Physics-informed and prior-data driven: A hybrid Stacking framework for rice growth parameter inversion

The PROSAIL-Dw model was developed for simulating the rice canopy reflectance in scenarios that the underlying surface was covered by a water layer. C-Vine Copula-based parameter simulation method could retain inter-parameter correlations, generating data more consistent with actual rice growth status. The proposed hybrid inversion framework significantly enhanced inversion accuracy and reliability for LAI, AGB, CCC, and CNC. Accurate monitoring of rice growth status is essential for scientific water and fertilizer management in paddy fields. Using remote sensing data combined with radiative transfer models and artificial intelligence algorithms can realize the semi-mechanism inversion. However, the commonly used hybrid inversion models have difficulties in adapting to paddy field scenarios covered with water layers. In addition, the data simulation methods often ignore the correlations between parameters, leading to distortion of the simulated data. To address these challenges, by developing the PROSAIL-Dw model considering the influence of the underlying surface moisture state on the canopy reflectance and proposing a multivariable joint prior knowledge data simulation method based on C-Vine Copula, this study proposed a novel hybrid framework based on Stacking model for retrieving rice growth parameters from multispectral imagery. The results indicated that, by introducing two parameters reflecting the presence and depth of the water layer, the PROSAIL-Dw model can more accurately simulate the NIR reflectance with water layer coverage (with R ² increased by 0.42 for low nitrogen treatment). The growth parameters simulated by the C-Vine Copula method could retain the correlations, thus effectively improving the accuracy of the Stacking model compared with conventional methods (with rRMSE decreased by 5.81%-15.00%, and R² increased by 0.19-0.30). The hybrid inversion framework constructed in this study has further improved the accuracy and reliability of rice growth parameter inversion, and has important practical value for the scientific management of water and fertilizer in early-stage paddy fields.