Evaluating the Utility of Weather Generators in Crop Simulation Models for In-season Yield Forecasting

Authors: ROHIT, NANDAN - Oak Ridge Institute For Science And Education (ORISE); Bandaru, Varaprasad; MEDURI, PRIDHVI - University Of Maryland; JONES, CURTIS - University Of Maryland; LOLLATO, ROMULO - Kansas State University

Submitted to: Agricultural Systems
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/18/2024
Publication Date: 7/29/2024
Citation: Rohit, N., Bandaru, V., Meduri, P., Jones, C., Lollato, R. 2024. Evaluating the Utility of Weather Generators in Crop Simulation Models for In-season Yield Forecasting. Agricultural Systems. 220. Article 104082. https://doi.org/10.1016/j.agsy.2024.104082.
DOI: https://doi.org/10.1016/j.agsy.2024.104082

Interpretive Summary: Forecasting crop yields is essential for ensuring food security, especially with the unpredictability of climate change. This study examined three distant random weather generator tools designed to predict the weather resembling historical climatology and tested their performance in forecasting wheat yields. The evaluated weather generators include Global Weather Generator (GWGEN), WeatherGEN, and R Multi-Sites Autoregressive Weather GENerator (RMAWGEN). Our findings revealed that RMAWGEN and WeatherGEN were particularly good at predicting rainfall accurately, with RMAWGEN performing better in dry months and WeatherGEN excelling in wet months. Overall, RMAWGEN consistently outperformed the other tools in forecasting crop yields, showing superior capabilities with less error. These findings are essential for selecting the most reliable weather prediction tool, critical for planning strategies to mitigate fluctuations in food supply, stabilize market prices, and address food insecurity.

Technical Abstract: Crop yield forecasting is crucial for ensuring food security and adapting to the impacts of climate change, as it provides early insights into potential harvest outcomes and helps farmers and policymakers make informed decisions in the face of changing environmental conditions. The accuracy of the crop model–based yield forecasting frameworks is affected by the uncertainty in future weather data, substituted with synthetic weather realizations generated by stochastic weather generators. This study aims to evaluate the performance of three recent stochastic weather generators—Global Weather Generator (GWGEN), WeatherGEN, and R Multi-Sites Autoregressive Weather GENerator (RMAWGEN) for crop yield forecasting. We utilized historical weather data from Daymet as an input for the weather generator and for evaluating the performance of the generated weather realizations. Furthermore, the weather realizations generated by these weather generators across multiple winter wheat field sites in Kansas and Oklahoma were employed in the calibrated Environmental Policy Integrated Climate (EPIC) crop model to assess the potential impact of variations in weather generators on the accuracy of crop yield forecasts. RMAWGEN and WeatherGEN excelled in accurately simulating rainy days and precipitation amounts, with WeatherGEN particularly effective in wet months and RMAWGEN performing best in dry months, showcasing their proficiency in diverse weather conditions. RMAWGEN showed 21 to 26 percent lower error (MAPE) for precipitation compared to the other weather generators. For solar radiation, RMAWGEN exhibited a mean MAPE of 4.6%, whereas WeatherGEN and GWGEN showed higher errors, with mean MAPEs of 5.6% and 7.8%, respectively. RMAWGEN demonstrated the lowest MAPE for maximum temperature at 26%, while WeatherGEN and GWGEN showed lower accuracy at 38%; however, for minimum temperatures, RMAWGEN exhibited the lowest MAPE at 87%. Except for GWGEN, RMAWGEN and WeatherGEN demonstrate good agreement with Daymet in replicating spatial variability patterns. RMAWGEN consistently outperformed the other weather generators, particularly during the forecasting period. Consequently, it showed superior capabilities in forecasting crop yields with the RMSE of 0.87 Mg/ha and MAPE of 20.1%. The findings of this study are crucial for selecting accurate weather data estimates for crop yield forecasting. Utilizing alternative sources such as ensembles of multiple weather generators or outputs from sub-seasonal multi-model forecast systems may further enhance the accuracy of crop yield forecasts.