利用随机森林和地理加权回归的黄河源GIMMS_(3g)NDVI降尺度方法

GIMMS_(3g) NDVI Downscaling Using Random Forest and Geographically Weighted Regression in Yellow River Source

针对现有AVHRR、SPOTVGT、MODIS产品难以构建长时序、高时空分辨率NDVI数据集的问题,提出了利用随机森林和地理加权回归模型对GIMMS_(3g) NDVI进行降尺度的方法。基于“关系尺度不变”假设,从不同空间分辨率和数据源角度将其空间分辨率从8 km提高至250 m,并利用MODIS数据进行精度评价。结果表明:降尺度数据的空间分辨率有较大提升,能真实反映源区内NDVI空间分布特征;降尺度数据与MODIS数据具有较好的一致性,除500 m分辨率下的RF降尺度外,其他降尺度结果的绝对误差≤0.1的比例达到70%;东南部高植被覆盖区的降尺度效果要优于西北部;RF模型在体现局部细节纹理特征方面更具优势;解释变量的不同组合会影响降尺度的精度;降尺度模型尺度的改变对RF模型降尺度结果影响较大。

The method of downscaling is presented for the problems that difficulty in constructing NDVI datasets with long time series and high spatio-temporal resolution by AVHRR,SPOTVGT and MODIS products,the original GIMMS_(3g) NDVI data is downscaled by using random forest and geographically weighted regression model.Based on the assumption of“relationship scale is invariant”,its spatial resolution has been increased from 8 km to 250 m by different spatial resolutions and data sources,and its accuracy is evaluated by MODIS data.The results show that:the spatial resolution of downscaled data has been greatly improved,which can truly reflect the distribution characteristics of NDVI in the area;the downscaled data and MODIS data show good consistency in space and time,except for RF downscaling at 500 m resolution,the proportion of absolute error≤0.1 reaches about 70% in other downscaling results;the downscaling effect is better in the southeast area with higher vegetation coverage than that in the northwest;downscaled data of the random forest method have more advantages in displaying local detail texture features;the combination of different explanatory variables effects the accuracy of downscaling;changing of downscaling model scale has great influence on RF downscaling results.