ARYA Operational Yield Forecasting Model at the National and County Level
The map above was produced by Dr. Natacha Kalecinski using USDA National Agricultural Statistics Service (NASS) data. It shows the yield anomaly in 2022, for maize (corn) at the county-level across the United States, compared to the 21 year average from 2001-2022. The colors indicate the range of the anomaly from minus -80% (blue), indicating below average, to +80% (red), indicating above average.
Project Summary
This study aims to compare and contrast optical and SAR data value to monitor crop yield across the USA and Europe for different spatial scale. Building on the robust performance of DVI, we developed an operational yield forecasting model, ARYA, that uses MODIS and VIIRS data, as well as growing degree days from the ERA5 reanalysis. ARYA is designed to provide frequent and accurate crop yield forecasts throughout the growing season, both at national and sub-national level. Since 2020, ARYA provides weekly yield and production forecasts for winter wheat, corn, soybeans, and sunflowers. The model has been successfully calibrated over a 20-year period and consistently achieves high accuracy, with a coefficient of determination (R 2 ) greater than 0.7 and a root mean square error (RMSE) ranging from 5% to 14%. Since MODIS is expected to be decommissioned in 2023, VIIRS data are being considered as a replacement. Our study investigated the impact of this transition, and the results are promising. VIIRS achieved an error rate of 7% compared to 6% with MODIS over a ten-year period for winter wheat estimation in Ukraine. This transition ensures the continuity of yield forecasts and maintains their accuracy. Local scale yield forecasts show a decrease in R 2 and coefficient of variation. We conducted an evaluation using high-resolution surface reflectance from HLS and Planet data, to estimate crop yield at field scale in Iowa, USA. We examined which optical indices were most predictive. Despite achieving a coefficient of determination greater than 0.8, optical indices saturate for high yields and cannot robustly capture the variability. Additionally, optical indices only offered relatively short forecast windows, with the highest correlation occurring in the middle of the growth cycle. Estimating yields as early as possible in the season remains a major challenge. To address this issue, we evaluated the effectiveness of synthetic aperture radar (SAR) data compared to optical satellite data for crop yield monitoring. We examined various sensors, including Sentinel-1, RadarSat-2, and L-band UAVSAR, and their ability to explain variations in crop yields (corn, soybean and rice) between fields in Arkansas, USA, in 2019. We developed a Random Forest regression model that combined SAR parameters with DVI, resulting in a 50% improvement in yield predictions for all growth stages. However, optical and SAR models still struggle to effectively explain high-value crop yields, consistent with previous research. Their studies will help us gain better insights on how trends and extreme weather influence crop yield across the U.S.
Study Area:
United States
Earth Observations Used
MODIS | VIIRS | ERA 5
Lead Institution
Project Lead
Sergii Skakun
University of Maryland
Natacha Kalecinski
University of Maryland