FL-ToLeD: An Improved Lightweight Attention Convolutional Neural Network Model for Tomato Leaf Diseases Classification for Low-End Devices

Abstract

The agricultural sector is still a major provider of many countries' economies, but diseases that continuously infect plants represent continuous threats to agriculture and cause massive losses to the country's economy. In this study, a faster and lightweight tomato leaves diseases detection model was proposed for tomato disease classification based on a soft attention mechanism with a depth-wise separable convolution layer. With a model size of 2.5 MB and 221,594 trainable parameters, the proposed model achieved 99.5%, 99.10%, 99.04% for training, validation and testing accuracy respectively, and 99 % for each of precision, recall, and f1-score, it also achieved 99.90% for ROC-AUC with average inference time of 2.06924~mu s. The proposed model outperformed Ulutaş and Aslantaş (2023) by 2.2% in terms of accuracy, precision, recall and f1-score. Additionally, it performed better than Agarwal (2023), Abbas (2021), and Verma (2020) in terms of accuracy, precision, recall, and f1-score by 8%, 2%, and 6%, respectively. It also outperformed Arshad (2023) by 4.77%, 8.92%, 35.18% and 5.11% in terms of accuracy, precision, recall and f1-score, respectively. Furthermore, the proposed model is 90 times smaller than Verma (2020) and 2.5 times smaller than Ulutaş and Aslantaş (2023) in terms of model size. All this makes the proposed model more suitable for low-end devices in precision agriculture.