Inverse modeling with neural networks for the retrieval of cloud parameters

This paper presents the solution of an inverse model problem for the retrieval of cloud parameters by means of neural networks. Information about clouds is extracted from measurements in and around the oxygen A-band at 760 nm. The average transmittance through this band defines a relationship between cloud top height, cloud fraction and cloud optical thickness. The reflectance spectrum contains 79 single spectral points, but the mapping in remote sensing inverse problems is usually better specified in a lower-dimensional space. Dimensionality reduction or feature extraction is performed using non-linear principal component analysis. A neural network is used for this task, the dimensionality of the data is reduced to 4 principal components. Radiative transfer model simulations are used to compute oxygen A-Band reflectance for several viewing geometry and geophysical scenarios. A second neural network is trained to solved the inverse problem based on the model simulations. The new Inversion approach coupling two neural networks is extremely fast and robust and can be used in near-real-time applications