DisTrad(Disaggregation procedure for radiometric surface temperature)模型是常用于遥感地表温度空间分辨率提升的主要模型之一。DisTrad模型常面向空间范围有限、地形相对平坦的研究区域,且常选用植被参数(如植被指数或植被覆盖度等)作为关键参数。然而在空间范围较大、地形起伏地区,地表温度的空间变异可能无法完全通过植被参数解释。本研究选取四川盆地及毗邻地区为研究区,通过模拟数据研究DisTrad模型在地形起伏区地表温度空间分辨率提升中的适用性。数字高程模型(Digital Elevation Model,DEM)、归一化差值植被指数(Normal-ized Difference Vegetation Index, NDVI)等参数,采用滑动窗口逐步回归,将空间分辨率为6km的地表温度提升至空间分辨率为1km。研究结果表明,改进的模型在平原及海拔较低的高原地区提升获得的地表温度空间分辨率具有较高精度,均方根误差(Root Mean Square Error, RMSE)为0.5K左右;在地形起伏较大的地区,RMSE为4K,验证了改进的模型提升的可行性。
英文摘要:
Accurate temporal and spatial estimation of land surface temperature (LST) is important for evaluat-ing climate change, global hydrological cycle and monitoring urban heat islands (UHI). LSTs with high quality can be routine by using satellite remote sensing. However, characters of both high spatial and temporal resolu-tions have been difficult. Cloud cover further reduces the useable observations of surface conditions. Monthly LST product (MOD11C3) composited and averaged temperature values at 0.05 degree latitude/longitude grids (CMG) have coarse spatial resolution (~5.5 km). An alternative to the lack of high-resolution observations is to disaggregate LST data using other products of MODIS of 1 km observations. Historically, disaggregation of LST at high resolutions (1 km) has relied on vegetation index, e.g. NDVI (Normalized Difference Vegetation Index). However, this downscaling method is not adequate for areas encompass basin and upland. We applied Digital El-evation Model (DEM), NDVI, Enhanced Vegetation Index (EVI), Albedo, and slope to resolve this drawback by utilizing stepwise regression method with a moving window. The following is the algorithm. Land surface param-eter (LSP) data are sampled to the coarser thermal resolution. A stepwise regression is performed between the monthly temperature product and sampled land surface parameters, then a function f (LSP) framed. The parame-ters of the regression function are applied to LSP data at high, target resolution. Coarse-scale residual field repre-sent variability in temperature driven by other factors other than vegetation and DEM is added back into the high-resolution base map. So, we utilize LSP to sharpen original images. A reasonable rectangle box that making certain pixel be center is outlined for stepwise regression. Function is obtained by stepwise between LST and LSP. Loop and downscale the other pixels until image processed. Coefficients and intercept are saved as images. The disaggregation LST is achieved by substituting images at target resolution to function. The size of the box flowed over the image in this paper is 19 by 19. Stepwise disaggregation algorithm is applied to the resample MOD13A1 and DEM data. The fitting parameters vary with different window scenes. In contrast, the number of DEM entered function is much larger than NDVI. That indicated DEM is more significant than NDVI, EVI, albe-do and slope in most fields of the study area, especially in mountain area. The RMSE of downscaling LST is 4.93K. Image sharpening is therefore not a replacement for high-resolution thermal imaging sensors. Neverthe-less, in the absence of thermal imagery because of cloud in Sichuan, DisTrad seems to be able to enhance the res-olution of MOD11C3 product.