Improved prediction of rice yield at field and county levels by synergistic use of SAR, optical and meteorological data

Abstract

Timely and accurate rice yield prediction over large regions is imperative to making informed decisions on precision crop management and ensuring regional food security. Previous studies on crop yield prediction mainly used of either optical or synthetic aperture radar (SAR) data alone, and few of them investigated the synergistic use of these two data sources. It remains unclear whether and how synergistic use of optical and SAR imagery would improve crop yield prediction over a large region. In addition, various machine learning (ML) algorithms have been used in regional yield prediction, but the existing findings are mostly inconsistent about the optimal ML algorithm that prevails others across years and spatial levels. To solve these issues, this study designed a meta-learning ensemble regression (MLER) framework and developed an efficient method of integrating multi-source data for accurate prediction of rice yield at field and county levels over Jiangsu Province of China.The results demonstrated that SAR data from Sentinel-1 could compensate for the lack of optical data from Sentinel-2 for rice yield prediction, which was further improved by adding the meteorological data. The MLER algorithm exhibited the highest field-level prediction accuracy for Xinghua and Suining (R2 = 0.89, RMSE = 0.54 t/ha), as compared with individual base algorithms (RF: R2 = 0.80, RMSE = 0.72 t/ha; XGBoost: R2 = 0.76, RMSE = 0.80 t/ha; SVR: R2 = 0.74, RMSE = 0.83 t/ha) and the long short-term memory (LSTM) (R2 = 0.76, RMSE = 0.80 t/ha) with a ten-fold cross-validation. Furthermore, integrating the MLER algorithm with the best-performing data combination could reduce the influence of regional phenological variability for building yield prediction models applicable to the entire province. The MLER models for predicting the rice yield across the province exhibited robust performance in the leave-one-year-out accuracy assessment at both levels (field level: R2: 0.50∼0.61, RMSE: 0.85∼1.22 t/ha; county level: R2: 0.42∼0.67, RMSE: 0.23∼0.27 t/ha) and could be transferred to an extreme-heat year (2022) (R2 = 0.49, RMSE = 1.52 t/ha). This research opens up new possibilities for predicting the rice yield of smallholder fields over large regions at multiple levels by integrating publicly available multi-source data and machine learning.