Potato Research Agronomy and Crop Science
Stacked Learning for Disease Severity Prediction of Late Blight in Potato Using AI/ML with Long-term Weather Variables
Abstract
Potato, the 4^th important crop in the world is highly affected by late blight disease. To control this disease, farmers apply high doses of pesticides, which is damaging the ecosystem. Hence, there is a need to optimise the pesticide applications. To achieve this, a reliable and timely forecast of late blight disease severity is required. Stacked learning is an emerging machine learning technique that gives predictions with high accuracy and reliability. Therefore, the present study aims to develop stacked model for predicting late blight disease severity for the location-Srinagar, India using weather-based indices. Weather variables used for developing indices were maximum and minimum temperature, maximum and minimum relative humidity, precipitation, wind speed, and shortwave irradiance. The trend analysis of weather variables was done using Mann–Kendall and Pettitt tests. These variables along with disease severity data of 18 years (2006–2023) were used for calculating weather-based indices for different planting conditions (early, normal, late and combined). Indices were selected using Mutual Information. These indices were used as inputs for developing base learners- RF, SVR, GPR, and XGB. The out-of-fold predictions from these base learners were used as meta features to develop the stacked model to predict late blight disease severity in potato. Trend analysis revealed that 5 variables showed significant shifts over the 18 years. Furthermore, of a total of 56 indices, 10 indices were retained for each planting condition. Thirty-seven of the 40 retained indices were weighted with precipitation, temperature, and wind velocity as major components, demonstrating the importance of temperature, moisture, and aeration in disease epidemiology. Furthermore, on an independent test set, the stacked model outperformed the base models with the lowest nRMSE of 0.172 under early and 0.139 under combined planting condition and comparable to the best base learner (RF) under normal planting, the difference in nRMSE being only 0.003. However, GPR performed better with nRMSE of 0.191 under late planting condition. Hence, the developed model can be used to develop planting condition based disease prediction models and can be integrated into a decision support system to provide timely warnings to the farmers and minimize the pesticide application in managing the late blight disease.