset containing 2,815 leaf images from seven plant species was organized into 14 joint species-health classes and partitioned into fixed group-disjoint training, validation, and held-out test sets containing 1,970, 422, and 423 images, respectively. Preprocessing and architecture selection were conducted using only the training and validation partitions, leading to the AgroX-Swin classifier based on Swin Transformer-Tiny with 384 × 384 Original RGB inputs. Explainability was quantitatively evaluated using Grad-CAM++, SHAP, LIME, and their complementary fusions on a predefined class-balanced subset of 42 held-out images. Across five independently initialized training runs, AgroX-Swin achieved a mean held-out accuracy of 87.19% and macro-F1 of 0.8300 ± 0.0122, while the five-model probability ensemble achieved 87.94% accuracy and 0.8355 macro-F1. Deterministic aggregation of the same 14-class predictions yielded 99.29% species-recognition accuracy and 88.42% binary health-status accuracy, with a diseased-class F1-score of 0.9160. Temperature scaling reduced the representative model's expected calibration error from 0.0792 to 0.0310. At the 20% perturbation level, Grad-CAM++ + LIME produced the largest mean confidence drop (0.2540), while Grad-CAM++ showed the highest transformation stability (0.8371) and SHAP + LIME showed the highest attribution-burden stability (0.7954), demonstrating complementary rather than uniformly superior behavior among individual and fused explanation methods. Model-randomization analysis further examined the dependence of Grad-CAM++ explanations on learned model parameters. AgroXAI provides an integrated framework combining leakage-aware model development, multi-seed training-stability assessment, held-out classification, calibration, statistical uncertainty analysis, and quantitative evaluation of complementary XAI methods. The resulting attribution maps are interpreted as prediction-relevant visual explanations rather than validated lesion localization or quantitative disease-severity estimates. External robustness and real-world agricultural generalization remain to be established using independently acquired field datasets.