Retrieval of atmospheric and land surface parameters from satellite-based thermal infrared hyperspectral data using a neural network technique

Land surface temperature (LST), land surface emissivity (LSE), and atmospheric profiles are of great importance in many applications. Radiances observed by satellites depend not only on land surface parameters (LST and LSE) but also on atmospheric conditions, and it is difficult to retrieve these parameters simultaneously from multispectral measurements with high accuracies. This work aims to establish a neural network (NN) to retrieve atmospheric profiles, LST, and LSE simultaneously from hyperspectral thermal infrared data suitable for various air mass types and surface conditions. The distributions of surface materials, LST, and atmospheric profiles were elaborated carefully to generate the simulated data. The simulated at-sensor radiances were divided into two sub-ranges in the spectral domain: one in the atmospheric window and the other in the water absorption band. Subsequently, the radiances were transformed in the eigen-domain in each sub-range, and then the transformed coefficients were used as the inputs for the network. Similarly, the atmospheric profiles, LST, and LSE were used as outputs after the eigen-domain transformation. The validation of the NN using the simulated data indicated that the root mean square error (RMSE) of LST is approximately 1.6 K, and the RMSE of the temperature profiles is approximately 2 K in the troposphere. Meanwhile, the RMSE of total water content is approximately 0.3 g cm⁻², and that of LSE is less than 0.01 in the spectral interval where the wave number is less than 1000 cm⁻¹. Two experiments using actual thermal hyperspectral satellite data were carried out to further validate the proposed NN. All of these studies showed that the proposed NN is capable of retrieving atmospheric and land surface parameters with compromised accuracies. Because of its simplicity, the proposed NN can be used to yield preliminary results employed as first estimates for physics-based retrieval models.