Due to old or damaged components in temperature sensors and power outages at observatories, there are sometimes missing and abnormal temperature values from ground observatories. The incomplete temperatures have a significant impact on society and human life. Although we can obtain the near-surface air temperature by collecting other relevant meteorological information, it is challenging to invert the temperature precisely. In this paper, we make meteorological datasets with data from four backgrounds and propose a neural network, PC-Net, to invert the near-surface air temperature. We also design the parallel channel block, a parallel branch of convolutional layers, which improves the ability of the network to capture detailed characteristic information. PC-Net automatically extracts features from meteorological data and reconstructs these features for the near-surface air temperature. It successfully improves the accuracy of temperature inversion under complex weather conditions. PC-Net achieves a Pearson correlation coefficient of 0.993 and outperforms other structures, and it still reduces the RMSE of the best-performing U-Net by 3.87 \(\%\) (0.820 \(^\circ \) C vs. 0.853 \(^\circ \) C).