Abstract: A well-known approach for remote sensing inversion of the mean dry air mole fraction of methane(XCH₄)is weighted function modified differential optical absorption spectroscopy(WFM-DOAS). The ability to distinguish between the spectral structures of "broadband absorption" and "narrowband absorption" is one of its primary technologies. In the meantime, the inversion of XCH₄ is greatly influenced by the digital elevation model(DEM). Currently available methane inversion products mostly fit broadband structures with polynomials; however, the selection parameters for polynomial order are not well defined, and the broadband structure fitting is not precise enough. In local areas, the high-precision inversion requirements cannot be met by the accuracy of the employed DEM. This paper selects the Qinghai Tibet Plateau region, where the Waliguan Atmospheric Background Reference Observatory is located, as the research area. A higher precision digital elevation model(GLO-30)is used, and a fully connected neural network is used instead of a low order polynomial for “broadband structure” fitting. Furthermore, a “skip” structure is added to the traditional fully connected neural network, and a dropout strategy is used to optimize the network. By Comparing the experimental results with the XCH₄ inverted using the global multi resolution terrain elevation data 2010(GMTED2010)and low order polynomial fitting methods. The results show that the improved fully connected neural network can better fit the broadband spectral structure, and combining it with a higher precision DEM can improve the inversion accuracy of XCH₄, with the highest correlation coefficient increased to 0.92. The joint optimization method used can be used for remote sensing inversion of XCH₄ in oil and gas production areas, thereby better serving the investigation of methane abnormal emissions in oil and gas production areas.

Key words: WFM-DOAS, XCH₄, DEM, broad-band spectra fitting, artificial neural network, satellite remote sensing