Research On Data Reconstruction And Time Series Analysis Method Of MODIS Land Surface Products

Massive amounts of remote sensing data have been accumulated in the past 40 years in which period space remote sensing has experienced an enormous upswing. However, the effective observation data of the ground surface is hardly to obtain when the process of earth observing using remote sensing was affected by the cloud-cover. It may create the problems that discontinuous space and irregular time span of the remote sensing data. Affected by these problems, the application level of the remote sensing time-series data were declined, and the deep investigation of the rule hiding in the temporal dimension of abundant data was limited. In this context, how to reconstruct the missing or low-quality remote sensing data and how to analysis the reconstructed data has become an unavoidable problem that needs immediate settlement.To solve the above problems, the work of data reconstruction and time-series analysis were done based on the NDVI(Normalized Difference Vegetation Index) and LST(Land Surface Temperature) which are representative in the MODIS Land Surface Products. Based on the spatial-temporal distribution features of The NDVI and LST data, fractal interpolation algorithm and stepwise regression model was designed to reconstruct the NDVI data and LST data respectively. And on this basis, the reconstructed data were analyzed deeply to seek the information contained in their time dimension. Detailed research contents and results are listed in the following:(1)The spatial-temporal characteristics of cloud-cover in the study area were analyzed and the necessity of reconstructing data was expounded quantitatively. There are four indexes proposed to describe the spatial-temporal characteristics of cloud-cover which are mensal probability of cloud-free, mensal average ratio of cloud-free, > 80% mensal period ratio of cloud-free and mensal mode of continuous cloud-free period respectively. Results show that cloud-cover have significantly effect on the remote sensing data in the study area. The value of mensal average ratio of cloud-free ranges from 18.95% to 57.19% and significantly changes with time and space. The north region is less affected by the cloud-cover than the south region of the study area which was divided into northern and southern halves by the 36°N. From the time series analysis, cloud-cover had a relatively minimal impact in March and April.(2)The special stationarity and fractal feature of NDVI were researched. Results show that NDVI has special stationarity in low elevation area, but it has remarkable non-stationary characteristic in high elevation area such as western Sichuan, southern Shaanxi and northeast Tibet. In addition, in the urban built-up area, river area and farmland or woodland mixed crisscross area, NDVI also has remarkable non-stationary characteristic. NDVI line(row) can be regarded as a fractal set because the section line of NDVI line(row) has obvious fractal feature. It is exactly based on this fractal set that the section line of NDVI line(row) was sampled and calculated, resulting in that box dimension is between 1.30 to 1.60.(3)The fractal interpolation reconstruction algorithm of NDVI data was designed based on the fractal feature of NDVI. In the algorithm, firstly the initial point set was determined by grouping method, then di was determined by analytic method, at last precision checkout point set C was designed to control the accuracy of Iterated Function System(IFS). Precision analysis result show that the precision of fractal interpolation has little influence on the respond rules of spatial missing scale of NDVI. The precision of fractal interpolation is similar with interpolation precision of OK when the spatial missing scale is small, while the precision of fractal interpolation is superior to the precision of OK and IDW. What’s more, fractal interpolation can keep more texture details of features compared with spatial interpolation method. NDVI reconstruction data is the base data of the LST reconstruction data.(4) LST time series reconstruction algorithm was designed based on the correlation analysis of LST and elevation, NDVI, longitude and latitude. Backward regression method was applying in the algorithm to screen independent variables. The optimal regressive function was fitted though compression filter of the total regression and expansion filter of single factor model using AIC principle. Results show that error of reconstruction data is small. 71.8% of the data error in the two time-point of the day and 78.2% of the data error in the two time-point of the night can be limited to the level of 3℃. More than 90% of the overall data error can be limited to the level of 5℃. Measured data from meteorology station need a scales expansion when the accuracy of reconstruction is verified. For all this, a method of scales expanding the measured data is proposed in this paper.(5) Error correction model to reconstruction LST data in cloud-cover area was designed which increase the estimation accuracy of LST in cloud-cover area. This model utilizes the jump features of the effect of sunshine duration on LST data. Day of low sunshine duration in a period is regarded as the criterion condition in this model.(6) Proximity fuzzy classification algorithm was designed based on adding-windows DTW distance. Firstly, standard time series curves of all land-use types were got using sample data and iterative computation. Then standard fuzzy data set was determined though DTW distance and each pixel were classified by the proximity method. Results show that total accuracy is high and overall accuracy can reach 83.8% and Kappa coefficient is 0.77.This method can be used as vegetation information extraction of the time series data in the years without sampling data.(7)The time series characteristic of LST reconstruction data was analysed for different height region and different land types in the study area. Results show that annual mean LST of different elevation region show a parallel distribution characteristic. The annual mean LST in the region where the elevation is in the range of 1~2km is 2℃lower than the LST in the region where the elevation is under 1km and is 2.6℃ higher than the LST in the region where the elevation is above 2km. The annual mean LST in different elevation region all showed a trend of slow growth with the time. The ranking from high to low for ground temperature of different land types is paddy-land, dry-land and wood-land. Though ground temperature variance analysis, it was found that the ground temperature data of the 7-11 phase and the 24-26 phases can assist to classify the paddy-land and dry-land. Moreover, seasonal effect affects LST more than NDVI.(8) The time-lag relationship between LST and NDVI was researched by developing vector regression model of LST and NDVI time series data. LST and NDVI time series data are both first-order integration variables. VAR(7),VAR(5) and VAR(2)model were developed for paddy-land, dry-land and wood-land respectively. With Granger causality analysis, it was found that time-lag variables of LST and NDVI have stronger ability to explain NDVI. Impulse analysis indicates that LST would bring homonymous impact to NDVI when it suffers a shock from external conditions. Meanwhile, impact duration and intensity is not the same for different land types.The reconstruction of the two kinds of representative data which are NDVI and LST in the study area from the year of 2005 to 2014 have been realized using reconstruction algorithm proposed in this paper. The work improves the continuity of the two kinds data in time and space. The spatial-temporal distribution features of NDVI are utilized fully to extract the vegetation information, thus get a high total accuracy. Vector auto-regression method is adopted to analyze the time-delay relationship of LST and NDVI time series data of different land types, founding that the lagged variables of LST and NDVI have remarkable effects on the NDVI.