Evaluating Performance of Machine Learning Models to Predict Climate Zone-Based Crop Yields

In light of the challenges posed by climate change, as highlighted by Wing et al. (2021), which projected a reduction in global crop yields by 3–12% by mid-century and 11–25% by the century's end under severe warming scenarios, six machine learning models have been tested. Leveraging data from the United Nations (UN)’s Food and Agriculture Organization (FAO), this tool integrates parameters such as temperature, rainfall, pesticide usage, and cropland areas for key crops across six climate zones: polar, temperate, arid, tropical, Mediterranean, and mountains. Utilizing advanced machine learning techniques—including Random Forest, XGBoost, Recurrent Neural Network, Artificial Neural Network, Long Short-Term Memory, and K-Nearest Neighbor—the tool evaluates model performances based on R2 and RMSE metrics. The ARIMA model demonstrated the highest accuracy for Maize and Soybeans in the Tropical climate zone, while the Random Forest model excelled for Potatoes, Rice (paddy), Sweet Potatoes, and Wheat. In the Dry climate zone, the ARIMA model outperformed others for Maize, Potatoes, Rice (paddy), and Sweet Potatoes, with Random Forest best for Wheat and Soybeans. In the Temperate climate zone, the Long Short-Term Memory model provided the best predictions for Maize, while Random Forest excelled for other crops, and ARIMA was most effective for Rice (paddy). This study underscores the importance of a hybrid machine learning approach, combining the strengths of various models to address climate change complexities. Hybrid models offer more robust and reliable predictions, supporting farmers and decision-makers in adapting to changing conditions and ensuring sustainable agriculture.