Improved Soybean Leaf Disease Detection by Utilizing Separable-CNN and Global Average Pooling

Abstract

Soybean is one of the most important oilseed and protein-rich crops, and its productivity is seriously affected by leaf diseases that reduce both yield and quality. Both timely and prompt detection of the disease is thus a requirement in proper crop management especially in farming conditions of India where timely diagnosis at a field level can provide the opportunity to act fast. This paper proposes a lightweight deep-learning model, which is a Separable Convolutional Neural Network (Separable-CNN) with Global Average Pooling (GAP), to detect soybean leaf disease. The depthwise separable convolutions decrease the level of computation, and the GAP-based aggregation of features tends to reduce overfitting and enhance compactness of the model. The suggested approach was trained and tested on a dataset of 40,000 leaf images of soybean cultivated in the field of eight classes, the seven classes comprising of disease classes, while the remaining one was a healthy one. It was observed in experiments that the model obtained a test accuracy of 95.70% with only a small architecture of about 0.78 million parameters and a model size of 9.04 MB. Relative analysis of the existing deep-learning architectures proved that the proposed architecture provides a good tradeoff between the classification performance and the computational efficiency. These results suggest that the suggested model can be effectively used in agricultural devices with limited resources to detect the soybean disease in real-time.