Estimation of canopy water content for wheat through combining radiative transfer model and machine learning

Canopy water content (CWC) is an important indicator of crop growth. Currently, the hybrid approach, which combines the physical model and machine learning (ML) method, is commonly used for inversing CWC at the regional scale via optical imagery. Although this approach has good inversion accuracy, it suffers from ill-posed issues from the physical model due to the difference between actual and simulated scenes. Meanwhile, the ability of existing ML methods to resolve this difference remains unclear. To fill the above study gaps, we added different degrees of Gaussian noise into the simulated dataset to reduce the differences in spectral reflectance between that simulated by the PROSAIL-5B model and that measured from optimal imagery. Furthermore, this study also compared the performance of different ML approaches (neural network, NN; Gauss process regression, GPR; and kernel ridge regression, KRR) in the retrieval of wheat (Triticum aestivum L.) CWC based on PROSAIL-simulated datasets. Additionally, we also compared the inversion results for the CWC generated by the SNAP Toolbox with those estimated by the hybrid approach. The results showed that the aforementioned hybrid approaches performed well in the retrieval of wheat CWC when Gaussian noise with a mean of zero and a standard deviation of 0.06 (η: 0, 0.06) was added to the simulated dataset. The hybrid model based on GPR had the highest accuracy of estimation for the CWC (RMSE = 0.0096 g/cm2, RRMSE = 22.2%), when compared with NN or KRR (NN: RMSE = 0.0105 g/cm2, RRMSE = 26.9%; KRR: RMSE = 0.0099 g/cm2, RRMSE = 24.3%). It also outperformed the existing CWC-retrieval algorithms among the SNAP Toolbox (RMSE = 0.027 g/cm2, RRMSE=44.1%). Our results demonstrate that the hybrid approach in combination with GPR and the physical model can accurately retrieve the CWC of crops, thus confirming that the approach is suitable for estimating the CWC.