In order to achieve the direct estimation of the vertical distribution of subsurface chlorophyll concentration from the physical quantities observed by remote sensing in the ocean surface layer, this paper applies a Convolutional Neural Network (CNN) to the inversion study of subsurface chlorophyll maxima (SCMs) for the first time. Our CNN model consists of three modules which are input module, feature extraction module and chlorophyll profile estimation module. Multi-source chlorophyll data at depths 0-200 m in the South China Sea, including OCCCI remote sensing data, BGC-Argo observation data and CMEMS ocean model data were used for training the CNN model. The model takes the local surface chlorophyll concentration, monthly features and SST as input predictors. We first pre-trained the CNN model with CMEMS model data, and then fine-tuned the model parameters with remotely sensed surface data and BGC-Argo data by transfer learning. Our CNN model have an average MSE of 0.01 mg/m³ and a correlation coefficient of 0.84 in the test set. The modelling results show that the CNN can effectively reproduce the SCM intensity and its depth in the basin of South China Sea.