Development of a CNN classifier with XAI to detect interpretable water stress in sweet potato using RGB images

Recent abnormal climate conditions have resulted in a decline in both the yield and quality of sweet potatoes (Ipomoea batatas L.). To overcome this, various deep-learning-driven monitoring techniques have been developed. High-cost hyper- or multispectral imagery integrated with done applications is intensively used with large-scale datasets to accomplish this. While high-cost hyperspectral or multispectral imagery integrated with drone applications is commonly used with large-scale datasets, these methods can be limited by their high costs and operational and maintenance challenges. Therefore, the current study developed a cost-effective monitoring system for evaluating water stress levels using RGB imagery and deep-learning models. A Convolutional Neural Network (CNN) model was served as the base model, and its several hybrid models were produced by combining the CNN with Random Forest (RF), Support Vector Machine (SVM), and Vision Transformer (ViT) were developed. As a result, the CNN-ViT hybrid model has achieved the highest accuracy of 0.99. In addition, to address the low-dimensional input issue, the feature maps extracted by the CNN were utilized for the ViT model. This approach enabled feature visualization of the water stress levels in the RGB imagery of sweet potatoes. Consequently the developed cost-effective RGB imagery monitoring system has demonstrated potential as a practical diagnostic tool for agricultural field monitoring • RGB images were used to classify sweet potato water stress into three levels. • A C-ViT model achieved 99 % test accuracy under AD and AZD augmentation conditions. • Misclassification under dry stress was due to weak visible drought symptoms. • Attention maps and LIME improved the interpretability of model predictions. • A GUI tool was developed for real-time water stress diagnosis using RGB images.