i-PomDiagnoser: A Real-Time Pomegranate Disease Management System

Objectives: Designing and developing i-PomDiagnoser: a real-time pomegranate disease management system for disease detection, classification, prediction, recommending preventive measures, and analyzing abrupt climatic changes and their impact on pomegranates. Methods: A data collection framework has been designed and developed using an agriculture drone, sensors, camera, and other equipment to collect real field pomegranate images and micro-level parameters. Comprehensive Exploratory Data Analysis (EDA) and Feature Selection (FS) processes were carried out to improve the accuracy of disease classification and forecasting models. ML-based Binary, Multimodel, and Multilabel classifiers were implemented for disease classification. The models were trained on 11 years of historical data and tested on 5 months of actual field data. A hybrid pomegranate disease forecasting model has been developed for accurately forecasting micro-level parameters for the next 45 days to predict diseases. Findings: Micro-level (weather, soil, water) parameters specific to the agro-climatic zone were collected. The five most prominent distinct diseases are considered for experimentation namely Bacterial Blight (Telya), Anthracnose, Fruit spot, Fusarium Wilt, and Fruit borer. The proposed Improved Ensemble Multilabel Classifier (i-Ensemble-MLC) with a modified voting scheme has achieved a high classification accuracy of 95.82%, addressing model overfitting and data imbalance. Moreover, the hybrid pomegranate disease forecasting model, combining LSTM and i-Ensemble-MLC, demonstrated better performance with minimal error rates (MSE: 0.003, RMSE: 0.071, MAE: 0.048, R2: 0.7) compared to the existing model1 (MSE:0.037, MAE:0.028). Novelty: The novelty lies in the creation of the all-in-one model, i-PomDiagnoser. This innovative system helps the farmers to correctly detect and predict the most prominent diseases of pomegranate. Keywords: Pomegranate, Agriculture, Disease Forecasting, Machine Learning, Deep Learning