Technique of CO2 concentration profile retrieving from satellite data by means of neural network approach Kataev M.Yu., Andreev A.G.
Goals of work: To develop the scientific computer program for modeling GOSAT NIR spectrums and their processing for the purpose of retrieving of the CO2 concentration profile by neural-networks method and total content by the differential absorption method. 2
The main trajectories of the solar beams in the system « earth + atmosphere ». Green is the reflected from surface solar radiation Blue is the scattering in atmosphere solar radiation 3
Reflected and scattering parts of the signal I 2 ( λ ) I 1 ( λ ) 2 τ Θ 0 τ Θ 0 τ atmosphere 1 1 Θ 0 I 1 ( λ ) =E 0 τ Θ 0 cos( Θ 0 ) ρ λ τ r λ ρ ( λ ) 4
Radiative transfer (RT) (F-M. Breon, S. Boffies J.Appl.meteorol., 1996, Jan, pp. 69-76) P ⎧ ⎫ 2 1 ⎪ ⎪ T ( , , P , P ) exp ( ( , p ) ( , p ) ( , p )) dp λ µ = − ∫ α λ + α λ + α λ ⎨ ⎬ 1 2 gas aer mol µ ⎪ ⎪ ⎩ P ⎭ 1 S ( ) λ 0 I ( ) R ( ) T ( , , 0 , P ) T ( , 1 , 0 , P ) λ = ⋅ λ λ µ λ 1 surf sun surf surf π P S ( ) 1 surf λ I ( ) 0 ( ( , p ) F ( ) ( , p ) F ( )) λ = ⋅ ∫ α λ γ + ψ α λ γ ⋅ 2 mol mol aer aer aer π µ sun 0 T ( , , p , P ) T ( , 1 , p , P ) dp λ µ ⋅ λ sun surf surf 1 / 2 cos( ) [( 1 )( 1 )] cos( ) γ = − µ µ − − µ − µ ϕ sun sat sun sat 5
Databases for spectra calculation in global world l 1 Solar spectra, Kurucz spectral resolution 1 cm-1, l 2 Spectroscopic information HITRAN2004 spectral parameters database, l 3 Meteorological parameters – NCEP Reanalysis data 6 hours 2.5x2.5 grad, l 4 Gases concentration profiles H2O – NCEP data, CO2 – NIES (S.Maksyutov) data l 5 Aerosol extinction – SPRINTARS (Toshihiko Takemura,RIAM,Kyushu Univ.) 6
Variations of the CO2 concentration profile for West Siberia region
Data from NCEP with comparison of AFGL climatological model
Example of transmittance calculation for Channel 2 of GOSAT
SEARCH OF THE OPTIMAL SPECTRAL CHANNELS 1 T ( , ) T ( , ) ⎡ − ⎤ ν µ ν µ 1 o 2 L ( , , ) 1 ν ν µ = ⋅ ⋅ ⎢ ⎥ 1 2 T ( , ) T ( , ) σ ν µ ν µ ⎣ ⎦ 2 o 1 where - dispersion of the measurements, - transmittance for all components of air T ( , ) σ ν µ 1 , 2 T ( , ) besides gas under investigation, - transmittance (1 – lie in absorption lines of gas under ν µ o 1 , 2 investigation and 2 – very small absorption of gas under investigation). 1.0 0.8 Transmittance 0.6 CH 4 0.4 All 0.2 OFF ON 0.0 6040 6042 6044 6046 6048 6050 6052 6054 6056 6058 6060 Wavenumber, cm-1
Results of searching of the CO2 spectral channels The 753 wavenumber with help of algorithm when selection criterion for retrieving error =0.1% was found.
Clusterization of the CO2 data The clusterization is based on close structure analysis of matrix of the distances between objects. The algorithm is realized using of performance of graph theory.
Neural network approach for inverse task solving n Mathematical model of the neuron y f ( g ) f ( w x w ) ∑ = = + i i 0 i 1 = w 0 x 1 a 1 x 1 w 1 x 2 a 2 x 2 f (g) w 2 y ∑ x n a n x n w n n The neural network of one layer; f(g) is the activation function; inputs: х 1, х 2, ..., х n; weight factors: w11, w12, …, outputs: y
Scheme of Neural network in our work
Maximal relative error on training of NN for retrieving of the CO2 concentration profile
Resuls of retrieving of the CO2 concentration profile from test samples (A) and training samples (B). B A
Relative error of CO2 concentration profile retrieving with changes of the aerosol optical thickness
l Technique of CO2 concentration profile retrieving from satellite data by means of neural network approach Kataev M.Yu., Andreev A.G. Tomsk state university of control systems and radioelectronics
Result of the CO2 total amount retrieving with help of Neural Net
Recommend
More recommend
