Lunar surface chemical compositions are essential for understanding the geological evolution of the Moon. One of the mission objectives of Lunar Reconnaissance Orbiter (LRO) Diviner is to obtain the global Christiansen feature (CF) product and then employ it to estimate surface oxide abundances. However, the early CF products are mixed with the information on the viewing geometry, space weathering and compositions and cannot be directly used for surface composition inferences. Recently, the new Normalized to Equatorial Noon (NEN) and OMAT corrected NEN CF (OMATCF) images were calculated and provided for the correction of the viewing geometry and space weathering effects. Therefore in this paper, six Back Propagation Neural Network (BPNN) models are firstly established based on the relationships between the OMATCF values and ground truths of oxide abundances at 48 lunar sampling sites. Then, these BPNN training models are applied to the OMATCF map and six Diviner oxide products with the resolution of 32 pixels/degree (ppd) and the coverage of >99% (70°N/S) are calculated and presented. The comparisons with the previous four results indicate that the prediction accuracy of Diviner oxides products are the highest for SiO (1.79), AlO (2.04), MgO (0.88) and CaO (1.10) except for TiO (3.17, the second lowest) and FeO (1.93, the second highest). Meanwhile, a satisfactory consistency is observed between Diviner results and Clementine or Chang'E (CE)-1 results. Considering higher spatial resolution (128 ppd) CF product in the future, the Diviner oxides will be the better data sources for lunar geological applications.