Study On The Reconstruction Of Time-series NDVI Data Set

Vegetation typically elicits dynamics at the seasonal and annual level. Time-series of Normalized Difference Vegetation Index (NDVI) datasets, such as the Pathfinder AVHRR Land (PAL) NDVI, MODIS and SPOT/VEGTATION NDVI dataset, have proven to be appropriate for the detection of long-term vegetation cover changes in regional, continental or global scales. They are also successfully applied to extract the biophysical parameters of vegetation cover. Normally, there are quite frequently fluctuations with marked rises and falls because of variable cloudiness, data transmission errors, incomplete or inconsistent atmospheric correction and bi-directional effects in the NDVI dataset. Even though the Maximum Value Composite (MVC) applied on the NDVI is commonly used to reduce these error sources, there is still so much noise which would affect further analysis and applications. So it is meaningful to study on NDVI reconstructing algorithms in development.At first, several widely used NDVI reconstruction algorithms were introduced by analyzing the aspect of principle, mathematical model. Then six algorithms, including Mean-value iteration filter (MVI), the modified best index slope extraction (BISE), Fourier Transform (FT), Savitzky-Golay filter (SG), Harmonic analysis of time series and Asymmetric Gaussian function fitting (AG), were compared and evaluated. The SPOT/VEGETATION NDVI data set were used as the original time-series data resource. The land cover type data were also used in this analysis. The reconstructed results were validated and assessed by using some in-suit NDVI measurements carried out during late May to early August,2009. Eight observation points are covered with different kinds of underlying surface (e.g. farmlands with corn, wheat, and rape, grassland, and desert).Compared with other five methods, we know:the MVI method detects most of the discrete noise, but ignores the continuous abnormal points. The AG fitting method recognizes the abnormal low values and generates a smooth curve, but it cannot recognize low values caused by harvest. Modified BISE and S-G filter method can fix abnormal low points, but they cannot recognize abnormal high points. FFT have poor performance because of a large deviation with the original data. HANTS algorithm is an improvement upon FFT filter which can deal with the problem of time-series data with different intervals, and recognize most abnormal values of the series, but change most of the original data. The analysis results indicate that the AG and MVI fitting methods have better reconstructed NDVI value.Finally, according to the results of the above analysis, we come up with two integrated approaches based on standard deviation weight and characteristics of noise points respectively. In the first integrated approach, the revision probability (RP) for each point was calculated. A point with bigger RP has larger probability to be a noise value. A RP threshold was set, and a point would be considered as a noise when its RP exceeds this threshold. The weighted average of the revised NDVI values reconstructed by these five methods was inserted to the new time-series instead of the noise. In the second integrated approach, the characteristics of noise points were analyzed at first. A best one of these five reconstruction methods was chosen to modify the noise according to the characteristics of each noise point. The criterion of noise is same with the former integrated approach. The result shows that these two integrated methods are better than the six separate methods above. They do not only retain most of the original data, but also modify the noise to the utmost extent. NDVI time series datasets produced by these two approaches combine with the advantages of the five reconstruction methods, which can be used to carry out the researches on global and regional environmental change, vegetation dynamic, and so on.