Evaluation of machine learning-dynamical hybrid method incorporating remote sensing data for in-season maize yield prediction under drought.

Effective yield forecasting is a key strategy for adaptation when facing food loss to climate variability. Currently, solar-induced chlorophyll fluorescence (SIF) is an emerging remote-sensing index owing to its high relevance to plant photosynthesis, and sensitivity to drought. Despite many studies have focused on drought monitoring and production assessment by SIF, little puts it into practice for in-season yield prediction. In this study, we combined multi-source satellite and meteorological data, especially coupling with subseasonal-to-seasonal (S2S) dynamic atmospheric prediction climate model (IAP-CAS FGOALS-f2), with an addition of SIF, to predict maize yields in the U.S. Corn Belt, based on the developed machine learning dynamical hybrid model (MHCF). By comparison, we found that SIF performed well in the correlation analysis with yield, with average correlations up to 0.719 in August. Then we utilized different algorithms, different models (S2S data for MHCF, climate data for the Benchmark), and different input combinations to train and predict maize yields. All four algorithms using SIF significantly improved prediction performance. S2S + VIs + SIF combination (FGOALS-f2、NDVI、EVI、SIF) can achieve the best performance, while the XGBoost algorithm reached 0.897 of R2. With the best combination, it can achieve 4 months before maize harvest (with R2 value of 0.85, and RMSE < 13 bu/acre). In 2012, the year had a severe drought, although predictive capability decreased in all the predictions, the models with SIF still maintained robust and improved the prediction (improved R2 by 5.92%, and RMSE decreased by 18.08% of XGBoost). According to the study, it can be expected, the combination of MHCF and SIF will play a greater role in subseasonal yield prediction. We also provide an operational proposition of hybrid yield forecasting method to fully integrating climate prediction and machine learning for early notice of crop production losses. [ABSTRACT FROM AUTHOR]