Fig. 1. Overview of the study area

Fig. 2. The cold years in Northeast China

Fig. 3. Comparison between simulated and observed variables of soybean

Fig. 4. illustrates the simulated profiles of soybean LAIs determined from 512 simulations. The LAI values vary considerably from 2 to 7, which is largely due to the fact that soybean growth processes are variable under different weather conditions and farming management systems. Note that the early and late windows (denoted by shaded rectangles in the figure) basically cover the critical growth period of each simulation. These simulation results thus show a broad range of variability in training the regression model given in Eq. (1).

Fig. 5. The coefficient of determination (R²) for the regression models

Fig. 5. Daily leaf area index outputs from 512 simulations of the CROPGRO-Soybean model (the shaded rectangles indicate the “early-season” and “late-season” windows used for image observations)

Fig. 6. The maximum LAI and its specific observation dates of early and late growing season windows obtained from Sentinel-2 in 2018 (a and c represent the maximum LAI and its specific observation dates of early growing season windows; b and d for late growing season windows, respectively)

Fig. 7. Estimated yields by calibrated CROPGRO-Soybean model under different cold injury scenarios (scenarios S1, S2 and S3 are set as reducing 3℃, 2℃ and 1℃ at the whole growth stages, respectively. Scenario N4 is set as actual weather. Scenarios S5, S6, S7 and S8 are set randomly as minimum temperature of 0℃ for 5 consecutive days during the periods from seedling emergence to flowering stage, flowering stage to pod bearing stage, and pod bearing stage to filling stage, and filling stage to maturity stage, respectively.)

Fig. 8. The comparison between actual and simulated yield losses

Fig. 9. Spatial distribution of estimated yields in normal year (2017) and cold year (2018)

Fig. 10. The spatial distribution of yield losses in 2018 relative to 2017