Chroma-Sense: a Memory-Efficient Plant Leaf Disease Classification Model for Edge Devices

CNN models used in disease management systems for disease detection are typically stored and executed on remote cloud servers, communicating with sensors over the internet. This paradigm raises concerns about data security and reliability, especially in remote farmlands. Edge computing introduces a computing layer closer to sensory nodes, enabling more reliable communication and faster results. These edge devices are often resource constrained and cannot run CNN models that need high memory and computation. Since plant disease semantics are simple and consistent across channels, Chroma-Sense proposes processing the R, G, and B channels independently using the same feature extractor. Individual processing reduces the width of the feature extractor, lowering RAM usage and number of computations performed. In addition, reusing the same feature extractor minimizes the parameter count, the Flash memory required to store the model, enabling efficient edge deployment. Proposed Chroma-Sense was trained on a subset of the PlantVillage dataset and achieved a 25% reduction in peak RAM usage and a 60% reduction in flash memory when tested on three different edge devices with varying heap and storage capacities.