随着卫星技术的进步，监测地表动态变化的要求也逐渐提高。然而由于卫星传感器的空间分辨率和时间分辨率的限制，现有的大部分遥感时间序列产品都不能满足这一需求。目前的时间序列产品通常具有较高的时间分辨率（如MODIS和AVHRR），然而其空间分辨率却往往较粗糙，无法刻画异质区域的空间细节。与之相反的是，Landsat系列传感器能提供细空间分辨率的影像产品。然而由于重访周期长和云污染等问题，这些产品的时间覆盖十分稀疏。为了结合两种数据集的优势，近年来已有数十种时空融合算法被开发用于重建高时空分辨率的遥感时间序列数据。虽然这些算法在发表时均号称在预测准确性、计算效率和对输入数据的要求等方面具有各自的优势，但是由于不同研究中实验设置和输入数据的差异，用户仍然很难确定哪种方法在各方面表现较为优秀。
现有针对时空融合算法的比较研究基本都忽略了一些在实际应用中不可避免的影响因素，导致如何针对特定需求选择合适算法成为一个难题。为了解决这个的问题，本研究结合理论分析和模拟实验，比较了六种经典的时空融合算法对于不同影响因素（输入数据的时空异质性、几何配准误差、辐射不一致性误差和空间分辨率）的敏感性。本研究选用的六种经典的时空融合方法分别为UBDF、LMGM、STARFM、Fit-FC、OPDL和FSDAF，主要结论如下：
（1） 在不同的时空异质性的输入条件下，Fit-FC和FSDAF两种方法始终能比其余四种方法生产出更高精度的NDVI时间序列；
（2） 相比于其余的时空融合模型，采用回归加权模型的Fit-FC对于几何误差具有最好的鲁棒性，然而这种回归模型在一定程度上保留甚至是放大了输入数据的辐射误差；
（3） 采用增量加权模型的FSDAF能部分消除输入数据中的辐射不一致性，然而其融合结果对于输入数据中的几何误差较为敏感；
（4） 在较大的空间分辨率之比的输入情形，Fit-FC和FSDAF两种方法仍能取得较为不错的结果，因为这两种方法都能很好地捕捉NDVI数据中的空间细节；
（5） 结合增量加权和回归加权两种方法的优势有望在未来开发一种更加鲁棒的算法。
以上发现可以帮助用户根据他们自己独特的需求，选择适合于不同遥感数据集的时空融合算法，同时为算法开发人员提供一定的指导。

With the advance of satellite sensors, high requirements are called for monitoring land surface dynamics with details. However, due to the limitation of spatial resolution and temporal frequency of satellite sensors, most of the existing remote sensing time-series products cannot satisfy this need. Normally, the time-series product (MODIS, and AVHRR) possesses a high temporal frequency but at a coarse spatial resolution, lacking of spatial details for heterogeneous areas. By contrast, Landsat series and Sentinel-2 can provide fine spatial resolution imagery, but their temporal coverage is sparse due to the long revisit cycle and cloud contamination. Therefore, to combine the advantages of these two datasets, in recent years dozens of spatiotemporal fusion methods have been developed to reconstruct time-series data with both high spatial resolution and frequent coverage. Although each developed method was claimed to have unique advantages in terms of prediction accuracy, computation efficiency, and input data requirements, it was still difficult for users to reach a consensus on which method outperforms all the others under a particular condition because of the differences in experimental settings and input data. 
Although several current studies comparing the different fusion methods have been conducted, selecting the suitable fusion methods to meet different requirements is still challenging, as inevitable influential factors tend to be neglected. To address this problem, this study combined theoretical analysis and simulation experiments to comprehensively compare the sensitivities to different influential factors (i.e., spatiotemporal variations, geometric misregistration errors, radiometric inconsistencies, and spatial resolution ratio) of six typical spatiotemporal fusion methods, including the Unmixing-Based Data Fusion (UBDF), Linear Mixing Growth Model (LMGM), Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), regression model Fitting, spatial Filtering and residual Compensation (Fit-FC), One Pair Dictionary-Learning method (OPDL), and Flexible Spatiotemporal DAta Fusion (FSDAF). The main conclusions of this study are as follows:
(1) Fit-FC and FSDAF can achieve better performances in producing NDVI time-series than the other four fusion methods considering various spatiotemporal conditions;
(2) Fit-FC employs the regression weighting to achieve the strongest resistance to geometric errors among the six fusion models, but it is sensitive to radiometric inconsistency because such a model maintains or even amplifies radiometric errors to some extent. 
(3) FSDAF utilizes the increment weighting to mitigate radiometric inconsistency but its fusion result is sensitive to geometric errors;
(4) Fit-FC and FSDAF can generate satisfactory results even with a large spatial resolution ratio of input data because they can well capture spatial details of NDVI;
(5) Combining the advantages of increment-weighting and regression-weighting methods might develop a more robust fusion method in the future. 
These findings could help users determine the appropriate method for different remote sensing datasets according to their unique requirements and provide guidelines for developers in the future development of novel methods.