Lettuce Disease Detection for a Home-Based Smart Farm System

Food security and agricultural growth continue to become pressing concerns globally, which calls for the development of smart farm systems. A component that must be integrated with smart farms is machine learning (ML), an artificial intelligence (AI) tool. This ensures successful harvest through a lettuce disease monitoring system. Instead of cloud computing implemented by previous systems, edge computations can solve challenges such as patchy internet connectivity, which requires added power. The study highlights the use of ML, not only on strong desktop GPUs, but also on resource-constrained or edge devices. Hence, the use of TinyML allows for ML deployment on edge devices with a low-power implementation. Notable Convolutional Neural Network (CNN) Keras-compatible models that can run on the Raspberry Pi 3B+, a popular choice for ML, are MobileNetV2, EfficientNetV2S, and ResNet50V2. A lower-end device, such as the STM32 NUCLEO-F767ZI, has difficulty in running much larger models. Hence, model variations, STEfficientNetLCV1 and ResNet26V1, are used for the STM32. This study analyzes the models’ architecture, accuracy, and resource usage. The results highlight that MobileNetV2 is the most efficient quantized model on the target devices, with the fastest predictions, lowest memory usage, and lowest maximum power consumption. The model boasts 88% and 87% accuracy on the Raspberry Pi and STM32 device respectively. However, a noticeable difference is in the power consumption at 2.24 W and 0.44 W while achieving the same accuracy. This, then, shows the STM32’s ability to run the same ML application as the Raspberry Pi with a TinyML approach.

I. Introduction

Smart Hydroponics is a leading application of IoT-based technologies that integrates real-time monitoring with automation, collection, and process management, cultivating soilless culture using water and nutrient solutions [1]. The smart farm project relies on sensor nodes to transmit data in real-time by sensors deployed in the system via wireless connection, or wireless sensor networks (WSN). Edge devices, on the other hand, bridge the gap between the end point of a network’s reach and the external world. In mitigating high power demands, an edge application to detect plant diseases, with a WSN, shows great promise to enhance the project.