Research On Land Cover Classification Method Using Annual Time Series Landsat Data

Landsat satellite data has been the main source of large scale and mesoscale land cover classification because of its high quality and accessibility.However,some traditional classification methods are more dependent on the single-phase data,the classification result is not good,and the classification effect of different types of vegetation cover areas such as farmland,grassland,deciduous forest land and evergreen forest land is poor.Although many studies have also tried to adopt multi-temporal images into classification tasks,they mostly depend on simple algebraic operation or linear combination for characteristic structuring,making the discriminant of feature not significant.In fact,some Landsat images have the "strip" problem,and some areas often have cloud/snow covering.These complex data conditions pose great challenges to related research.In this regard,a land cover classification method based on annual time series Landsat data(LandUTime)is proposed,which aims to overcome the above limitations,improve the classification accuracy,and ensure the applicability,stability and robustness of the algorithm.Around the classification method,main research contents and achievements are as follows.(1)A feature construction algorithm based on regression of annual time series data is proposedThe algorithm uses annual time series Landsat images as input,uses regression analysis techniques to model the change pattern of land surface,and constructs the classification features from time series models.In this paper,a series of verification experiments were carried out in two research areas.The proposed algorithm obtained overall classification accuracy of 88.76% and 89.44% and improved the accuracy of 8.85% and 9.51% compared with traditional methods based on single image.The results show that it is feasible to construct classification features from annual time series remote sensing images based on regression analysis.This method is effective for improving the classification accuracy of land cover.(2)Analyzed the importance of time series model selection for ensuring the validity of featuresIn the experiment of using time series images to construct features,by comparing various methods,we found that the method based on statistical values(e.g.mean and median)has a certain effect on improving the classification accuracy,but it does not fully excavate the information contained in time series data,and causes a certain amount of wasting data.The accuracy is no more than 85%.Among the methods based on regression analysis,the polynomial model does not consider annual change patterns of land surface,neglects the missing data,and seeks to minimize the fitting error of existing data,which leads to the poor classification result.However,the model proposed in this paper has a high accuracy in modeling phenological changes and a strong ability to integrate and extract the feature information.(3)An ensemble classification algorithm based on parallel feature subspace is proposedBased on "feature blocks",the original temporal feature space is cut into parallel subspace,and then the base classifier is integrated hierarchically,maintaining the relationship between temporal features.At the same time,the heterogeneity and integration of base classifiers is enhanced by using Bootstrapping and subspace strategies.In a series of control experiments conducted in two research areas,the overall accuracy of the proposed method is 90.46% and 91.67%,which is higher than the other similar algorithms.The efficiency of proposed algorithm is also high.Compared with random forest and k-nearest neighbors algorithm,it makes a balance between the accuracy and operation efficiency.(4)The discriminant ability of LandUTime for different land cover types is summarizedExperiments show that LandUTime can effectively improve the accuracy of land cover classification by modeling annual variations of the ground and integrating multiple classifiers.When comparing the results of different categories,it was found that LandUTime has a strong ability to identify deciduous forest,grassland,farmland,and water bodies,but it is somewhat inadequate in distinguishing developed land,bare land,and evergreen forest.Further analysis shows that the distance metric adopted in base classifiers has a certain influence on the category judgment.The comprehensive research results of this paper show that the LandUTime land cover classification method is effective in solving the problems encountered by traditional methods.The method is stable and efficient,and has wide application prospect.