Explainable AI for Black Sigatoka Detection in Constrained Resource Settings
View/ Open
Date
2023-06-30Author
Kalungi, Joshua Edward
Yiga, Gilbert
Kayanja, Emmy William
Kyagaba, Jonah Mubuuke
Metadata
Show full item recordAbstract
Agriculture is a crucial sector in Uganda’s economy, serving as the primary source of income for many Ugandans. However, plant diseases are a major issue in agriculture as they have an impact on both the quantity and quality of food produced. The death of the plants, a decline in farmer profits, and increased production costs during the control process are just a few of the detrimental effects of plant diseases. Banana Black Sigatoka (BBS) Leaf Disease, which is brought on by fungi of the genus Pseudocercospora, poses the greatest challenge of all the diseases, causing significant economic losses worldwide. Therefore, early detection and timely intervention are crucial for preventing the spread of the disease.
In recent years, Machine Learning (ML) has shown great potential for detecting and diagnosing plant diseases, including BBS. However, the lack of transparency and interpretability of ML models raises concerns about their responsible use. This study has put a lot of emphasis on the Explainable Deep Learning model for banana black Sigatoka detection in order to give local farmers and experts in resource-constrained situations more clarity and understanding.
To achieve this goal, a transfer learning technique where knowledge is gained from large trained datasets was utilized. We used a framework based on the pre-trained models MobileNetV2, AlexNet CNN, and ResNet1 to detect the disease. AUC/ROC, F1 score, confusion matrix, and its derivatives were used to evaluate the performance of algorithms.
Additionally, this study examined some of the most popular Explainable AI (XAI) techniques like Saliency, Input x Gradient, Layerwise Relevance Propagation, Integrated Gradients, Guided Backpropagation, Occlusion, Grad-CAM, Guided Grad-CAM, Lime, DeepLift, as well as their replacements that use the SmoothGrad approach.
We demonstrated the effectiveness of our model through extensive experiments and showed that it outperformed existing state-of-the-art models for BBS detection. Our model not only provided accurate and interpretable results but also promoted responsible AI practices for plant disease diagnosis.