Smart solutions for maize farmers: Machine learning-enabled web applications for downy mildew management and enhanced crop yield in India

Abstract

Increasing use of machine-learning (ML) algorithms in plant disease forecasting is one-way to reduce the global crop yield losses caused by plant pathogens. This study focuses on forecasting maize downy mildew (MDM) and developing a web application to disseminate the information for taking early precautions. The susceptible maize genotype, African Tall, was planted each month from October 2018 to September 2022 in downy mildew sick soil maintained at the maize research plots, V.C Farm, Karnataka, India, yielding 48 disease cycles. A tripartite analysis involving host, pathogen, and weather parameters revealed that maximum temperature was the most influential factor with a feature importance score of 0.76 in the Random Forest algorithm. Other factors scored below 0.2, indicating relatively weaker contributions. Six machine-learning algorithms namely Decision Trees, Random Forests (RF), Support Vector Machines, K-Nearest Neighbors, Bagging Regression and XGBoost Regression were evaluated to forecast MDM using eight performance indicators. The RF algorithm has given the best forecasting task with an R² of 0.97, a Mean Absolute Error (MAE) of 3.78, a Mean Squared Error (MSE) of 11.83, a Root Mean Squared Error (RMSE) of 3.44, a Mean Absolute Percentage Error (MAPE) of 9.09 %, a Symmetric Mean Absolute Percentage Error (sMAPE) of 8.65 %, an Explained Variance Score (EVS) of 0.96, and a Mean Bias Deviation (MBD) of −0.29. JASS, a web tool for forecasting MDM outbreaks, was created using the Random Forest model. It provides real-time, weather-based forecasts to assist with proactive crop management. This study highlights the potential of ML in MDM forecasting and underscores the significance of user-friendly platforms like JASS in enhancing maize yield and ensuring food security. The web application is accessible at https://mdmpdi.pythonanywhere.com.