Rare earth resource is a national strategic resource, which plays an essential role in the field of high technology research and development. In this paper, we aim to use remote sensing quantitative inversion prospecting technology, use surface-to-surface mode, and model inversion and evaluation through convolutional neural network model to achieve a new research method for large-scale, low-cost, rapid and efficient exploration of ion-adsorbed rare earth ore. The results show that the RE2O3 content of samples has significant negative correlation with the second, third and fourth band of GF-2 image, but has no significant correlation with the first band of GF-2 image; the convolution neural network model can be used to reconstruct the RE2O3 content. The content distribution map of RE2O3 obtained by inversion is similar to that of geochemical map, which indicates that the convolution neural network model can be used to invert the RE2O3 content in the sampling area. The quantitative inversion results show that the content distribution characteristics of ion adsorption rare earth ore in the study area are basically consistent with the actual situation; there are two main high anomaly areas in the study area. The high anomaly area I is a known mining area, and the high anomaly area II can be a prospective area of ion adsorption type rare earth deposit. It shows that the remote sensing quantitative inversion prospecting method of ion adsorption type rare earth deposit based on Convolutional Neural Networks (CNN) model is feasible.