Research On The Construction Methods Of Long Time Series High Spatial And Temporal Resolution NDVI Dataset

Long time series NDVI with high spatial resolution plays significant role in examining regional long-term dynamics of the vegetated land surface and climate impacts,etc.. However,observed data could be severely missed for the persistent presence of atmospheric contamination and snow cover in the NDVI time series. Various data fusion models have been developed to figure out these problems,but most of which failed to reconstruct NDVI with long time interval and long time series. To achieve the goal,we proposed two long time series and high spatial-temporal NDVI reconstruction method named Long Timeseries Linear Fusion Model, LTLFM,and Long Time-series Spatial and Temporal Adaptive Fusion Model,LTSTAFM,both of which take a fourier fitting model with 8 parameters as basis,but separately using a linear fusion method and spatial and temporal adaptive method in contrast. To assess the model accuracy,multi-source data form GIMMS、Landsat and MODIS covering Ruoergai area were used to carry out the experiment. Research shows that:1) Fourier fitting method shows capacity in reconstructing time series NDVI dataset with a certain precision and well simulating the NDVI seasonal and regular change information. But due to the limitations of the data fitting algorithm itself, the method cannot reflect enough anomaly information;2) The LTSTAFM model which developed from STARFM is capable to simulate regularly and anomaly vegetation sign with a relative high accuracy. In comparing with observations, the model results showed a low root mean square error(RMSE)(0.0378), a high coefficient(0.08651 in average) and an average slope of 0.9993, which quite well meet regional scale research.3) There is a perfect agreement between the LTLFM model results and the observations in tone and texture features with the NDVI of ice,snow or cloud covered areas reconstructed,and show a good consistency with the sounding areas without any patch in the image which reflecting coarse resolution data features. Besides, it showed a high correlation and low RMSE between the model results and the observations,coupled with a normal distribution of the error of which up to 75.04% pixels in average within ±0.05 and up to 97.65% in average within ±0.1.4) Compared with Fourier fitting method, LTSTAFM and LTLFM model trend to produce images with clearer texture and detail structure, and a better correlation in time series with observations(MODIS).5) Although the difference between LTLFM and LTSTAFM model results is not quite significant, the LTSTAFM model shows a better performance in some severe disturbance area such as river, mountain, farmland, etc., while the LTLFM model more suitable for the long timeseries NDVI reconstruction of homogeneous areas. Besides, LTSTAFM model produce an overall slight lower results than LTLFM.6) Though both the two long time series and high spatial-temporal NDVI reconstruction models achieve a relatively high accuracy result, issues in data source, accuracy verification, applicability and ability in change detection remain to discussed.