A lightweight explainable hybrid convolutional neural network for potato leaf blight detection

Potato (Solanum tuberosum), the fourth most abundant food crop in the world, is subjected to significant challenges due to diseases such as late blight, causing global annual yield loss exceeding $6.7 billion. Current methods of detection are time-consuming and frequently fail to identify early-stage infections. Deep learning, and particularly the Convolutional Neural Networks (CNNs) provide fast and scalable automation of disease classification compared to traditional methods. In this work, potato leaf diseases were classified by many CNNs such as XceptionNet, DenseNet121, a 5-layer CNN, a 6-layer CNN, and a custom hybrid CNN model (SpudNet5-R3). The datasets we used are as follows: (1) PlantVillage dataset was used and (2) PLD dataset which contains healthy, early blight, and late blight potato leaves. A full preprocessing pipeline was carried out which includes resizing, colour normalization, and Contrast Limited Adaptive Histogram Equalization (CLAHE). Targeted data augmentation strategies were applied to tackle the class imbalance. The models were trained, validated, and tested individually on three datasets and the input images were resized to 224 × 224. We analysed performance measures such as accuracy, precision, recall, F1-score, and inference time for finding the best model. In addition, Grad-CAM visualization was used to provide interpretable insights into the model predictions, showing the leaf regions that contributed to the classification. Experimental results show the strong competitiveness of our custom SpudNet5-R3 architecture, obtaining testing accuracy of 99.07% and macro F1-score of 99% on the PlantVillage (D1) potato dataset. It achieved also the testing accuracy of 95.56% and a macro f1-score of 96% on PLD (D2) dataset, demonstrating the successfulness of CNN architectures customized for robust and accurate recognition of the potato diseases.