Comparison of CNN models for application in crop health assessment with participatory sensing

Abstract

Timely and robust diagnosis of plant diseases and nutrient deficiencies play a major role in management of crop yield. Automation is a low cost alternative to human experts and can help to detect early onset of crop diseases which aids faster decision making and in giving recommendations to farmers to curb yield loss. We have developed a smart-phone based participatory sensing application for agriculture which is used by farmers to scout their fields for events of interest, especially those related to crop health. Recently, deep convolutional neural networks (CNN) have emerged as a prominent technique in computer vision related challenges and such deep-learning based models could prove as an important tool to do just-in-time assessment of crop health. With a view to building state-of-the-art diagnostic capabilities on the phone, we present analysis of CNN models in terms of accuracy, memory, and inference time. Effects of change in hyperparameters have been evaluated in terms of accuracy. The trained model gives 99.7% classification accuracy with satisfactory inference time and model size which assures the application of CNN architectures for real-time crop state diagnosis on a large scale with limited hardware capabilities.