Fig. 1. Weighting functions of water vapor channel and infrared channel of GOES satellite, quoted from http：//cimss.ssec.wisc.edu

Fig. 2. U-net architecture used by generator of pix2pix. Dimensions of data are shown as (channels, width, height)

Fig. 3. Test error for experiments using 1 h satellite image intervals, pix2pix architecture, high resolution data, and without visible channels. (a) 40000 iterations; (b) 500000 iterations

Fig. 4. Comparison of wind speed retrieved by neural network and wind speed of NCEP/NCAR reanalysis. (a) 850 hPa; (b) 200 hPa

Fig. 5. Spatial distribution of MVD of wind retrieved by neural network. (a) 850 hPa; (b) 200 hPa

Fig. 6. Comparison of wind field retrieved by neural network and wind field of NCEP/NCAR reanalysis. (a) 850 hPa wind direction (streamline) and speed (shaded) retrieved by neural network at October 9, 2019 12:00 UTC; (b) same as (a), but with NCEP/NCAR reanalysis data; (c) 200 hPa wind direction (streamline) and speed (shaded) retrieved by neural network at October 9, 2019 12:00 UTC; (d) same as (c), but with NCEP/NCAR reanalysis data; (e)-(h) same as (a)-(d), but at January 26, 2018 18:00 UTC

Table 1. Results of wind field retrieve with different input satellite image intervals, channels, and neural network types

Table 2. Results of wind field retrieve with different resolutions of data

Table 3. Comparison of results between neural network and H8 AMV product