Forecasting Risk of Crop Disease with Anomaly Detection Algorithms

Area of Science:

• Agricultural Science
• Data Science
• Epidemiology

Background:

• Crop disease surveillance often omits data from healthy crops, hindering accurate risk forecasting.
• Existing models struggle with imbalanced datasets, lacking information on disease absence.

Purpose of the Study:

• To develop and evaluate anomaly detection techniques for forecasting crop disease risk using only outbreak data.
• To assess the performance of various anomaly detection algorithms in predicting disease outbreaks.

Main Methods:

• Anomaly detection algorithms were trained on weather conditions associated with historical crop disease outbreaks.
• Five algorithms (robust covariance, one-class k-means, Gaussian mixture model, kernel density estimation, one-class SVM) were tested for forecasting accuracy.
• A case study using potato late blight data from Great Britain was conducted.

Main Results:

• Gaussian mixture model achieved the highest forecast accuracy (97.0%), followed closely by one-class k-means (96.9%).
• Ensemble methods combining algorithms provided more accurate and robust forecasting.
• The developed models demonstrated high accuracy in hindcasting historical outbreak events.

Conclusions:

• Anomaly detection offers a novel and effective approach to crop disease risk forecasting with imbalanced outbreak data.
• Gaussian mixture models and ensemble approaches show significant promise for agricultural decision support.
• These techniques are adaptable to other crop diseases and geographical regions.