An original convolutional neural network, i.e. U-net approach, has been designed to retrieve simultaneously local soot temperature and volume fraction fields from line-of-sight measurements of soot radiation fields. A five-stage U-net architecture is established and detailed. Based on a set of N2 diluted ethylene non-premixed flames, the minimum batch size requirement for U-net model training is discussed and the U-net model prediction ability is validated for the first time by fields provided by the modulated absorption emission (MAE) technique documenting the N2 diluted flame. Additionally, the U-net model’s flexibility and robustness to noise are also quantitatively studied by introducing 5% & 10% Gaussian random noises into training together with the testing data. Eventually, the U-net predictive results are directly contrasted with those of Bayesian optimized back propagation neural network (BPNN) in terms of testing score, prediction absolute error (AE), soot parameter field smoothness, and time cost.