Assessment model for low temperature stress in wheat based on RGB and hyperspectral images

【Objective】 This study investigated the effects of low temperature stress on chlorophyll fluorescence parameters, RGB image parameters and hyperspectral indexes of wheat leaves. It aimed to establish an assessment model for wheat under low temperature stress, providing reference for disaster prevention and mitigation in wheat production.【Method】 Using winter wheat variety Jimai 22 as the research material, a controlled low temperature stress experiment was conducted during the jointing stage. Three treatments were applied:daytime（8:00–20:00）/nighttime（20:00–8:00on next day） mean temperatures of 8 °C/0 ℃（T1）, 6 °C/-2 ℃（T2） and 4 °C/-4 ℃（T3）, each lasting for three days.Potted wheat grown under natural field conditions（23 °C/8 ℃） served as the control（CK）. The changes in chlorophyll fluorescence parameters, RGB image parameters, and hyperspectral indexes of wheat leaves were analyzed on 1, 3 and 6 d after the low temperature stress treatments. Wheat low temperature stress assessment models were developed using univariate linear regression, random forest（RF） and artificial neural networks（ANN）. 【Result】 The chlorophyll fluorescence parameter DI_o/RC was identified as an effective indicator for assessing wheat low temperature stress. In the univariate linear regression model, the model using the enhanced vegetation index（EVI） performed the best, with a regression equation of y=-1.261x+1.401, yielding a coefficient of determination（R²）, root mean square error（RMSE）, mean absolute error（MAE） and mean relative error（MRE） of 0.536, 0.058, 0.045, and 11.31% respectively. For RF and ANN models, models based on RGB image parameters outperformed those based on hyperspectral indexes. The RF model achieved R², RMSE, MAE and MRE values of 0.771, 0.042, 0.033 and 8.57% in the test set, with an R² improvement of43.78% and reductions in RMSE, MAE and MRE by 28.31%, 28.06% and 24.21% respectively compared to the univariate linear regression model. The ANN model achieved R², RMSE, MAE and MRE values of 0.742, 0.046, 0.037 and9.01%, with R² improving by 38.34% and RMSE, MAE and MRE decreasing by 20.33%, 18.06% and 20.32% respectively compared to the univariate linear regression model. 【Conclusion】 The RF model based on RGB image parameters demonstrates the best performance and the highest accuracy, making it suitable for evaluating low temperature stress in wheat.