- Explainable Deep Learning for Field-Level Crop Disease Detection: A Simulated Precision- Agriculture Evaluation

This article presents a simulated empirical study in agricultural science, computer vision, and explainable artificial intelligence. The simulated dataset included 6,720 training images, 1,440 validation images, and 1,440 hold-out images, with deliberate class imbalance and domain shift. A simulated image-classification benchmark was constructed to represent 9,600 crop-leaf images across eight disease and healthy categories. Metadata captured illumination, camera distance, occlusion, leaf age, and background complexity. Three model families were compared: a standard CNN, a transfer-learning network, and a lightweight explainable model with saliency-based localization checks. Performance was evaluated on a clean validation set and a separate field-complexity hold-out set. The transfer-learning model achieved the highest simulated clean-set accuracy, while the lightweight explainable model retained the strongest field hold-out macro-F1 and the lowest inference latency. Explanation-consistency scoring indicated that some high-confidence CNN predictions relied on background regions rather than symptomatic tissue. The explainable model reduced this failure pattern and maintained a 0.89 field macro-F1. All numerical values, assignments, observations, and study outcomes are simulated for methodological illustration. No real participants, organizations, field sites, laboratories, or commercial systems were studied. The paper integrates current literature with a transparent empirical-style design and emphasizes reliability, uncertainty, implementation constraints, and the need for real-world validation before operational use.