基于改进S-G滤波和非监督分类局部核回归的中国LAI时序数据融合研究

LAI Time Series Data Fusion in China Based on Improved S-G Filtering and Unsupervised Classification Local Kernel Regression

叶面积指数Leaf Area Index(LAI)是表征植被冠层结构的重要参数,其遥感产品常常因云、气溶胶、积雪、算法和仪器问题等因素影响,导致数据质量差或缺失,严重影响LAI数据集的应用。本文提出了一种基于改进S-G滤波和非监督分类局部核回归的LAI时序数据融合方法,并利用2014—2020年MODIS LAI、PROBA-V LAI、VIIRS LAI产品数据,开展归一化融合研究,以提高产品的一致性、连续性和精确性。结果表明,融合LAI与源产品及其他LAI产品(MCD15A2H、MOD15A2H、VNP15A2H、PROBA-V)的LAI值频率分布、时序变化表现出良好的一致性,并有较好的相关性,R2分别为0.85、0.77、0.84和0.89,与这4个产品相比,数据缺失频率总体下降,时间连续性有所提高,相较于MCD15A2H LAI(19.59%)、MOD15A2H LAI(25.54%)、VNP15A2H LAI(23.33%)和PROBA-V LAI(9.64%),融合LAI平均缺失频率降低为5.04%。与其他产品比较,融合LAI与地面实测值的相关性最好,决定系数(R2)达0.76,比其他产品高0.03~0.2,均方根误差(RMSE)为1.16 m^(2)/m^(2),比其他产品低(0.1~0.66)m^(2)/m^(2),具有较高的精度。

Leaf Area Index(LAI)is an important parameter for characterizing the canopy structure of vegetation.Remote sensing products of LAI are often affected by factors such as clouds,aerosols,snow,algorithms,and instrument issues,which result in poor data quality or missing data,seriously affecting the application of LAI datasets.In this paper,we propose a LAI time-series data fusion method based on improved S-G filtering and unsupervised classification local kernel regression.The normalized fusion is carried out using MODIS LAI,PROBA-V LAI,and VIIRS LAI products from 2014 to 2020 to improve the consistency,continuity,and accuracy of the products.Results show that the fused LAI shows good consistency in frequency distribution and temporal variation with the source products and other LAI products(MCD15A2H,MOD15A2H,VNP15A2H,PROBA-V),with a R2of 0.85,0.77,0.84,and 0.89,respectively.Compared to other four products,the frequency of data missing is generally reduced,and the temporal continuity is improved.The averaged missing frequency of the fused LAI is reduced to 5.04%,compared to MCD15A2H LAI(19.59%),MOD15A2H LAI(25.54%),VNP15A2H LAI(23.33%),and PROBA-V LAI(9.64%).Also,the fused LAI shows the strongest correlation with ground measurements,with a R2of 0.76,which is 0.03~0.20 higher than other products,and a root mean square error(RMSE)of 1.16 m^(2)/m^(2),which is lower than other products(0.10~0.66)m^(2)/m^(2),indicating a higher accuracy.