A Remapping Technique of FY-3D MWRI Based on a Convolutional Neural Network for the Reduction of Representativeness Error

The assimilation of spaceborne passive microwave measurements often suffers from representativeness errors due to the mismatch between the observation footprints and numerical weather prediction (NWP) model grids. In this article, a new brightness temperature remapping technique based on a deep convolutional neural network (CNN) is proposed to reduce the representativeness error in FengYun-3D (FY-3D) microwave radiation imager (MWRI) observation data assimilation. The remapping technique uses an adapted dataset construction method in which the training data consist of synthetic NWP-model-grid-based MWRI brightness temperature (
$T_{B}$
) images and synthetic MWRI-observed antenna temperature (
$T_{A}$
) images. The synthetic
$T_{A}$
and
$T_{B}$
, which are generated through radiative transfer for TOVS (RTTOV) model and MWRI degradation model, make the network learn the spatial remapping relationship between observation and background
$T_{B}$
in data assimilation. In addition, land–sea mask information is input into the CNN to help the network better analyze the coastline area data. The CNN-based remapped MWRI observation data are evaluated through observation minus background (OMB) diagnosis with the Global/Regional Assimilation and PrEdiction System (GRAPES) four-dimensional variational (4D-Var) system. The experimental results illustrate that the bias and standard deviation of OMB with the CNN-based remapped MWRI observation are quantitatively reduced compared with the raw measurements in GRAPES 4D-Var.