Random forest model that incorporates solar-induced chlorophyll fluorescence data can accurately track crop yield variations under drought conditions

Abstract

Timely and reliable crop yield estimation is vital for ensuring both global and regional food security. Previous studies have primarily used process-based crop models or statistical regression-based models for crop yield estimates. However, these model types possess limitations, particularly in accounting for specific extreme climate events that occur during the growth stage. In this study, remote sensing data, climate data, and soil moisture data from the winter wheat growth period in northern China from 2003 to 2017 were used to construct a crop yield simulation model based on the Random Forest (RF) algorithm. The effect of drought on winter wheat yield was quantitatively evaluated by calculating the fitting accuracy of the RF model, analyzing the importance of the factors influencing yield simulations, identifying a typical drought event, and determining the yield estimation accuracy as well as the percent yield loss (PYL) under drought conditions. The results indicated that solar-induced chlorophyll fluorescence (SIF) could characterize drought stress on winter wheat yield. The fitting accuracy of the RF yield simulation model was relatively high (R² = 0.72). Among all climate factors, SIF, enhanced vegetation index, and soil moisture were significant factors affecting wheat yield, exerting greater effect than those of all other climate factors. Furthermore, 2011 was identified as a typical drought year in the winter wheat area of northern China. The RF model simulated the accuracy of winter wheat yield for 2011 with an R² of 0.80. The RF model simulation revealed that the yield simulation accuracy of winter wheat under drought conditions was 90.64 %. The mean simulated PYL due to drought was 5.6 %, aligning closely with the mean actual PYL of 6.1 %. This suggested that the RF model was feasible for simulating crop yields and tracking yield variations by incorporating environmental variables, especially SIF data, under drought conditions.