From images to insights: Using a convolutional neural network to improve powdery mildew severity detection in mungbean

To efficiently identify powdery mildew (PM) severity in mungbean leaves, we developed a Convolutional Neural Network (CNN) approach and validated its effectiveness against human evaluation. We fine-tuned an EfficientNet-B3 pre-trained model, which, in our related studies, performed better than re-implemented Inception V3 models. The CNN was trained on 90% of the images (2 880) for training and 10% (320) for validation, with data augmentation applied using Python and TensorFlow. The model obtained 82.10% and 73.03% as training and validation accuracies after 14 epochs, respectively. Further analysis with an additional 15 datasets revealed PM disease indices ranging from 2.03 (resistance) to 6.45 (high susceptibility). The concordance between AI-predicted and human-assessed PM severity was 74.4% (adjusted R2: 72.4%), with an average root mean squared error (RMSE) of 0.854 and a mean absolute error (MAE) of 0.715, indicating moderate predictive error. Comparison of our developed AI-based application prototype on smartphones with expert evaluations yielded a strong correlation (r = 0.992**, R2 = 0.984), suggesting that this tool can effectively estimate PM severity across mungbean cultivars. The application shows considerable promise, and further optimisation and strategic dissemination efforts will enhance its adoption among farmers.