Application Of Spatiotcmporal Fusion In Regional Land-Cover Time Scries Classification

Accurate land cover classification results are great significance to promote the balance and protection of the ecological environment and the effective planning and adjustment of the land resource utilization structure.Land cover classification based on multi-temporal remote sensing data has better accuracy performance than classification based on single-temporal data.However,due to the inherent limitation of the imaging principle of remote sensing satellites,the frequency of repeated observations of medium and high resolution sensors is low,and the number of images is small,which seriously restricts the classification accuracy based on time series data.The spatio-temporal fusion technology is an effective method to solve the lack of time series observation data,but its application in the classification research based on time series data has not been fully verified.How to use the spatio-temporal fusion model to obtain higher accuracy for land use classification based on time series data is a crucial issue.To address this problem,this article takes Huludao City,Liaoning Province as the research area,takes Landsat-8 and MOD09A1 image data as the main remote sensing data sources,and takes STARFM(Spatial and Temporal Adaptive Reflence Fusion Model),ESTARFM(Enhanced Spatial and Temporal Adaptive Reflence Fusion Model)and Semi-Physical spatio-temporal fusion model are the means for generating annual Landsat-8 resolution time series data.Random Forest(RF),Maximum Likelihood Classification(MLC)and Support Vector Machine(SVM)Three methods are used as classifiers based on time series data classification,and perform a collaborative analysis of classification accuracy on the combination of different spatiotemporal fusion models and classifiers.The specific effects of the spatiotemporal fusion model,the number of time series data sets and the time span on the accuracy of multiple classifiers are explored.Research indicates:(1)In the experiment,the reconstruction accuracy of the data reconstructed by multiple spatiotemporal fusion models for different classifiers is greater than that of the classification accuracy of time series data.It shows that the spatiotemporal fusion model can effectively improve the classification accuracyof surface coverage,and is not sensitive to the choice of classifier.However,its improvement in classification accuracy is smaller than the classifier itself.(2)The spatial-temporal fusion model improves the classification accuracy of different features differently.Among them,the classification accuracy of water and artificial building features is relatively small,and the classification accuracy of cultivated land,forest land and grassland is relatively large.It indicates that the classification of surface cover based on time series data improves the accuracy mainly on vegetation and other features that are greatly affected by the season.(3)The quality of ESTARFM model reconstruction data is the best.The STARFM model is next,and the quality of semi-physical model reconstruction results is relatively poor.Among the corresponding classification results,the ESTARFM model reconstruction data has the largest improvement in accuracy for various classifiers,followed by the STARFM model and the semi-physical model.It shows that the improvement of classification accuracy is positively correlated with the quality of the reconstructed data of the spatiotemporal fusion model.(4)In the case of continuously increasing the number of time-series images,the classification accuracy of various classifiers shows a trend of increasing first and then tending to a stable.In the classification based on time series data,when the frequency of the time series data is one scene per month,or when the time range covered by it is a time period in which the phenological characteristics change greatly,the classification accuracy will reach a relative maximum value.