Accurately estimating land surface temperature (LST) is a critical concern in thermal infrared (TIR) remote sensing. According to the thermal radiance transfer equation, the observed data in each channel are coupled with both emissivity and atmospheric parameters in addition to the LST. To solve this ill-posed problem, classical algorithms often require the input of external parameters such as land surface emissivity and atmospheric profiles, which are often difficult to obtain accurately and timely, and this may introduce additional errors and limit the applicability of the LST retrieval algorithms. To reduce the dependence on external parameters, this paper proposes a new algorithm to directly estimate the LST from the top-of-atmosphere brightness temperature in Landsat-9 two-channel TIR data (channels 10 and 11) without external parameters. The proposed algorithm takes full advantage of the adeptness of the ensemble learning method to solve nonlinear problems. It considers the physical radiance transfer process and adds the leaving-ground bright temperature and atmospheric water vapor index to the input feature set. The experimental results show that the new algorithm achieves accurate LST estimation results compared with the ground-measured LST and is consistent with the Landsat-9 LST product. In subsequent work, further studies will be undertaken on developing end-to-end deep learning models, mining more in-depth features between TIR channels, and reducing the effect of spatial heterogeneity on accuracy validation.

Land surface temperature (LST) is a critical physical parameter affecting energy and water exchange that has attracted much attention in various fields, such as environmental protection, agriculture, and climate change. Studies on spatially continuous and high-resolution LST retrieval methods, which can be efficiently acquired using thermal infrared (TIR) remote sensing technology, have been developed for many years, resulting in various LST remote sensing products. The typical mechanism thermal radiative transfer model is based on the assumption that the land surface is flat, with the TIR remote sensing image of the spatial resolution of the enhancement of the ability to observe the land surface of the three-dimensional geometric structure of the fine observation, due to the terrain caused by the topographic effect caused by the topography of the undulation becomes non-negligible, the assumption of flat surface may cause apparent errors. Some LST retrieval algorithms considering topographic effects have also been proposed recently. However, they are still inaccessible due to dependence on emissivity or atmospheric parameters, which limits the accuracy and timeliness of the retrieval algorithms. In addition, various machine learning algorithms for end-to-end LST retrieval have been proposed, which utilize their ability to handle complex nonlinear relationships to retrieve LST without external parameters. However such models currently do not fully consider the topographic effect due to a lack of account of the radiative transfer process in undulating terrain conditions. In this study, utilizing the ability of convolutional neural networks to extract spatial features from adjacent pixels, a radiative transfer model-driven convolutional neural network (CNN) model is proposed to realize the end-to-end retrieval of LST, considering the topographic effect. During training, a computational method based on ambient radiance scattered from the surrounding adjacent pixels in the improved radiative transfer model is used to obtain a local-scale simulation dataset covering different LSTs, emissivity, terrain undulations, and atmospheric conditions. The proposed CNN model is trained on this basis, and the theoretical accuracy is evaluated using the simulation dataset. The model has been applied to long-time-series Landsat-9 TIR remote sensing images. The accuracy is verified using terrain-corrected (TC) LST products. The results show that the new method proposed in this paper can effectively eliminate the topographic effect in TIR remote sensing observations and obtain accurate LST retrieval results, requiring only brightness temperature and digital surface model data.

Land surface emissivity is a key parameter that affects energy exchange and represents the spectral characteristics of land cover. Large-scale mid-infrared (MIR) emissivity can be efficiently obtained using remote sensing technology, but current methods mainly rely on prior knowledge and multi-temporal or multi-angle remote sensing images, and additional errors may be introduced due to the uncertainty of external data such as atmospheric profiles and the inconsistency of multiple source data in spatial resolution, observation time, and other information. In this paper, a new practical method was proposed which can retrieve MIR emissivity with only a single image input by combining the radiance properties of TIR and MIR channels and the spatial information of remote sensing images based on the Sentinel-3 Sea and land surface temperature radiometer (SLSTR) data. Two split-window (SW) algorithms that use TIR channels only and MIR and TIR channels to retrieve land surface temperature (LST) were developed separately, and the initial values of MIR emissivity were obtained from the known LST and TIR emissivity. Under the assumption that the atmospheric conditions in the local area are constant, the radiance transfer equations for adjacent pixels are iterated to optimize the initial values to obtain stable estimation results. The experimental results based on the simulation dataset and real SLSTR images showed that the proposed method can achieve accurate MIR emissivity results. In future work, factors such as angular effects, solar radiance, and the influence of atmospheric water vapor will be further considered to improve performance.

The thermal infrared (TIR) band of the Landsat-9 satellite has a spatial resolution of 100 meters, which is coarser than the optical multispectral bands of the 30-meter resolution and the panchromatic band of the 15-meter resolution. Existing image fusion algorithm studies primarily focuses on multispectral and panchromatic bands. For the TIR band with the lowest spatial resolution, although Landsat-9 land surface temperature (LST) products are resampled to 30m using the Cubic algorithm to match the multispectral bands, the physical mechanism of sharpening methods remains unclear, and they fail to fully leverage the higher spatial resolution advantage provided by the panchromatic band. This study proposes a practical LST sharpening method (TOA-MsPS) by integrating the correlation between top-of-atmosphere TIR brightness temperature (BT) and the fused reflectance of Multispectral and Panchromatic bands. The TOA-MsPS method comprises three steps: panchromatic and multispectral band fusion, correlation modeling of reflectance and BT, and LST end-to-end retrieval. Compared to existing methods, the TOA-MsPS method sharpens the spatial resolution of both TIR band BT images to 15 m without relying on external parameters, simultaneously deriving the LST data. All input data for the method consists of remote sensing observations at the TOA, reducing external parameter uncertainties and error propagation inherent in the preprocessing steps of current LST retrieval algorithms, such as atmospheric correction, resampling, and emissivity estimation. Qualitative visual interpretation based on visual inspection indicate that TOA-MsPS derived LST images exhibit significantly enhanced detail and reasonable local spatial distribution. Quantitative validations using ground site measurements demonstrate that the sharpened LST images achieve comparable accuracy to Landsat-9 LST products while substantially improving spatial resolution, with root mean square errors of 2.45K (LST product) and 2.35K (LST sharpen), respectively. Furthermore, the TOA-MsPS method can be directly applied to remote sensing data from multiple other remote sensing data sources, further enhancing the precision of long-term land surface thermal radiance monitoring.