基于地理加权回归模型的典型山地卫星反演降水产品降尺度研究

Spatial Downscaling of Remotely Sensed Precipitation Using Geographically Weighted Regression Algorithms in Typical Mountainous Areas, China

降水是陆地水循环的关键变量,高分辨率降水数据的获取是准确模拟陆地水循环过程的前提。虽然卫星反演降水产品具有较强的空间代表性和连续性,但其空间分辨率较低的问题限制了它的应用。以太行山、横断山和喀斯特山区为研究对象,基于降水与高程(DEM)、植被指数(NDVI)之间存在较好相关关系的假设,构建了GPM降水(Global Precipitation Measurement Mission)与高程、植被指数的地理加权回归模型,得到了2014—2016年研究区1km分辨率GPM降水数据。研究结果表明:地理加权回归模型能有效地提高GPM数据的空间分辨率。降尺度后,GPM数据精度在太行山和横断山区略有提高。年尺度上,相比于原始GPM数据,太行山和横断山区降尺度数据站点实测数据的确定系数分别提高了0.06和0.08,RMSE分别降低了0.45%和3.89%,MAE分别降低了0.16%和1.70%;月尺度上,太行山区67%的月份,横断山区83%的月份GPM产品降尺度后更加接近于站点实测数据。喀斯特地区GPM数据降尺度后精度略有下降,降尺度后,年尺度的降雨数据与实测数据的RMSE和MAE分别增加了10.00%和8.00%,R^2降低了0.06,月尺度上仅8月和9月降尺度后的精度更高。降雨与地形和NDVI的关系较弱是喀斯特地区降尺度效果较差的主要原因。

Precipitation is a key factor in terrestrial water cycle. The acquisition of high-resolution precipitation data is a prerequisite for simulating terrestrial water cycle with high precision. Although satellite-based precipitation has high spatial representativeness and continuousness, the relatively low spatial resolution in the product limits its applications in terrestrial hydrological simulations. Based on the assumption that there exists a strong correlation between precipitation, altitude and vegetation index, a Geographically Weighted Regression (GWR) model for the precipitation, elevation and vegetation index was developed, and the monthly and annual Global Precipitation Measurement Mission (GPM) data with 1-km resolution in three typical mountainous areas (i.e. Taihang mountainous area, Hengduan mountainous area and Kasite mountainous area) from 2014 to 2016 were obtained. The results showed that the GWR model could effectively enhanced the spatial resolution of the GPM data. The resolution of the GPM data slightly increased in Taihang mountainous area and Hengduan mountainous area after downscaling. At annual scale, after downscaling, the coefficient of determination (R^2) between observed data and GPM increased by 0.06 and 0.08, the root-mean-square error (RMSE) decreased by 0.45% and 3.89%, and the mean absolute error (MAE) decreased by 0.16% and 1.70% in Taihang mountainous area and Hengduan mountainous area, respectively. At monthly scale, the downscaled precipitation was closer to the observed precipitation in more than 67% of the months in Taihang mountainous area and 83% of the months in Hengduan mountainous area. However, the resolution of the GPM data slightly degraded in Kasite mountainous area: at annual scale, the R^2 decreased by 0.06, and the RMSE/MAE increased by 10.00%/8.00% after downscaling;at monthly scale, the downscaled precipitation showed higher precision than original GPM data only for August and September. The poor performance of the downscaling algorithm in Kasite mountainous area was mainly due to a weak correlation between precipitation, vegetation index and altitude.