Enhancement of Cassava Leaf Disease Detection Using InceptionNet and Squeeze and Excitation Attention Mechanism

Automated cassava leaf disease detection using images has gained immense importance as it is consumed by 800 million people as a staple crop across the globe. This paper proposes a novel approach to cassava leaf diseases classification based on InceptionV3 convolutional model augmented with Squeeze and Excitation attention mechanism for enhanced feature representation. The proposed work also attempts to improve robustness through extensive data augmentation, balanced sampling techniques and adaptive learning rate adjustment mechanism. For the evaluation of the performance of the proposed model, different techniques for explainability have been adopted, including Grad- CAM and LIME, which bring to the fore the mechanism of decision-making by our model. The proposed model is compared against six fine-tuned convolutional architectures and the results demonstrate the highest accuracy and F-score values for disease classification, thus showing the efficacy of the proposed approach. Over-all this research contributes to the general area of agricultural disease diagnosis but also sheds light on the interpretability of deep learning models for complex classification tasks.