Effective estimation of crop yields at a regional scale holds significant importance in facilitating decision-making within the agricultural sector, thereby ensuring grain security. However, traditional ground-based measurement techniques suffer from inefficiencies, and there exists a need for a reliable, precise, and effective method for estimating regional rice yields. In this study, we employed four machine-learning techniques: partial least squares regression (PLSR), support vector regression (SVR), random forest regression (RFR), and back propagation neural network (BPNN). We combined these methods with multi-temporal rice NDVI (normalized difference vegetation index) data for rice yield estimation. Following an accuracy evaluation and a spatial analysis, the key findings of our study are as follows. (1) The RFR model emerged as the most accurate for rice yield estimation, achieving an R2 of 0.65, an RMSE of 388.79 kg/ha, and an rRMSE of 4.48%. While PLSR and SVR demonstrated comparable accuracy, they were both inferior to RFR. (2) Using the top seven predictors with the highest importance rankings as inputs for the RFR model (NDVI values on the 6th, 17th, 33rd, 44th, 71st, 90th, and 106th days after the rice transplanting stage) achieved comparable accuracy while reducing information redundancy. (3) The proposed model demonstrated good spatial applicability (MI = −0.03) for rice yield estimation in Jiangsu, China. (4) A high spatial resolution yearly rice yield dataset (1 km) spanning from 2001 to 2020 was generated using the proposed model and is accessible on the Zenodo database. In conclusion, this study has demonstrated the efficacy of combining multi-temporal remote sensing data with machine-learning techniques for accurate rice yield estimation, thereby aiding agricultural authorities and production enterprises in the timely formulation and refinement of cropping strategies and management policies for the ongoing season.
Read full abstract