喀斯特地区地表温度空间降尺度方法初探

In karst areas, the terrain fluctuates greatly, and conventional downscaling methods and selected factors are not applicable to it. According to the characteristics of karst areas, this paper selects reflectivity, remote sensing indices and elevation factors as scale factors, and establishes a nonlinear relationship between MODIS band 31 and 32 radiance data and scale factors through a random forest model to construct a Karst Random Forest (KRF) model suitable for karst areas. The thermal infrared radiance with a spatial resolution of 1 km is successfully downscaled to 100 m, and finally the land surface temperature with a spatial resolution of 100 m is retrieved using the split-window algorithm. Comparing the KRF method with the multi-factor random forest regression model (MTVRF) considering only conventional factors and the thermal sharpening algorithm (TsHARP), the results show that: 1) In karst areas with different elevation differences, the KRF method can significantly improve the downscaling accuracy of land surface temperature. The root mean square error (RMSE) is 2.46 K and 1.45 K in the northwest of Zunyi City and the area south of Guiyang City respectively, which is 0.1419 K and 0.2928 K lower than that of the MTVRF model respectively, and 0.6204 K and 0.6953 K lower than that of the TsHARP algorithm respectively, and performs better in karst mountainous cities with lower terrain undulation (south of Guiyang City); 2) The KRF method also has a good effect on different land types in karst areas, with the best performance in the vegetation area, where the RMSE is 1.41 K, and the RMSE in the fragmented bare soil area is 1.84 K. The research shows that the KRF method considering special scale factors can improve the downscaling accuracy of land surface temperature in karst areas and provide a more refined and reliable land surface temperature product for research based on land surface temperature in this area.

喀斯特地区地形起伏大,常规的降尺度方法及所选择的因子对其不适用。该文根据喀斯特地区的特点,选取反射率、遥感指数及高程因子为尺度因子,通过随机森林模型建立MODIS第31、32波段辐射亮度数据和尺度因子之间的非线性关系,构建适合喀斯特地区的随机森林(Karst Random Forest,KRF)模型,成功将空间分辨率为1km的热红外辐射亮度降至100m,最后利用劈窗算法反演得到空间分辨率为100m的地表温度。将KRF方法与仅考虑常规因子的多因子随机森林回归模型(MTVRF)和热锐化算法(TsHARP)对比,结果表明:1)在不同高差的喀斯特地区,KRF方法可较大程度提高地表温度降尺度精度,均方根误差(RMSE)在遵义市西北部和贵阳市以南地区分别为2.46K和1.45K,较MTVRF模型分别降低了0.1419K和0.2928K,较TsHARP算法分别降低了0.6204K和0.6953K,且在地形起伏度较低的喀斯特山区城市(贵阳市以南)表现更好;2)在喀斯特地区不同地类上,KRF方法效果也较好,其中植被区域最优,RMSE为1.41K,破碎的裸土区域RMSE为1.84K。研究显示,考虑特殊尺度因子的KRF方法可提高喀斯特地区地表温度的降尺度精度,为该地区以地表温度为基础的研究提供更精细可靠的地表温度产品。