| The monitoring of snow parameters is of great significance to the prediction of snow disaster,the early warning of snow disaster,the agricultural irrigation,the transportation in regional areas.It's also significant to the meteorology,hydrology,climate change in global areas,and other fields.Remote sensing techniques play an important role in large-scale and high-frequency snow monitoring.Nowadays,among many remote sensing methods of snow monitoring,NDSI(Normalized Difference Snow Index)is the most common snow index for daily snow recognition in non-forest areas.However,the existing studies use fixed NDSI threshold in the same area to identify snow in non-forest areas,ignoring the temporal variation of snow spectral information,which caused by the change of external factors and snow physical properties,this leading to monitoring errors of snow recognition.NDFSI(Normalized Difference Forest Snow Index)is practical application index of remote sensing for daily snow recognition in forest areas.However,the existing methods use the fixed NDFSI threshold for snow recognition in forest areas,ignoring the different influence of forest canopy structure and forest coverage rate to snow spectral signals,this leading to the commission errors of snow recognition in forest areas.The area with the same snow depth but larger snow grain size has stronger scattering effect on the microwave signal.And the above area has larger bright temperature difference of different frequencies received by the sensor.If the influence of snow grain size is not considered,the snow depth will be overestimated.However,the existing semi empirical brightness temperature gradient algorithms do not fully consider the influence of spatial variation and the daily variation of snow grain size on snow depth inversion.So,the accuracy of snow depth inversion needs to be improved.In this paper,we propose an adjust method to identify daily snow in non-forest areas by dynamic NDSI thresholds,which reduces the impact of spectral fluctuation on snow recognition and improves the recognition accuracy in non-forest areas.Wepropose an adjust method to identify daily snow in forest areas by dynamic NDFSI thresholds,which improves the monitoring accuracy in forest areas.We establish the model of snow depth inversion by considering the spatial variation and daily variation of snow cover rate and snow grain size.The main studies and conclusions are as follows:(1)Taking the Sanjiangyuan ecological environment protection area with bare land and withered grass during snow cover period as the study area.The monitoring results based on 30 m resolution Landsat OLI(Operational Land Imager)data are used as the “ground truth” to calibrate the best NDSI threshold for snow recognition based on 500 m resolution MODIS(Medium Resolution Imaging Spectrometer)data of same period.We establish a correlation function model between the optimal NDSI threshold for snow determination based on MODIS data and the average NDSI value of pure permanent snow pixels in the same period MODIS data,and the determination coefficient of the model reaches 0.86.Compared with the monitoring method with fixed NDSI threshold of 0.33 based on MODIS data,the overall classification accuracy and the value of F with dynamic NDSI thresholds based on MODIS data can be improved by 2.62% and 5.05% at most,and 0.14% and 0.13% at least,respectively.When the dynamic NDSI thresholds are lower than the fixed NDSI threshold value of 0.33,the omission error of the dynamic NDSI thresholds method can be reduced by 17.76% and 1.49% at most and at least,and the recall rate can be increased by 12.14% and 1.49% at most and at least,respectively.When the dynamic NDSI thresholds are higher than the fixed NDSI threshold value of 0.33,the commission error of the dynamic NDSI thresholds method can be reduced by 5.23%and 0.34% at most and at least,and the snow classification accuracy can be increased by 3.84% and 0.31% at most and at least,respectively.The spatial differences between the dynamic NDSI thresholds method and the fixed NDSI threshold value of0.33 are mainly concentrated in the mixed pixels in the snow edge areas.(2)The Haihe river basin with abundant forest resources is taken as the study area.From the high resolution and cloudless Google Earth satellite image,the snow samples with the size of 500 m×500 m in evergreen forests and deciduous forests areselected.In the MODIS image of the same period,the spectral characteristics of the pixels where the above snow samples are located are counted.In the land cover data,the forest coverage of the pixels where the above snow samples are located is counted.There is a good linear correlation relationship between the NDFSI threshold of snow recognition based on MODIS data and the forest coverage rate in the evergreen forest areas or in the deciduous forest areas,and the determination coefficients are 0.91 in the evergreen forest areas and 0.92 in the deciduous forest areas.The results of snow recognition by dynamic NDFSI thresholds based on MODIS data have higher accuracy.The overall classification accuracy,the value of F and snow classification accuracy of the dynamic NDFSI thresholds are 3.16%,2.68%,5.01% higher than the fixed NDFSI threshold of 0.38 in the evergreen forest areas,and 5.00%,4.61%,8.35% higher than the fixed NDFSI threshold of 0.37 in the deciduous forest areas.Compared with the fixed NDFSI threshold,the commission error of the dynamic NDFSI thresholds recognition method is reduced by 6.52% in the evergreen forest areas,and 9.30% in the deciduous forest areas.The dynamic NDFSI thresholds can significantly reduce the commission error.(3)The Sanjiangyuan area covered by less forest coverage is taken as the research area.The measured snow depth data of the ground meteorological station,the MWRI(Microwave Radiation Imager)data of 25 km resolution,the snow coverage rate of 25 km scale and the snow grain size of 25 km scale retrieved from optical MODIS data are used for multiple regression analysis.And the 25 km scale snow depth inversion model under the influence of snow cover rate and snow grain size is established with a determination coefficient of 0.79.The error of snow depth inversion using above model is smaller and RMSE(Root Mean Squared Error)is only1.24 cm,which is better than the other model.According to the idea of linear decomposition of mixed pixels,a snow depth inversion model of 500 m scale is derived by downscaling the 25 km scale model under the influence of snow cover rate and snow grain size. |
PDF Full Text Request