Simultaneous Estimation of Land Surface and Atmospheric Parameters From Thermal Hyperspectral Data Using a LSTM–CNN Combined Deep Neural Network

Thermal infrared (TIR) remote sensing observation signal is influenced by both atmospheric and land surface conditions that are difficult to separate with conventional multichannel TIR data. Because of the advantage of channel wealth, hyperspectral TIR data can simultaneously estimate the land surface and atmospheric parameters using neural network models or integrating them with physical models. However, the commonly used neural network models do not fully explore the correlation between different channels by treating the input data as discrete features. Thus, this study aims to develop a new deep neural network (DNN) by combining the long short-term memory (LSTM) network and convolutional neural network (CNN) for estimating land surface temperature (LST), emissivity, atmospheric transmittance, upward radiance, and downward radiance more accurately. By applying on the thermal airborne hyperspectral imager (TASI) simulation dataset covering global atmospheric conditions with 32 channels in $8.0- 11.5\,\,\mu \text{m}$ , the proposed model achieved results with the LST error of 0.95 K, the emissivity error of less than 0.012 for each channel, and the accuracy of three atmospheric parameters has also been improved compared with the current neural network models. Our model has been applied to a real TASI image, and its validity was further proved by the ground measurement validation data. Therefore, it can provide more reliable initial values for physical optimization models.