Accurate and rapid regional-scale crop yield estimation can provide effective data support for the formulation of national food security policies. Compared with complex mechanism models, sampling statistical surveys and empirical models based on multi-source data have better reliability and operability for county-level or city-level yield estimation. Previous studies have proposed many factors related to winter wheat yield, but systematic research on the selection and analysis of multi-source factors is lacking. On the basis of remote sensing, meteorological, and statistical data, this study systematically explored the influence of key-phase growth-environment-landscape features on winter wheat yield estimation at the county-level and determined the best time phase and characteristic parameters.The considered features included crop condition, environmental forcing (e.g., precipitation, light, and temperature), and farmland landscape. The key phases were the key periods of winter wheat yield formation (P1—P5), which were extracted from the NDVI curve of the crop growth process. Random Forest (RF) regression models were developed using different combinations of phases and features to simulate statistical wheat yield data and evaluate the importance of different combinations via an accuracy assessment. The performance of the models built from each layer combination was compared using the Mean Relative Error (MRE), Root Mean-Squared Error (RMSE), Normalized Root Mean-Squared Error (NRMSE), and coefficient of determination (R²). Data on years 2014—2017 were used to build the models, and 2018 data were utilized for validation.Results showed that P2, P3, and P4 resulted in higher accuracy than P1 and P5 in terms of single phases. The model accuracy using multi-phase features was higher than that of using single phases, and the combination of P2 and P4 was the best. Among all the features, crop growth features had the greatest impact on yield estimation accuracy, whereas the addition of environmental forcing factors (e.g., water, light, and temperature) did not significantly improve the accuracy. The addition of farmland landscape features could effectively improve the accuracy of yield estimation. Moreover, five important features (PROP, NDVI_P2, B2_P2, ED, and B1_P4) were selected, and a yield estimation model was established to obtain the county-level yield of winter wheat in Hebei Province. The MRE of wheat yield estimation at the county level in 2018 was as low as 2.85%, and the RMSE, NRMSE, and R² were 253.25 kg/ha, 4.09%, and 0.83, respectively.Conclusion Multi-phase performance is better than single-phase performance. Combining crop growth features with farmland landscape features (RMSE of 247.79 kg/ha) provides more accurate estimates than using crop growth features alone (RMSE of 295.95 kg/ha). Furthermore, the RF model produces good yield estimation results. This study provides insights into and new methods for nationwide estimation of winter wheat yield at the county level.

remote sensing;yield estimation;winter wheat;statistical data;NDVI;Random Forest;Hebei Province