Research On Integrating Soil Moisture Based On Machine Learning To Improve Satellite Precipitation Accuracy

Precipitation is the main driving factor of runoff changes and an important part of the regional hydrological process.Flood forecasting mainly relies on hydrological models to simulate hydrological processes.Precipitation plays an important role in it.Therefore,it must have sufficient accuracy and real-time.However,precipitation has high temporal and spatial variability,so obtaining high-quality precipitation data is indispensable for flood control and disaster mitigation.Although the precipitation observation data of the surface rain gauges meets the accuracy and real-time,the uneven distribution and high sparseness of its distribution limit its further application in hydrological simulation.Meteorological satellites can provide a wider coverage of precipitation information,so two methods of precipitation observation and estimation,"top-down" and "bottom-up" have emerged.The former analyzes the cause of precipitation,such as the water content of the cloud top,obtains the instantaneous precipitation value,and the latter obtains the cumulative precipitation value by analyzing the results of precipitation,such as soil moisture.The two methods can be organically combined and used as a supplement to the ground precipitation.However,because the real-time satellite data has not undergone post-processing,there are system errors and data missing,so the satellite-rainfall station data fusion is carried out to improve the precipitation observation accuracy.There have been many studies on precipitation fusion,but most methods only consider single satellite data and do not effectively deal with the missing data.This paper proposes different machine learning fusion models,including a variety of neural network models: Multilayer Perceptron(MLP),Long Short-Term Memory(LSTM)and integrated learning models: Extreme Gradient Boosting(XGBoost),Light Gradient Boosting Machine(Light GBM),Categorical Boosting(Cat Boost),combined with Tropical Rainfall Measuring Mission(TRMM)3B42-RT satellite real-time precipitation,Advanced Scatterometer(ASCAT)-NRT satellite real-time soil moisture and rainfall station data,to obtain the optimal precipitation fusion model,and use the noise reduction encoder at the same time Two missing value filling models(Denoising Autoencoder,DAE)and Generative Adversarial Imputation Networks(GAIN)solve the problem of missing data.The experimental results on the data of 29 rainfall gauges in the Jingjiang River Basin show that:(1)After adding satellite soil moisture data,compared with only using satellite precipitation and rainfall station data,the accuracy of the integrated precipitation products can be significantly improved.(2)Through quantitative evaluation and comparison,it can be seen that the integrated learning model is better than the neural network model as a whole.Among them,the root mean square error and correlation coefficient of XGBoost are the best,while the average absolute error of Light GBM is the best.The fusion effect of these models shows that the fusion precipitation distribution obtained by Light GBM is more suitable for the real precipitation in the entire area,so it is used as the optimal precipitation fusion model for the study area.(3)The Light GBM precipitation fusion model has the best results after inputting the soil moisture filled with DAE,which can reduce the average absolute error and root mean square error of the original TRMM data in the study area by 17.95%and 23.76%,respectively,and increase the correlation coefficient to 0.773,an increase of 17.83%.(4)After fusion,the daily-scale precipitation data from 2007 to 2013 are obtained,which has higher resolution(0.1°)and accuracy in the Jingjiang River Basin,which can provide data support for areas lacking data.