Tunable diode laser absorption spectroscopy tomography (TDLAT) is one of the combustion diagnosis methods. Aiming at the problem that the existing TDLAT reconstruction algorithm is difficult to reconstruct gas parameters rapidly and accurately in limited-data absorption spectroscopy, the existing TDLAT reconstruction algorithm based on convolutional neural network (CNN) is improved by combining with the deep learning theory, which improve the accuracy of reconstructed flame temperature distribution to a large extent, and is suitable for different flame characteristics and a variety of beam arrangements. The CNN training and structure optimization methods for TDLAT reconstruction are studied. The necessity of preprocessing the integrated absorbance data and the temperature data is discussed. A hierarchical learning model is proposed, which effectively utilizes the prior smoothness information of gas parameters. The verification results on gaussian phantoms show that the average reconstruction error of this algorithm is only 0.24% when there is no noise. The algorithm is further validated on turbulent methane plume. Finally, a flat flame burner temperature measurement system is built. The experimental results show that this algorithm can rapidly reconstruct the two-dimensional temperature distribution at the z=1.5 cm cross section of the flame under different combustion states.