Improved downscaled inversion of soil moisture based on the BGKS methodology

Accurate prediction of soil moisture can help optimise irrigation decisions, reduced labor intensity, maintain hydro-ecology, and help monitor and prevent disasters. In the current the SM(Soil moisture) inversion studies, most of them suffer from a single method, low resolution of inversion models, difficulty in predicting a large range of SM, and poor robustness and generalisation. Downscaled SM inversion, as a method that can improve the prediction resolution, mostly relies on a single machine learning or a single mathematical derivation model in existing studies. To address these issues, and enhance the precision of soil moisture measurements, we advocate for a BGSKs downscaling inversion approach, construct a BG-Keras based SM estimation model by combining the predicted values of BP (Back Propagation)and GBDT(Gradient Boosting Decision Tree) with in-situ SM measurements, simultaneous. The results confirmed the superior performance of our proposed BG-Keras model compared with other models. The correlation coefficient (R) of BG-Keras reached 0.9716 at the optimum. the mean absolute error (MAE) as well as the root mean square error (RMSE) decreased to 0.0212 ∼ 0.0513 and 0.0335 ∼ 0.0592, respectively. And we obtained the Tibetan Plateau Soil moisture predictions with a resolution of 1 km and SM maps with a resolution of 36 km were obtained, and the accuracy of the predictions was explored on the basis of the inversion results. The results show that the BG-Keras fusion model effectively solves the problem of low prediction accuracy of SMAP SM inversion at low scales, and has high accuracy and ensures the spatial and temporal consistency of prediction.