A mathematical modelling-based interpretable deep learning approach for lettuce disease detection in extreme environmental conditions

Abstract

Lettuce is a widely consumed crop with significant nutritional value. However, leaf diseases in the lettuce can degrade plant health, diminish plant yield, and lead to substantial economic losses. Therefore, detection of these diseases at early stage is extremely vital. To address the challenge of disease identification in real-world field conditions, we introduce a multi-level feature extraction framework, CNN-WOPNet. This study utilized a lettuce NPK dataset cultivated under extreme environmental conditions in a hydroponics system. The proposed model utilizes a mathematical Walrus Optimization algorithm for CNN hyperparameter tuning, and a parallel network (ParNet) attention module to develop a novel classification network (CNN-WOPNet). This network processes the multi-level deep features from the optimized CNN and attention module, effectively emphasizing crucial locations in plant disease images. CNN-WOPNet model classified diverse range of plant diseases with an impressive performance metrics such as accuracy 99.54 %, precision 99.60 %, F1-score 99.61 %, and recall 99.61 %. ParNet module demonstrated the shortest training and testing times, 755.82 s and 0.01 s, respectively, while delivering competitive performance compared to existing methods. An ablation study was also conducted, demonstrating the efficacy of proposed model.