Algorithms Of Snow Cover Extraction From Multi-source Remote Sensing Data Based On Adaboost And Its Application

Snow disaster is one of the most frequently happened disasters in China. It causes heavy influences on the development of animal husbandry, people's production and lives, and national economy. Extracting the snow cover rapidly and accurately is important for assessments of the snow disaster intensities and elements. It is also helpful for disaster reduction in pastoral areas. Remote sensing technique has many advantages such as high spatial resolution and large coverage. It plays an important role in snow cover extraction, especially for pastoral areas.Based on remote sensing data acquired by MODIS instruments and HJ-1-B satellite, this dissertation concentrates on the developments of algorithms and models and their applications. Firstly, the algorithms for extracting snow cover at sub-pixel and pixel scales are proposed. Secondly, models for snow disaster assessments are developed based on GIS techniques of spatial analysis. The main research contents include:(1) An algorithm of extracting the sub-pixel snow cover from MODIS data is proposed. Using the 30-m HJ-1-B data as ground truth to calculate snow fraction for MODIS data pixel by pixel, an algorithm is presented for MODIS data to obtain sub-pixel snow cover information based on a regression relationship between the normalized difference snow index (NDSI) and snow fraction. This method is applied and evaluated by selecting Yushu located in Qinghai province as the study area. Results demonstrate that the relationship between snow fraction and NDSI is significant. The mean absolute error of snow fraction calculated based on NDSI is lower than 0.2. Furthermore, this research finds that both the land surface temperature (LST) and the normalized difference vegetation index (NDVI) are negatively correlated with the snow fraction. This finding provides the possibility for estimating snow fraction with multiple factors based on remote sensing data.(2) An Adaboost-based algorithm for extracting snow cover from HJ-1-B data is developed. AdaBoost is an adaptive machine-learning algorithm it constructs a "strong" classifier by combining a set of "simple weak" classifiers. Firstly, features for extracting the snow cover are constructed according to the fact that snow reflectance is high in the visible wavelengths (0.4～0.7μm) and low in the shortwave infrared range (1～4μm). The sensors on board the HJ-1-B satellite obtains eight bands, among which band1, band2 and band3 of the CCD locate in the visible range, while band 4 of CCD, band 1 and band2 of IRS locate in the near-infrared range. Therefore, one visible and one near-infrared band of the above six bands are selected to calculate the NDSI value and then 9 features are derived. Ultimately,15 features are chosen as inputs for the proposed algorithm, including 9 NDSI features and the first six HJ-l-B bands. Secondly, the best feature is selected as a weak classifier in each iterative learning step, and all the weak classifiers are used to construct the strong classifier. Thirdly, the strong classifier is applied to the simulated data sets to evaluate the proposed algorithm, and the results show that the AdaBoost algorithm based on the combination of classifiers can obtain better accuracy than using a single classifier. The proposed algorithm is further applied to extract snow cover from HJ-1-B satellite image. Results indicate that the Adaboost-based algorithm can not only improve the identification accuracy of snow pixels, but also automatically select the optimal combination of different bands for calculating the NDSI and determine the appropriate threshold.(3) Snow disaster assessment. Based on the previous researches about algorithms of snow cover extraction, five topics about snow disaster assessment are selected for further investigation according to the spatial and temporal changes of snow disaster and the workflow of disaster reduction, These five topics are assessments of the snow disaster intensity, snow disaster spatial extent,, the influence of snow disaster on the animal husbandry, the influence of snow disaster on the lifeline damages, and the influence of snow disaster on the population. For each topic, the corresponding evaluation model is proposed. Selecting the snow disaster in Altai pastoral area in 2009 as a case study, the algorithms and models developed in this research are applied and the snow disaster intensity and the influences on the animal husbandry are evaluated. It is demonstrated that the results of assessment are consistent with the actual situations.