Enhancing Crop Disease Detection Systems With Explainable AI Techniques for Deep Learning Models Using Spectral Imaging

Abstract

Recognizing crop diseases at an early stage is essential for modern agriculture because it greatly enhances crop output and decreases economic losses. Manual examination and specialized expertise are the backbone of traditional disease detection approaches, but they can be exhausting and error-prone. In response to these difficulties, Deep Learning (DL) models have become an effective means of improving agricultural disease detection systems. Through autonomous learning and feature extraction from cropped images, these models-especially Convolutional Neural Networks (CNNs) have shown to be quite effective in image categorization tasks. Train deep-learning models to accurately identify a broad range of illnesses by using massive datasets of categorized crop images. Improved food security and more sustainable farming practices are the end results of incorporating DL models into crop disease identification systems, which secures and improves diagnosis while giving farmers more agency to make well-informed choices. As for crop disease prediction, we also tested the efficacy of several fine-tuning TL models such Efficient-Net, Squeeze-Net, and Dense-Net-121. Models were trained using the open-source Plant Village dataset. The study found that the Dense-net-121 model achieved the highest accuracy rates, with 98.5% on the training dataset and 97.85% on the testing dataset.