The influence of non-uniform temperature on the back-up structure (BUS) is one of the important factors that causes the accuracy of the main reflector of the radio telescope antenna to decrease. Due to the complex topological structure of the BUS, there are shielding, heat conduction and heat radiation among the rods, which makes the temperature field of the BUS difficult to be accurately obtained and predicted by thermodynamic simulation. In this study, temperature sensors were installed on the BUS of Nanshan 26-meter radio telescope (NSRT) to obtain the BUS temperature data set. Three different unsupervised feature selection (UFS) methods were used to select 16 temperature sensitive points from 66 temperature measurement points, and then these three different sets of temperature sensitive points were used as inputs. MIMO-BP (Multiple Input and Multiple Output - Back Propagation) neural network model is used to train the predicted temperature values of 66 points corresponding to the output, and then the temperature prediction of global continuous points on the BUS is realized by interpolation algorithm. Through calculation and comparative analysis, it is concluded that the unsupervised feature selection method based on eigenvalue sensitive criterion (EVSC) has the best effect on selecting temperature sensitive points. Combined with the BP neural network and Barnes interpolation algorithm, only 16 measured temperature points are used to predict the temperature field distribution of global continuous points on the NSRT's BUS. The root-mean-square error (RMSE) of the predicted value is about 0.707 ^\circC. The research result provides an alternative method for the arrangement of temperature collection points and the acquisition and prediction of temperature field in the BUS of large aperture radio telescope.