Because of the effects of radiation, existing air temperature instruments used in the meteorological detection field can produce radiation errors of approximately 1 °C. We developed an atmospheric temperature measuring instrument to reduce radiation error. We used the computational fluid dynamics (CFDs) approach to optimize the ability of the instrument to block radiation and guide airflow to the sensor. We mounted two aluminum plates and two airflow deflectors above and below the sensor. The outer surfaces of the plates were covered with high-reflectivity silver film, which could effectively block direct and reflected solar radiation. The deflectors had a streamlined shape, which could effectively guide airflow to the sensor. To further improve the accuracy of the instrument, we used the CFD approach to quantify its radiation errors under different meteorological conditions. Then, we utilized a neural network algorithm to develop a high-precision and universal radiation error correction algorithm. Subsequently, we conducted experiments to evaluate the accuracy of the new instrument. The experimental results indicated that the new instrument had a root means square error (RMSE), mean absolute error (MAE), and correlation coefficient of 0.0027 °C, 0.023 °C, and 0.99, respectively. The mean value, the upper and lower 95% confidence interval of the measured radiation errors of the new instrument were 0.088 °C, 0.096 °C, and 0.079 °C, respectively. These findings suggest that the new instrument has the potential to reduce the measurement error to less than 0.1 °C.