Winter wheat yield prediction at a county scale using time series variation features of remote sensing spectra and machine learning

Abstract

To meet the challenges of population growth, decreasing arable land, and food demand, accurately predicting winter wheat yield is of great significance for national food security. Time series remote sensing data can continuously track the spectral changes during the crop growth cycle, providing a wealth of data to support yield prediction. However, relying solely on raw time-series vegetation index data is often insufficient to accurately reflect the crop's growth status and physiological changes, increasing the uncertainty of prediction results. To address this issue, based on the time series normalized difference vegetation index (NDVI) and normalized difference water index (NDWI) of winter wheat during the whole growth period, this study extracted time series variation features of spectral (dVI), such as the rate of NDVI change during the filling period and the mean NDWI during the growth season of winter wheat, etc. Then, we compared a traditional multiple linear regression (MLR) method and two machine learning (ML) methods, including random forest (RF) and Extreme gradient boosting (XGBOOST), to predict county-level winter wheat yield. The results showed that the prediction accuracy of county winter wheat yield in 2020–2021 was best using the dVI and RF model. R², RMSE and CCC were 0.67, 644.93 kg ha⁻¹, 0.80, respectively. Compared with the original time series VIs, the method of combining dVI with RF has been proved to be more reliable and robust. This highlights the importance of incorporating spectral variation features into yield prediction analysis to enhance precision and reliability. Our study illustrated that the combination of ML and dVI is an effective and promising method for crop yield estimation.