An augmented attention-based lightweight CNN model for plant water stress detection

Abstract

Recently, deep learning techniques specifically the Convolutional Neural Networks (CNNs) have reported outstanding results from the application for plant water stress detection based on computer vision system compared to other machine learning methods. However, the size of the conventional CNN models is generally too large for its deployment on resource-limited devices such as mobile smartphone or embedded devices. In this study, a lightweight CNN is proposed by incorporating attention mechanism as an augmentation module into the model. The model was trained, validated, and tested using plant images of Setaria grass undergone three water stress treatments. Experimental results show that the proposed method improved the interclass precision, recall, F1-score, and the overall accuracy by more than 9%. Compared to the established lightweight CNN models, the proposed lightweight CNN achieved faster computational time with comparable parameters. In addition, the proposed lightweight model is also efficient when trained on small plant dataset with limited overfitting.