The benefits of using a machine learning approach in a forward operator for visible satellite images are explored. In the conventional version of the forward operator, cloud-affected reflectances are determined by linear interpolation in a compressed, seven-dimensional look-up table (LUT) computed with standard radiative transfer (RT) methods. It is demonstrated that replacing the LUT by a feed-forward neural network can reduce the computational effort by an order of magnitude without detrimental impact on the accuracy of the method. The sensitivity of the mean reflectance error to parameters controlling the network structure and the training process is investigated. Best results are obtained for networks with between four and eight hidden layers. Moreover, for the training of the network only $1/1000$ of the data that has to be computed for the LUT using slow standard RT methods is required. The amount of memory required while generating synthetic images is reduced by a similar factor, compared to the LUT-based approach. The reduced requirements and increased speed strongly enhance the extensibility of the method. Adding more input parameters to account e.g. for traces gases, aerosols or more details in the cloud structure would be problematic for the conventional approach due to strongly increasing LUT sizes, but should be feasible in the neural network based version. A neural network inference code including tangent linear and adjoint versions was implemented to demonstrate further advantages of the new approach. In contrast to the LUT-based approach the derivatives computed with the adjoint of the neural network are continuous. Moreover, the adjoint code will not have to be changed when the network is trained with improved RT methods. The effort to keep the adjoint code in sync with the nonlinear code can thus be avoided.

探讨了在可见卫星图像的前向算子中使用机器学习方法的好处。在前向算子的传统版本中，受云影响的反射率是通过使用标准辐射传输 (RT) 方法计算的压缩七维查找表 (LUT) 中的线性插值确定的。结果表明，用前馈神经网络替换 LUT 可以将计算工作量减少一个数量级，而不会对方法的准确性产生不利影响。研究了平均反射率误差对控制网络结构和训练过程的参数的敏感性。对于具有 4 到 8 个隐藏层的网络，可以获得最佳结果。此外，仅用于网络的训练$1/1000$需要使用慢速标准 RT 方法为 LUT 计算的数据。与基于 LUT 的方法相比，生成合成图像所需的内存量减少了类似的因素。降低的要求和提高的速度极大地增强了该方法的可扩展性。添加更多输入参数以考虑例如微量气体、气溶胶或云结构中的更多细节，对于传统方法来说将是有问题的，因为 LUT 大小大幅增加，但在基于神经网络的版本中应该是可行的。实施了包括切线线性和伴随版本的神经网络推理代码，以展示新方法的进一步优势。与基于 LUT 的方法相比，使用神经网络的伴随计算的导数是连续的。而且，当使用改进的 RT 方法训练网络时，不必更改伴随代码。因此可以避免保持伴随码与非线性码同步的努力。