Integrating prior information for improving 3D model-driven GAI estimation with application to wheat crops

Green Area Index (GAI) is a key crop trait obtained through remote sensing with wide applications in agriculture. Although 3D model-driven approaches to retrieve GAI from multispectral reflectance observations are appealing, they are constrained by limitations in the realism of simulated datasets used for training. This study comprehensively explored how to integrate prior information—such as soil background, leaf optical properties, and canopy structure—into radiative transfer models to improve GAI retrieval. A suite of models (MARMIT-2 for soil reflectance, PROSPECT for leaf optical properties, ADEL-Wheat for canopy structure, and LESS for radiative transfer) was employed to generate five simulation datasets incorporating different combinations of prior information. Support Vector Regression (SVR) models were independently trained on these simulated datasets and validated against an extensive data set made of 310 samples of GAI ground measurements and the corresponding SuperDove satellite data. Our results show that stage-specific GAI retrieval integrating detailed prior information on soil and leaf properties (R 2 = 0.93, RMSE = 0.47) notably outperforms standard model inversion approaches (R 2 = 0.82, RMSE = 0.73). The improved realism of the training dataset stems from three key strategies was discussed in detail including: (1) employing models that integrates physical and biological knowledge; (2) narrowing the training space; and (3) minimizing distribution shifts. While this study focused on GAI estimation for wheat crops using SuperDove observations, the findings can be extended to other crops, vegetation variables, and satellite systems. • Integrating prior information on soil, leaf, and canopy properties significantly improved GAI retrieval accuracy (R 2 = 0.93, RMSE = 0.47). • Three key strategies—incorporating physical-biological models, narrowing the training space, and reducing distribution shifts—enhanced simulation realism. • The scalable methodology offers potential for application to various crops, vegetation variables, and satellite systems.