Improving regional wheat drought risk assessment for insurance application by integrating scenario-driven crop model, machine learning, and satellite data

Accurate estimation of yield losses from natural disasters on a regional scale can guide agronomic management and agricultural insurance, transfer disaster risk, and ensure food security. Conventional yield losses, however, mainly depend on historical events, for which detailed records of locations and losses are unavailable.The development of a disaster vulnerability model has consequently been hindered by the lack of sufficient samples. To improve regional yield estimates, a novel method for estimating yield losses and the pure insurance rate is thus strongly needed.Using 129 major wheat-producing counties on the North China Plain as an example, we developed a drought assessment system based on crop growth modeling, machine learning, and satellite data. The initial model used was the model to simulate the crop–weather relationship over a large area (MCWLA) applied to wheat (MCWLA-Wheat), which was then calibrated by multi-step-assimilation with multi-source data. We first established various drought scenarios to simulate the impacts of drought at different growth stages on wheat yield by driving the calibrated MCWLA-Wheat.According to our results, drought-sensitive stages of wheat varied by drought severity, with insufficient water supply during the Greenup–Heading stage having the most significant impact on grain yields. Based on the outputs of the simulation and three drought indicators—standardized precipitation index (SPI), standardized soil moisture index (SSMI), and relative leaf area index (RLAI), we then established vulnerability models coupling the MCWLA-Wheat with statistical models (random forest [RF] and multiple linear regression [MLR]). The vulnerability model MCWLA+RF showed higher accuracy, with a root mean square error (RMSE) of 6% (i.e., 365 kg ha⁻¹), relative to that of the MCWLA+MLR (RMSE = 1175 kg ha⁻¹). Ranked in descending order, the relative importance of drought indicators at Greenup–Heading was SPI, SMMI-20 cm, and RLAI.Our study has highlighted a potential method for quantifying crop disaster risk—integration of a scenario-driven crop growth model, machine learning, and satellite data—that can be applied to other crops and regions.