| Moderate-resolution imaging spectroradiometer(MODIS)snow-cover products have relatively low accuracy because of complex terrain and shallow,fragmented snow cover in the Tibetan Plateau.The snow distribution in the area is significantly affected by factors such as topography,geographical location,surface temperature,and vegetation index.In this study,fractional snow-cover(FSC)mapping algorithms were developed using the linear regression model,spectral mixture analysis model,back-propagation artifificial neural network model(BP-ANN),random forest model(RF)and support vector machine model(SVM)based on unmanned aerial vehicle(UAV)data and satellite remote sensing data(Landsat 8 OLI and MODIS of version 006).The accuracies of these models were validated against Landsat 8Operational Land Imager(OLI)snow-cover maps(True_FSC)and compared with the MODIS global FSC product(MOD10A1 FSC,version 005)for the purpose of finding the optimal algorithm for FSC mapping in the Tibetan Plateau.The aims are to explore: 1)The influence of factors such as topography,geographical location,temperature and vegetation on the snow cover distribution in the study area,and the influence of these factors on the accuracy improvement on the retrieval models of FSC;2)The adaptability and contribution of UAV technology on the detection of FSC over the Tibetan Plateau;3)The retrieval accuracy and adaptability of MODIS FSC models of the linear regression model,spectral mixture analysis model and machine learning model;4)Which kind of MODIS FSC model is the best model for the entire Tibetan Plateau with the best combination of retrieval accuracy,operational efficiency and stability.Results indicates:(1)The factors contributing to the FSC retrieval model in the study area are arranged in descending order,namely the Normalized Difference of Snow Index(NDSI),the surface reflectance of band 3th(R3),Normalized Difference of Vegetation Index(NDVI),and the Land Surface.Temperature(LST),altitude(DEM),surface reflectance of band 7th(R7),latitude,longitude and surface reflectance of band 5th(R5).Among them,NDSI,R3,DEM and Latitude are positively correlated with FSC,while R5,R7,NDVI,LST and Longtitude show negatively correlated with the FSC.(2)The UAV technology obtained snow cover data with spatial resolution up to centimeter level and the introduction of UAV technology has improved the retrieval accuracy of the MODIS FSC effectly.Combined with satellite remote sensing data,the retrieval accuracy of the unary exponential regression model(EXP_FSC)based on empirical algorithm is not only higher than univariate linear regression model based on Landsat 8 OLI data(Landsat_FSC),but also higher than that of the spectral mixture analysis model(MIX_FSC)based on spectral analysis algorithm.In addition,The accuracy of the three machine learning models(UAV_RF,UAV_SVM,UAV_ANN)based on UAV data(UAV_FSC)also are better than MIX_FSC model and the regression models(EXP_FSC and M6_FSC).(3)Snow,bare land,vegetation,water,soil,and cloud are the types of features corresponding to the endmembers in the entire study area extracted by PPI and n-dimensional visualization tools,respectively.The accuracy of the linear regression model(M6_FSC),random forest model(Landsat_RF),support vector machine model(Landsat_SVM)and BP neural network model(Landsat_ANN)based on large sample of Landsat 8 OLI data(Landsat_FSC)are better than the similar models based on small sample of UAV data(EXP_FSC,UAV_RF,UAV_SVM,UAV_ANN).The MIX_FSC is the worst model in this study,but still has better accuracy than the MODIS V005 version standard FSC product(Standard FSC).The accuracy of these models in grassland,shrub and bare land is higher than that in other areas,such as forests and farmland.But the percentage improvement in accuracy of forests,farmland and other areas is higher than that of grassland.(4)In this study,the accuracy of models from high to low is Landsat_RF(“ACC “ ranged from 72.49% to 78.49%),Landsat_SVM(71.00%~77.13%),Landsat_ANN(70.50%~76.59%),UAV_RF(69.49%~76.30%),UAV_SVM(68.94%~76.26%),UAV_ANN(67.83%~76.09%),M6_FSC(66.86%~75.27%),EXP_FSC(65.87%~73.66%),MIX_FSC(64.83%~72.92%),Standard FSC(46.60%~65.33%).The Landsat_RF model based on NDSI,R3,R5,R7,NDVI,LST,DEM,Longtitude and Latitude is the optimal model in FSC extraction both accuracy and stability,operational efficiency and operability.It is the relatively optimal FSC snow mapping algorithm for the large-scale snow monitoring in the Tibetan Plateau. |
PDF Full Text Request