Comparison of different machine learning approaches for tropospheric profiling based on COSMIC-2 data

Earth, Planets and Space

Table 1 Fitting results obtained on the training and test data sets

		Artificial Neural Network				Random Forest				CDAAC wetPf2	ERA5
		Bending angle		Refractivity		Bending angle		Refractivity		CDAAC wetPf2	ERA5
		L	T	L	T	L	T	L	T
Temperature (K)	RMSE	1.49	1.56	1.46	1.51	1.44	1.68	1.47	1.60	1.50	STD	4.82
Pressure (hPa)		1.33	1.34	1.20	1.22	1.10	1.42	1.12	1.26	1.05		4.23
Water vapor pressure (hPa)		0.44	0.44	0.43	0.43	0.43	0.48	0.43	0.46	0.45		0.93

Vertically averaged root mean square errors (RMSE) for the temperature, pressure, and water vapor partial pressure between ERA5 and obtained using different machine learning models on the training and test data sets indicated by columns ‘L’ and ‘T’, respectively or 1DVar approach stored in operational CDAAC wetPf2 product. The right column presents vertically averaged standard deviations for the corresponding meteorological parameters calculated from ERA5 reanalysis