地表温度(Land surface temperature, LST)是研究地表能量平衡中必不可少的参数，在分析地球资源和环境动态变化过程中起着关键的作用。遥感卫星可以获取大范围长时序的观测数据，但目前受到传感器性能指标的限制，使得获取的数据存在时空分辨率的矛盾，限制了数据更广泛的应用。通过LST降尺度算法可以提高热红外数据的空间信息，其中基于统计回归模型的算法凭借其操作简单且降尺度精度较高等优势得到了广泛的研究与应用。近年来，基于统计回归模型的LST降尺度算法经过不断优化取得了更高的精度，但是目前此类算法依然存在一些问题。首先，基于统计回归模型的LST降尺度算法在异质性较强的区域考虑了变量的空间非平稳性，但是对LST与辅助参数之间非线性关系研究较少；其次，目前统计回归模型降尺度算法建模过程中，通常是基于尺度关系不变性的假设，忽略了不同分辨率数据之间转换的尺度效应；最后，目前的研究区域多集中于平原和城市区域，对山地等复杂地形区域的研究较少且降尺度结果的精度较低。

针对上述提到的LST降尺度算法中存在的问题，本研究提出相应的解决方案对算法进行优化，主要的内容与结论如下：

(1) 提出基于非线性地理加权回归模型的LST降尺度算法。

本研究针对目前LST降尺度局部模型中对变量间非线性关系研究较少的问题，提出非线性地理加权回归模型进行LST降尺度研究。此模型对变量间不同的非线性关系进行探究，选择最优的辅助参数与LST建立非线性关系，进行LST降尺度的研究。本研究选取两个不同气候的城市区域各三景数据对提出的算法进行验证分析，结果表明：将辅助参数的二次项与LST建模可以得到更好地拟合结果；基于非线性地理加权回归模型取得了较好的降尺度结果，与参考的验证数据相比，降尺度LST的均方根误差(Root Mean Square Error, RMSE)和平均绝对误差(Mean Absolute Error, MAE)的统计值均在2℃以下且决定系数(Coefficient of determination, R²)均在0.9以上。本研究提出的基于非线性地理加权回归模型的LST降尺度算法在城市区域取得了较好的降尺度结果。

(2) 提出泰勒展开模型用于修正LST降尺度过程中尺度效应。

本研究针对目前LST降尺度过程中尺度不变性假设的局限性，对LST降尺度算法中的尺度转换问题进行研究，提出了基于泰勒展开模型的LST降尺度算法。泰勒展开的二次项纠正了模型中非线性带来的尺度效应，对非线性模型进行优化。本研究选择三个不同研究区域十景数据进行算法的适用性研究，结果表明：基于泰勒展开模型的LST降尺度算法取得了较高的精度，与参考的LST相比，十景数据降尺度LST的RMSE和MAE均小于1.5℃，且R²大于0.90。另外，本研究针对算法中引入的经验因子S与MODIS LST获取时间不一致的问题进行了深入的探索，当MODIS LST获取日期无法获取经验因子S的数据时，可以选择同一或者相似季节的经验因子S代替。本研究算法纠正了非线性关系中的尺度效应，并对算法中不同数据获取时间不匹配问题进行了考虑，为降尺度的研究提供了新的思路。

(3) 开展复杂地形区域LST降尺度研究。

针对目前山地等复杂地形区域LST降尺度研究较少且结果较差的问题，对此类区域LST降尺度算法进行研究。复杂地形区域LST的分布具有特殊性，LST不仅受到地物覆盖的影响，还会受到海拔、太阳辐射等的影响。因此在复杂地形区域进行LST降尺度建模时，需要对LST与地形、海拔以及太阳辐射的关系进行研究。本研究选取两个山地区域不同季节的四景数据进行LST降尺度研究，结果表明：将短波辐射加入LST降尺度建模过程中有利于降尺度精度的提高，RMSE小于1.5℃，且与参考数据拟合的R²均大于0.92。另外对选取的四景图像进行分析发现，加入短波辐射在秋冬季的结果提高相比于夏季更明显，改善了降尺度算法研究中冬季精度较低的弊端。

本研究对基于统计回归模型LST降尺度算法中存在的几个问题提出了解决方案，并将提出的算法用于不同研究区域、不同季节的数据中，证明了算法的普适性。通过对结果的验证分析可以发现本研究提出的算法具有较强的鲁棒性，在平原、城市以及山地等异质性较强的区域均可获得高时空分辨率的LST数据，可以为热红外的相关研究提供高时空分辨率的数据支持。

