Explainable CNN-BiLSTM Framework for Okra Leaf Disease Detection and Classification Using Grad-CAM and SHAP Visualization

Okra (Abelmoschus esculentus) is an economically important vegetable crop in Nigeria and across sub-Saharan Africa, yet it remains underrepresented in agricultural artificial intelligence research, and its leaf diseases are difficult to distinguish by eye. This study presents an explainable CNN-BiLSTM (ECBL) framework for automated okra leaf disease classification across six categories: Alternaria leaf spot, Cercospora leaf spot, Downy mildew, Healthy, Leaf curly virus, and Phyllosticta leaf spot. The model couples an ImageNet-pretrained EfficientNetB3 backbone for spatial feature extraction with a bidirectional long short-term memory (BiLSTM) layer and a soft attention mechanism that weights diagnostically relevant leaf regions. Training used frozen-backbone transfer learning with on-the-fly augmentation and inverse-frequency class weighting on a public okra dataset of 1,495 field images. Because a 150-image test partition cannot separate closely matched models, the central comparison is based on repeated training rather than on a single split. Across five independent runs, the proposed model reached a mean accuracy of 91.6% with a standard deviation of 2.4 points and a mean macro F1-score of 0.916, against 79.6% and 0.795 for the strongest baseline, a standard Convolutional Neural Network (CNN)-LSTM built on the same backbone; the difference of 12.0 accuracy points is significant at p = 0.026, and the proposed model's run-to-run spread is roughly a third as large. A single reference run, used for the detailed error and explainability analysis, reached 86.7% accuracy, with a 95% confidence interval of 0.807 to 0.920, a macro F1-score of 0.863, and a macro one-vs-rest area under the receiver operating characteristic curve of 0.999 with 12.49 million parameters. On that individual run, the baseline scored 89.3%, and the two models could not be separated (McNemar p = 0.571), which is itself the clearest illustration of why repeated runs are necessary. Grad-CAM and SHAP were integrated for interpretation, and a deletion test over 60 images confirms that the Grad-CAM maps carry information about the decision: removing the most important 20% of pixels reduced predicted-class confidence by 0.377 against 0.113 for a contiguity-matched control (p < 0.001). The results indicate that an attention-based CNN-BiLSTM is an accurate, unusually stable, and computationally efficient option for okra disease detection.