A Study On The Extraction Of Snow Covered Areas Based On Multi-feature In Northern Xinjiang

The snow is one of the most active natural elements of the Earth's surface. Snowcharacteristics (such as snow cover, snow grain size and snow water equivalent) isimportant input parameters in the global energy and water balance model and thehydrological cycle model. Snow cover can affect the weather, industrial andagricultural water resources, and a series of other elements associated with humanactivities and life. One hand Snow can provide extremely valuable the snowmelt waterresources and tourism resources; On the other hand a large area of snow may cause avariety of snowstorm (including avalanches, snow drift, etc.), and snowmelt water canalso cause floods and landslides. The common ground of two results are both due tothe distribution by a range of snow cover. Therefore, whether simulated snowmeltrunoff process, the feedback and impact of the climate, or the snowstorm early warningand forecast, snow cover mapping are the primary foundation and guarantee.Agricultural irrigation in arid and semi-arid areas is highly dependent on snowsnowmelt, snow water resources is extremely important for the fragile ecologicalenvironment. Snow resources in Xinjiang is very rich, the agricultural production has avery close relationship with the seasonal snow. But it is also one of the importantfactors that cause the most serious seasonal snow disasters in agriculture and animalhusbandry. Therefore, the snow monitoring has great significance for Snowmelt runoffsimulation and hydrological forecasts, early Warning of snow disaster. Northern regionof Xinjiang is a region of many rivers confluence, real-time analysis processing of thesnow information, and timely and accurate to provide snow condition monitoringreports to the relevant government departments is of great significance for floodforecasting and disaster prevention and mitigation. The high altitude and remotemountainous areas cannot be all-round monitored snow information and reflect thedistribution of snow. The application of remote sensing technology can just make upfor this deficiency, remote sensing technology with its fast, macro, comprehensive andmulti-temporal advantage to become the primary means of large-scale monitoring ofsnow covered.In this paper, take the middle of Tianshan in the northern region of Xinjiang forexample, the use of multi-source remote sensing data (MODIS data and Landsat-5TMdata), the fieldwork of the study area as well as a number of meteorological andhydrological data, using Remote Sensing and GIS means, to carry out the extraction ofsnow cover based on improved SVM. First using SNOWMAP algorithm to classifiedof the study area remote sensing image, while carrying out the study of fittingrelationship between NDSI, DSI, RSI and snow cover fraction, the use of three fittingmethods that linear fitting, exponential fitting, logarithmic fitting, the results found thatthe DSI and RSI can be used instead the NDSI to estimate the rate of snow cover anddistinguish between cloud/snow, and exponential fitting statistical relationshipbetween snow index and snow cover fraction is better than traditional linear fittingmethod. After the initial identification of the study area snow, based on analysis of thestudy area snow covered Characteristics and its affecting factors, the mountain fringe areas of the thin widowed snow covered is difficult to extracting, as well as mountainshadow effects that snow are hard to effectively extract out.These for established snowarea extraction methods provide the basis. The MNF transform scatterplot get the snow,rock, shadow three endmember of study area, take the three endmember into the linearspectral decomposition model, to get snow abundant degree map of the study area.Then use NWFE method for feature extraction, the method than the PCA method morefocuses on separability information of prominent surface features of interest.NWFEfor feature extraction is a better choice to further improve the accuracy of the featureextraction. Before using the improved SVM to extracted snow cover, firstly discussedand analyzed the combinations of different texture features, different band, differentwindow size.By the analyze, we obtain classification optimal texture features is mean,contrast, correlation, variance. The optimal band is the seven band together. Theoptimal window size is3×3. By SCF and linear spectral separation of snow abundantdegree as the spectral characteristics of the information input to the SVM model toclassify remote sensing images for the study area. While we use the before analysisconclusions that the optimal texture features, the optimal band combination with snowspectral feature by SVM method at the optimal window size to classification the studyarea, then compared with traditional SVM.Results indicated that by taking advantageof the snow spectral feature as the input characteristics of SVM can achieve superiorresults to the traditional SVM classification results, the overall classification accuracywas increased by0.2702%. Considered combining the texture feature extracted basedon GLCM for classification, the overall accuracy which improved1.081%was higherthan the traditional SVM method, the mapping accuracy of99.01%. This suggests thatthe SVM method that combined with snow spectral feature and texture featureextraction of GLCM could be effective in snow area extraction from MODIS data ofrelatively low spatial resolutions. This approach allowed the classification ofinformation becomes more abundant, thus facilitating the extraction of information.Classification method proposed in this paper can adapt to the nonlinear relationshipbetween the features combination. By adding spectral feature vector which havesignificant relation with snow area and texture features as the SVM input vector, it canadjust the snow area extraction where the land cover types is lack of training samplesand improve the overall accuracy of regional snow area estimation.