Exploring the superiority of solar-induced chlorophyll fluorescence data in predicting wheat yield using machine learning and deep learning methods

Abstract

Reliable forecasts of large-scale wheat yield are very important for global food security. Although solar-induced chlorophyll fluorescence (SIF) is more sensitive than traditional remotely sensed vegetation indices to photosynthetic capacity, the performance of SIF in wheat yield prediction should be further explored. In this study, five satellite variables (i.e., Global Ozone Monitoring Experiment-2 (GOME-2) SIF at 0.5° spatial resolution, global spatially contiguous SIF (CSIF) at 0.05° resolution, and three vegetation indices at 1 km resolution) from 2007 to 2018 were used to predict wheat yield using two linear regression methods (least absolute shrinkage and selection operator regression (LASSO) and ridge regression (RIDGE)), three machine learning methods (support vector regression (SVR), random forest regression (RF), and extreme gradient boosting (XGBoost)), and one deep learning method (long short-term memory (LSTM)) to predict wheat yield across the Indo-Gangetic Plains. The results showed that machine learning and deep learning methods outperformed the two linear regression methods in predicting wheat yield, while the LSTM did not perform better than SVR. The prediction using the high-resolution SIF product had better performance than that using the coarse-resolution SIF product among all years. Moreover, the high-resolution SIF had better performance than the three vegetation indices in yield prediction in 2010, which indicated that the SIF data had great superiority in predicting wheat yield under extreme weather events. Our findings highlight that developing high-quality SIF products in the future has the potential to improve crop yield predictions, and our method can predict wheat yield simply and effectively in cropping areas with limited data.