Deep Learning Based Efficient Lightweight Model with Channel and Spatial Attention Mechanism for Rice Disease Detection

Abstract

The accurate identification of rice leaf disease is necessary to protect the crop in time, and numerous deep learning models have high computational demands and are hard to implement in agricultural settings with limited resources. The proposed systems are introduce two small convolutional neural networks to classify rice disease. First one, Modified Lightweight Convolutional Neural Network (MLWCNN), employs depth wise separable convolution and pointwise convolution to minimize the computational cost. The second, Modified Lightweight Model with Channel Attention Convolutional Neural Network (MLWCACNN), is based on the lightweight backbone, but it incorporates channel and spatial attention to enhance the acquisition of disease-related features. The imbalance between classes was dealt with by controlled augmentation, resulting in balanced datasets of 6310 and 2650 images. The 5-fold cross- validation was performed on 224 × 224-sized images, Adam optimization, and conventional classification metrics. MLWCNN obtained average accuracy of 78% on the four-class rice dataset but MLWCACNN obtained accuracy of 98% on the six-class dataset. The models proposed needed 0.047M and 0.059M parameters, respectively, indicating a viable accuracy complexity trade-off to smart agricultural diagnosis.