Retrieving Atmospheric and Land Surface Parameters From At-Sensor Thermal Infrared Hyperspectral Data With Artificial Neural Network

Abstract

The radiances observed by satellites are influenced by both land surface and atmospheric parameters, and it is difficult to retrieve these parameters simultaneously from multispectral measurements with high accuracies. Even though several methods have been proposed, they focus on the retrieval of land surface or atmospheric parameters separately. Generally, these atmospheric parameters are atmospheric water vapor and temperature profiles. Thus, this study aims to establish a back propagation (BP) artificial neural network (ANN) to retrieve land surface emissivity (LSE), land surface temperature (LST), atmospheric transmittance, upward radiance, and downward radiance simultaneously from the hyperspectral thermal infrared (TIR) data, suitable for various air mass types and surface conditions. The principle component analysis technique is first used to compress and remove noise from the data. The evaluation of the ANN using the simulated data without noise indicated that the root mean square error (RMSE) of LST is approximately 0.643 K; the RMSEs of emissivity, transmittance, upward, and downward radiance are approximately 0.0046, 0.005, 0.72, and 2.95 K, respectively. When applied on the simulated data containing noise, the errors of LST, LSE, transmittance, upward, and downward radiance are 1.26, 0.01, 0.01, 1.54, and 4.57 K, respectively. When applied on the real atmospheric infrared sounder data, the retrieved accuracies become worse because of various unstudied reasons. However, the results show that the proposed ANN is promising in retrieving the land surface and atmospheric parameters simultaneously. Because of its simplicity, the proposed ANN can be used to produce preliminary results employed as the first estimates for physics-based retrieval methods.