Optimizing Convolutional Neural Networks for Plant Disease Detection on ARM-M Microcontrollers

Rice production in Malaysia is facing significant challenges due to plant diseases and environmental hazards, leading to a decline in the rice self-sufficiency ratio. To address these issues, this study explores the deployment of AI-driven plant disease detection models on ARM-M based microcontrollers, specifically the STM32H747I-DISCO Discovery kit. Utilizing Convolutional Neural Networks (CNNs) for image recognition, we aim to achieve real-time disease monitoring without relying on continuous internet connectivity. Our research evaluates five state-of-the-art models-FD-MobileNet, MobileNet V1, MobileNet V2, ResNet V1, and SqueezeNet V1-focusing on their suitability for resource-constrained environments. The results indicate that FD-MobileN et is the most resource-efficient model with the lowest number of Multiply-Accumulate operations (MACCs) and the shortest inference time, achieving an accuracy rate of 98.44%. However, SqueezeNet, despite having the highest accuracy at 99.22%, faces deployment challenges due to its large model size on the STM32H747I-DISCO board. This work highlights the potential of embedded AI applications in enhancing agricultural resilience and efficiency.