Land Cover Classification In High Latitude Region Based On MODIS Data

Land cover refers to the synthesis of vegetation cover and artificial cover on the surface of the earth,which is always changing.The development of human society accelerates the process of land cover change.Understanding the information of land cover change is of great significance to guide the production and life of human beings.Land cover mapping is the main means to understand land cover change information.In the early stage,land cover mapping was completed by manual mapping,but with the development of remote sensing technology,automatic and semi-automatic mapping based on remote sensing images gradually became the main mapping methods.Nowadays,the research of land cover classification method based on remote sensing image and regional large-scale land cover classification mapping have become the hot issues in the field of remote sensing application research.This paper takes high latitude area--Canada as the research area and MODIS(Moderate Resolution Imaging Spectroradiometer)land standard product MOD13Q1 in 2010 as the data source.Firstly,the classification feature importance assessment scheme is designed to select the optimal classification feature combination based on the optional feature combination.Then,the optimal classifier selection scheme was designed based on alternative models(including classical machine learning model and advanced deep learning model)to select the optimal classifier.Finally,the generalization ability of the proposed classification scheme was evaluated by conducting regional large-scale land cover classification mapping across the whole country of Canada in 2010.The results show that the combination of MODIS time series data of four surface reflectance bands can produce the highest classification accuracy compared with other bands.In this paper,combining the advantages of FCN(Fully convolutional networks)model based on image blocks with spectral and spatial residual blocks,the deep learning model SS-Dip(Spectral and Spatial Residual Neural Network Based on Deep Image Prior)proposed in comparison with other methods--SVM(Support Vector Machine),RF(Random Forest),SS-Res(Spectral and Spatial Residual Neural Network),Spec-Dip(Spectral Residual Neural Network Based on Deep Image Prior)can achieve the highest classification accuracy.The SS-Dip model has been trained and tested in different provinces respectively,which shows that the model has good generalization performance.The SS-Dip model can still achieve high classification accuracy when tested in the whole Canada,which indicates that the classification scheme proposed in this study are feasible and effective for regional large-scale mapping in high latitudes.In conclusion,the classification scheme proposed in this paper has great classification effect and good generalization performance,and has a broad application prospect in large scale mapping in high latitude areas.