Object-Oriented Classification Based On Time Series

Land use change reflects the development and utilization of earth surface resources,and remote sensing image is the data source to obtain land use information.It is of great significance to acquire the status quo of land use in a large area quickly through the interpretation of remote sensing images.With the rapid development of remote sensing,long time series satellite images have been accumulated.Time series data can provide more characteristics related to time.At the same time,object oriented classification transforms the classification target from a single pixel to a homogeneous object,which can provide a more realistic and multi-level ground object boundary.Therefore,object-oriented classification combined with remote sensing images time series data is of great significance for the interpretation of ground object types.In order to construct an object-oriented classification method suitable for Wuqi County with typical characteristics of hilly and gully region on the Loess Plateau,this paper firstly extracted multi-source NDVI(Normalized Difference Vegetation Index)data using Landsat 8OLI and MODIS13Q1 images as data sources.Spatiotemporal fusion algorithm was used to generate NDVI time series data of 23 issues in 2018 with high spatiotemporal resolution.The smoothing effect of different filtering functions is compared and analyzed,and the optimal filtering algorithm is selected.The filtered NDVI time series data were used to extract phenological images to provide more time series information for classification.Secondly,a multi-dimensional feature system including spectrum,shape,texture,remote sensing index,topography,and phenology was constructed,and the optimal feature combination was selected through feature optimization.Then,the key parameters of object-oriented classification were determined,the contribution degree of different bands was analyzed,and the shape standard,compactness factor and other segmentation parameters suitable for the characteristics of the study area were determined by the control variable method,and the multi-scale segmentation classification hierarchy was constructed.Four classification methods,including Bayes classification,CART decision tree classification,Random Forest classification(RF)and Support Vector Machine(SVM),were used to realize hierarchical extraction of different ground features.Finally,an object-oriented classification method suitable for the typical characteristics of the hilly and gully region of the Loess Plateau is constructed.The main conclusions are as follows:(1)The NDVI time series data after fusion retained the spatial and temporal resolution characteristics of Landsat and MODIS images.Double Logistic filtering algorithm is more suitable for Wuqi County terrain characteristics,smooth reconstruction effect is better.(2)Based on the evaluation of feature weights,search and optimization,and calculation of correlation coefficients,the dimensionality reduction of feature space is realized,and a comprehensive feature system covering spectrum,shape,texture,remote sensing index,topography,and phenology is constructed.(3)After the increase the weight of near infrared wave band to participate in the segmentation,to determine the shape factor of 0.7,compact factor 0.6 key segmentation parameters,formed a "split-classification-merge-split-classification" of the multi-level classification rules.A three-tier classification system of cultivated land-water area-forest land,grassland and construction land was constructed.Finally,the optimal segmentation scale of different land features was determined: the cultivated land and water were 115 and 95 respectively,and the forest land,grassland and construction land were 65.(4)Random forest classification was determined to be an object oriented machine learning method more suitable for Wuqi County.The overall classification accuracy reached92.13%,and the Kappa coefficient reached 0.90.The Bayesian classification had a good effect on construction land,water area and forest land.The CART decision tree classifier has a high distinction among forest land,water area and grassland.In the random forest classifier,cultivated land,forest land,grassland,construction land and water area all achieved high mapping accuracy.Support vector machine classifier is better in water area and cultivated land extraction.