Research　on　lunar　oxides　abundance　has　been　spotlighted　for　its　great　significance　in　reconstructing　the　evolutionary　history　of　the　moon.　In　recent　years,　artificial　intelligence　technologies　have　been　introduced　to　map　oxides　abundance　on　the　lunar　surface　for　their　reliability　and　robustness.　However,　there　are　still　some　shortcomings　in　existing　studies.　First,　the　majority　of　these　studies　rely　on　spectral　data　and　used　in　situ　(drilled)　ground　truth　samples　collected　by　satellite　missions.　The　detection　depth　of　spectral　sensors　and　the　drilled　depths　of　the　returned　samples　are　not　consistent,　lowering　the　reliability　of　the　results.　Moreover,　existing　machine/deep　learning　models　may　not　be　suitable　for　processing　the　data　acquired　in　lunar　exploration.　In　this　article,　we　propose　a　novel　deep　learning　model　named　multifrequency　brightness　temperature　feature　fusion　network　(MFBTFF-Net)　for　processing　Chang＇e-2　lunar　microwave　sounder　(CELMS)　data　and　it　exploits　the　thermal　radiation　features　related　to　various　drilling　depths　to　acquire　the　global　lunar　oxide　abundance　maps.　The　experimental　results　demonstrated　that　the　proposed　MFBTFF-Net　model　can　significantly　improve　the　estimation　precision　of　most　lunar　oxides.　The　proposed　method　achieved　root-mean-square　error　indices　of　1.4449,　1.4826,　and　0.9824　(wt.%)　on　estimating　Al2O3,　FeO,　and　TiO2,　which　outperformed　the　state-of-the-art　models　by　at　least　0.0674,　0.6217,　and　0.0578,　respectively.　Furthermore,　based　on　the　proposed　model,　we　generated　a　new　set　of　lunar　oxide　abundance　maps.　Compared　with　the　abundance　maps　derived　from　spectral　data,　some　discoveries　can　be　obtained　due　to　the　unique　penetration　depth-related　information　provided　by　Chang＇e-2　CELMS　data.　This　study　demonstrates　the　large　potential　of　Chang＇e-2　CELMS　as　a　powerful　new　tool　to　understand　the　vertical　structures　of　the　moon　under　the　regolith.