The radiative transfer coupling between the Earth's surface and the atmosphere is highly complex, resulting in limitations in the accuracy and efficiency of traditional atmospheric correction techniques. Empirical linear methods offer relatively high computational efficiency but suffer from low precision, while radiative transfer model-based approaches achieve higher accuracy at the cost of substantial computation, restricting the quantitative application of multi-spectral satellite remote sensing. With the development of neural networks, integrating physical models with data-driven approaches has provided new vision for atmospheric correction. This study proposes a deep neural network framework trained on datasets simulated by radiative transfer models, which not only ensures correction accuracy but also significantly improves processing efficiency. First, radiative transfer model is used to simulate satellite top-of-atmosphere radiance under different observation geometries and atmospheric–surface conditions, thereby constructing a forward-simulated multi-spectral remote sensing dataset for data-driven atmospheric correction. The neural network then learns the geographic, geometric, and spectral characteristics embedded in the dataset to achieve pixel-level atmospheric correction. The study demonstrate that the proposed method achieves an average absolute error of 0.028 and a relative error of 0.82 compared to pixel-wise correction using radiative transfer models, while improving processing speed by approximately six orders of magnitude. Compared with mainstream physical-model-based methods in terms of image quality, quantitative metrics, and consistency with in-situ measured vegetation spectra, the proposed method performs favorably. This research provides an efficient atmospheric correction pathway for the quantitative application of multi-spectral remote sensing imagery.