EdgePlantNet: Lightweight edge-aware cyber–physical system for plant disease detection using enhanced attention CNNs

Abstract

The advances in sensing and computing methodologies have allowed ubiquitous Cyber–Physical Systems (CPS) which have enabled intelligent monitoring and management of crop plants, leading to Smart Agriculture. Yet, the computational constraints of the edge-computing devices have been a roadblock for utilization of complex processing algorithms for real-time applications like leaf-disease detection, were immediate and highly accurate results are of paramount importance. To address this, we propose EdgePlantNet, a Lightweight Edge-Aware CPS for Plant Disease Detection using Enhanced Attention CNNs. It comprises a novel dual-branched Convolutional Neural Network (CNN) architecture that incorporates an improved multi-layered perceptron based spatial attention mechanism (MLP-ATCNN). The MLP-ATCNN is fed with both the original leaf image and its segmented copy, allowing it to simultaneously focus on the leaf image at two scales namely, the diseased regions, and the overall leaf. This allows it to learn robust discriminatory features corresponding to different diseases, even when trained with much lower samples of data. We validate the performance of the EdgePlantNet on two popular and diverse datasets that are the PlantVillage and the BPLD dataset. The novelty of our proposed CPS much lower computational complexity and high disease detection accuracy as compared to other state-of-the-art methods. We also implement the EdgePlantNet on a resource constraint IoT edge device, demonstrating its efficiency for mobile computing.