The radiative transfer model is considered a promising approach for atmospheric correction (AC). This approach requires inferencing a set of parameters using complicated models and tables, leading to uncertainty in the removal of atmospheric effects and sometimes produces negative remote sensing reflectance, \({R}_{rs}\left(\lambda \right)\) . In this study, a learning-based AC model named AC-Net, based on convolutional and fully-connected neural networks, is proposed to retrieve \({R}_{rs}\left(\lambda \right)\) for Landsat-8 imagery over inland waters in tropical regions. In AC-Net, the convolutional subnetwork extracts spectral features of the top-of-atmosphere reflectance while the fully-connected subnetwork integrates these spectral features with sun-sensor geometric angles and aerosol optical thickness to derive \({R}_{rs}\left(\lambda \right)\) . To overcome model overfitting and geographical sensitivity problems caused by an insufficient quantity of in-situ training samples, a large set of satellite-derived \({R}_{rs}\left(\lambda \right)\) in various trophic states is generated using an existing AC model. The satellite-derived \({R}_{rs}\left(\lambda \right)\) , along with a small set of in-situ \({R}_{rs}\left(\lambda \right)\) , are used to optimize thousands of unknown parameters in AC-Net. In addition, the sigmoid function is selected as the activation function in the output layer, which prevents the output of negative reflectance values. In experiments, AC-Net was compared with related AC models, including QUAC, ACOLITE, FLAASH, LaSRC, iCOR, and C2X. The experimental results demonstrated that AC-Net has better performance than the compared models, with the results of root mean squared error (RMSE) = 0.0039 \({\text{sr}}^{-1}\) , mean absolute percent error (MAPE) = 4.19%, and spectral angle (SA) = \({19.5}^{0}\) . The testing results showed that AC-Net can avoid the output of negative \({R}_{rs}\left(\lambda \right)\) reflectance and alleviate the geographical sensitivity problem.
