Integrated Agro-Diagnostic Framework with Deep Learning for Crop Disease Management

Abstract

Plant diseases are a constant issue in food production, resulting in worldwide losses of 20–40% on an annual basis and more than $147 billion in damages. In India, more than 150 million farmers have an annual contribution of nearly 18% to the country's GDP, and timely diagnosis of diseases is not easily accessible in rural areas because of the very limited availability of trained plant pathologists, which leads to the loss of yields in the field ranging between 25% and 50%. Most of the existing CNN-based systems are only concerned with the detection of diseases, and farmers are not provided with any practical suggestions for treatment or input procurement. In response, we present a lightweight CNN network called AgriNet, trained on the Plant Village dataset of 61,486 labeled leaf images that represent 38 different classes for 14 different crops, and present an Integrated Agro-Diagnostic and Remediation Framework (IADRF) to address this. Compared with other networks under the same experimental setup, AgriNet is compared with Xception, ResNet-50 and Inception V3 in terms of accuracy, precision, recall, F1-score, the number of parameters and inference latency on the same training, and on the same dataset. The proposed model not only reaches high accuracy of 98.68% but also yields a high F1-scores of 98.47% using much less trainable parameters than the conventional CNN architectures. The prototype IADRF combines disease predictions with a structured collection of information regarding disease prevention and management and presents it through a user-friendly web interface on a mobile device. The reported latency results prove the possibility of calculations on the server side and in low resource areas, but direct deployment of the smartphone and validation at the field level remains a topic for further research.

Keywords

Graphical Abstract

Novelty Statement

The prototype Integrated Agro-Diagnostic and Remediation Framework (IADRF) combines disease predictions with a structured collection of information regarding disease prevention and management and presents it through a user-friendly web interface on a mobile device.