Comparisons Of Two Spatial-Temporal Filter Based Data Fusion Algorithms By Using Remote Sensing LST Data

With the continuous advancement of technology,remote sensing has rapidly developed into a stage of multi-platform,multi-sensor and multi-angle observation.However,due to the constraints between the existing technologies and imaging indicators,the earth surface information observed by a single remote sensing system is not comprehensive.It is still very difficult to obtain data with high time-,spaceand spectral-resolution.Spatio-temporal data fusion is a kind of multi-source remote sensing data fusion methods,and an effective way to solve this dilemma.It can break through the performance of a single sensor,take advantages of multi-platform complementary observation effectively,and achieve more accurate and comprehensive land surface information.At present,spatial-temporal data fusion algorithms can be divided into three categories: fusion methods based on mixed pixel decomposition,fusion methods based on spatial-temporal filtering,and fusion methods based on sparse expression.The fusion algorithms based on spatial-temporal filtering is widely used because of its simple framework,easy operation and high prediction accuracy.Such algorithms generally perform mathematical operations through multi-source,multi-temporal remote sensing data.Because of the similar pixels were used in this kind algorithm to generate the fused image,the quality of it is one of the main factors affecting the prediction accuracy.While there are different selection methods to produce similar pixels and can be generally divided into 2 categories as local and non-local,the exsiting studies rarely discussed the influence of different selection methods on the fusion algorithms based on spatial-temporal filtering.Therefore,2 algorithms using local and non-local filtering methods respectively(ESTARFM algorithm and STNLFFM algorithm)were selected in this study.8 Landsat 8 OLI/TIRS images of different phases and the corresponding MODIS MOD11A1 data products(located in Henan Province and Fujian Province)were used to produce 4 sets of experiments.Through the experiments,the accuracy of the land surface temperature(LST)images which were fused by the 2 algorithms was compared,and the factors which may influence the the accuracy of the LST images were analyzed.In addition,the study also modified the original STNLFFM algorithm to improve its applicability when using MODIS land surface products.The results show that:(1)The modified STNLFFM algorithm can deal with the MODIS MOD11A1 LST products which has null values better than original algorithm,the results are significantly improved compare to the original algorithm's results;(2)from the perspectives of texture,value domain,etc,the LST fusion results of the local and non-local filtering algorithms are highly similar to the observed images and have the potential to replace the observed images expect some individual results.Statistical analysis indicates the results of the non-local filtering algorithm(STNLFFM)are more closely to the observed results;(3)For the residual,the LST residual distribution has significant regional differences for the same spatio-temporal data fusion algorithm;the differences of fusion algorithms tinily affected the LST residuals for the same region.(4)The LST residual distribution is closely related to the type of land cover,and it is positively correlated with vegetation coverage,but negatively correlated with construction land.