Explainable fine-grained feature enhancement framework for strawberry diagnostic decision making

As the number of smart agricultural automation systems increases, accurate disease diagnosis is essential to maximize production efficiency. However, early-stage symptoms of strawberry diseases (e.g., Tipburn, Phytophthora blight, and Fusarium wilt) often manifest as low-contrast micro-symptoms that are visually similar to the inherent complex texture of strawberry leaves, leading to high technical difficulties in achieving an accurate diagnosis. Furthermore, providing an explanation for the diagnosis remains challenging, which hinders trust in field applications. We present EN-FDEA, a deep learning framework that integrates an EfficientNetV2 backbone with a Convolutional Block Attention Module (CBAM) and a Fine-grained Discriminative Enhancement (FDEA) module combining depthwise separable convolution, parallel feature amplification, and channel recalibration, and we evaluate it together with a provenance audit of the corpus on which it is trained. Parsing the acquisition metadata embedded in every file name showed that the 1,000 abnormal images of each disease originate from only 33-73 individual plants, and that normal and abnormal images share no greenhouse; under the image-level cross-validation used in the original design, 91.9-99.8% of abnormal plants appear in both the training and the test partition of a fold. All results reported here are therefore obtained under plant-disjoint cross-validation with five random seeds, with a site-disjoint arm as the strictest comparison. Removing plant leakage lowers balanced accuracy by up to 0.082 and removing site overlap by up to 0.173. Under the leakage-free protocol, EN-FDEA performs comparably to, and not reliably better than, its EfficientNetV2 backbone (mean paired balanced-accuracy differences of -0.013 to +0.012 across datasets and conditions); the only Holm-corrected significant advantage is Phytophthora blight under overexposure (+0.012, p = 0.047). Forensic analysis further shows that the elevated performance previously observed under synthetic overexposure is a property of the transform: a logistic-regression classifier using only global exposure statistics reaches balanced accuracies of 0.81-0.92, and models trained on original images transfer poorly to synthetic overexposure while transferring well to genuinely over-illuminated frames selected from the corpus itself. We report explanation faithfulness, calibration, and measured deployment cost, and we state where the corpus structure, rather than the model, bounds what can be concluded. Code, split definitions, and per-image predictions are released so that every number can be regenerated.