Land surface temperature (LST) is a very important parameter in geoscience and plays an important role in the process of surface-atmosphere energy exchange. Thermal infrared remote sensing satellite can obtain thermal infrared data with large range and long time series. At present, due to limitation of sensors, it makes the obtained thermal infrared data have the contradiction of spatio-temporal resolution, which limits the wider application of data. LST downscaling algorithm is an effective way to improve spatial resolution of thermal infrared data, and the algorithms based on statistical regression models have been widely studied and applied due to their advantages of simple operation and high downscaling accuracy. In recent years, LST downscaling algorithm based on statistical regression model has obtained higher precision through continuous optimization, but there are still some problems in these algorithms. Firstly, the LST downscaling algorithm based on statistical regression models considers the spatial non-stationarity of variables in heterogeneous regions, but there is little study on the nonlinear relationship between LST and auxiliary parameters. Secondly, most statistical regression models assume scale invariance, which makes the downscaled LST inaccurate. Finally, most of the current research focuses on the plain and urban areas, with few studies on complex terrain areas, and the accuracy of downscaling results is lower. In view of the above-mentioned problems in the LST downscaling algorithm, this study proposed a series of solutions to improve the algorithm. The main research content and conclusions are as follows:

An LST downscaling algorithm based on nonlinear geographically weighted regression (NL-GWR) model has been proposed.

In this study, aiming at the problem that there is less research on the nonlinear relationship between LST and auxiliary parameters, a nonlinear geographical weighted regression model is proposed for LST downscaling study. This model studies the nonlinear relationship between auxiliary parameters and LST and considers the spatial nonstationarity between variables. This study selected three scene data from two urban areas with different climates to validate and analyze the proposed algorithm. The results show that the NL-GWR model has achieved good downscaling results. Compared with the reference validation data, the root mean square error (RMSE) and mean absolute error (MAE) of downscaling LST are both below 2℃, and coefficient of determination (R²) is above 0.9 by fitting with reference data. The LST downscaling algorithm based on NL-GWR model proposed in this study has achieved good downscaling results in urban areas.

An LST downscaling algorithm based on Taylor expansion model has been proposed.

Aiming at the irrationality of scale invariance hypothesis in the LST downscaling process, we studied the scale conversion problem in the LST downscaling algorithm, and proposed an LST downscaling algorithm based on Taylor expansion model. The quadratic term based on Taylor expansion corrects the scale effect caused by model nonlinearity. In this study, ten scene data of three different research areas were selected for the applicability study of the algorithm, and an exploration was conducted to address the inconsistency between the empirical factor S introduced in the algorithm and the acquisition time of MODIS LST data. The results showed that when the empirical factor S was consistent with the acquisition time of MODIS LST data, the algorithm proposed in this study achieved higher accuracy. Compared with the reference LST, the RMSE and MAE of the ten scenes data downscaling LST were both less than 1.5 ℃, And R² is greater than 0.90. When the acquisition time of empirical factor S is inconsistent with that of MODIS LST, empirical factor S in the same or similar season can usually be cross applied. Therefore, when the data of empirical factor S cannot be obtained on the acquisition date of MODIS LST, empirical factor S in the same or similar season can be selected to replace it. The algorithm in this study not only considers the scale effect, but also explores the mismatch problem of different data acquisition times in the algorithm, and achieves good results.

Study on LST downscaling in complex terrain regions.

Aiming at the problem that there are few studies on LST downscaling in mountainous and other complex terrain regions and the results are poor, LST downscaling algorithms for such regions are studied in this study. The distribution of LST in complex terrain regions has its particularity. LST is not only affected by ground cover, but also affected by altitude, solar radiation, and so on. Therefore, when performing LST downscaling modeling in complex terrain areas, it is necessary to study the relationship between LST and terrain, altitude, and solar radiation. In this study, four scene data of two mountainous regions in different seasons were selected for LST downscaling research. The results showed that adding short-wave radiation to LST downscaling modeling process is conducive to improving the downscaling accuracy, with RMSE less than 1.5 ℃, and R² fitting with reference data greater than 0.92. In addition, after analyzing the selected four scene images, it was found that adding shortwave radiation improved the results more significantly in autumn and winter than in summer, which also improved the disadvantage of the previous downscaling algorithm that had a poor effect in winter.

This study proposes corresponding solutions to several problems in LST downscaling algorithms based on statistical regression models, and applies the proposed algorithm to data from different study areas and different seasons, demonstrating the universality of the algorithm. Through verification and analysis of the results, it can be found that the algorithm proposed in this study has strong robustness. High spatiotemporal resolution LST data can be obtained in areas with strong heterogeneity such as plains, cities, and mountains, which can provide high spatiotemporal resolution data support for LST related research.