Downscaling Of TRMM Precipitation Products Based On Deep Feedforward Neural Network

The study of the seasonal spatial and temporal distribution of precipitation is of great significance to the ecological protection and agricultural production in Northeast China.In order to obtain precipitation distribution with high spatial and temporal resolution,it is necessary to downscale TRMM products with lower resolution.Therefore,based on the correlation of vegetation index,terrain factor and precipitation,this paper uses three algorithms of deep learning,multiple linear regression and random forest to build a model to reduce the average of TRMM 3B43 products from 2009 to 2018 in October,April,July and October Scale to 0.01°?approximately 1km?,and fill the area above 50°N that is not covered by TRMM,obtain the spatial and temporal distribution of seasonal precipitation in the northeast region.Finally,this paper uses site measured data for accuracy correction,and studies the accuracy and applicability of downscale models to predict precipitation nationwide.The main conclusions of the study are as follows:?1?The fitting and accuracy of the deep learning downscale model to predict precipitation relative to the measured precipitation is better than multiple linear regression and random forest,which can effectively obtain the precipitation distribution with higher spatial resolution and accuracy in each season in the northeast region.The overall fit and accuracy performance of the model Better?R²=0.814?0.933,RMSE=2.389mm?16.98mm,MRE=10.16%?77.22%?.After the correction,the global precipitation accuracy is further improved?R²=0.853?0.956,RMSE=0.931mm?16.20mm,MRE=7.782%?31.91%?,in which the correction effect of the ordinary Kriging interpolation method based on Gaussian function is better than the IDW interpolation method,The correction results can more accurately reflect the spatial distribution of precipitation in the study area.?2?The prediction accuracy of the deep learning downscaling model in the northeast region has a seasonal difference.The simulation effect in April and October is better_month7=0.814).The downscaling model was used to fill the areas above 50°N that were not covered by the original TRMM satellite data.It was found that from January to October,the determination coefficient R²showed an increasing trend.The simulation accuracy in April,July and October was relatively good.The worst month.A study of the model accuracy of the Northeast Plains found that compared with the global accuracy of the Northeast Plains,the decision coefficient R²of the Northeast Plains before and after the downscaling in January improved overall,and the accuracy before and after the downscaling in July decreased slightly.?3?Compared with the accuracy of the TRMM data,the prediction accuracy of the precipitation forecast of each geographical division model across the country has been greatly improved?the average increase of R²is 23.4%,the average reduction of RMSE and MRE is19.0%,16.6%?.The applicability is good in central and southwestern regions?R²=0.710?0.925?,and the model has good performance and reliability in fitting,deviation and relative error in each month;the seasonal precipitation of models in eastern,southern and northeastern regions The difference is large,and the applicability of precipitation in the North China and Northwest regions is relatively poor.Among them,the model fitting degree of each month in North China is average,and the model error in Northwest China is large.