| Accurate and real-time precipitation identification and estimation are of great significance to hydrometeorological monitoring,flood disaster prevention,and environmental change management.With the development of global meteorological satellite observation technology in the direction of automation,intelligence,and network coordination,meteorological observation data is growing rapidly.How to efficiently extract rich precipitation information from massive meteorological satellite observation data is a major challenge for improving the capability of satellite precipitation identification and estimation.The middle and lower reaches of the Yangtze River have unique geographical location and climatic conditions,and the precipitation factors in the rainy season are also intricate and varied.To improve the accuracy and timeliness of identifying and estimating precipitation in this region,the key is to establish the nonlinear relationship between high spatial-temporal resolution meteorological satellite observation data and precipitation.This paper aims to develop algorithm for precipitation identification and estimation applicable to the middle and lower reaches of the Yangtze River by using data from the new generation of geostationary meteorological satellite FY-4A and deep learning technology.The main research of this paper is as follows:(1)A precipitation cloud identification model based on multi-feature fusion is proposed.To address the problems of insufficient learning of precipitation cloud features in satellite cloud imagery and easy omission of small areas,this method uses dilated convolution to perform multi-scale sampling on deep features in order to better capture the spatial correlation of precipitation clouds.An attention mechanism is added during upsampling to better focus on precipitation areas.In addition,visible light and near-infrared channel information is added to assist in improving the accuracy of daytime precipitation cloud identification.Experimental results show that this method can effectively capture precipitation cloud information and spatial distribution details in satellite cloud maps.The addition of features improves the accuracy of daytime precipitation cloud identification,with a Probability of Detection(POD)and Critical Success Index(CSI)reaching 0.651 and 0.554,respectively,and the identification of precipitation cloud edges becomes smoother.(2)A precipitation estimation model based on an improved Unet,named UPENet(Unet based Precipitation Estimation Network),is proposed.To address the issue of uneven distribution of rainfall samples in the Meiyu season in the middle and lower reaches of the Yangtze River,this method uses a convolutional neural network to learn spatial features in satellite cloud images that are closely related to precipitation.Residual modules are employed between the encoder and decoder to avoid the problems of gradient vanishing and exploding.Furthermore,a spatial channel attention mechanism is introduced to increase the weight of precipitation regions and suppress feature responses in non-precipitation regions.Experimental results demonstrate that UPENet can effectively characterize the spatial distribution of precipitation,with mean absolute error(MAE)and correlation coefficient(CC)of 0.432 mm/h and 0.592,respectively,which outperforms both the baseline precipitation estimation models and gridded precipitation products.(3)A methodology applicability analysis was conducted for the extreme rainstorm event of 20 July in Henan Province.Applying the precipitation cloud identification and estimation method to precipitation cases in the central and northern regions of Henan Province,this method is able to effectively identify the spatial distribution of precipitation clouds and the location of precipitation extremes compare to other models and precipitation products.The estimated extremes are close to the reference truth,indicating that this method is suitable for real-time monitoring of precipitation under extreme weather conditions. |
PDF Full Text Request