Urban Forest Vegetation Classification In Consideration Of Vegetation Phenology Based On High Spatial Resolution Satellite Imagery

Healthy, stable and sustainable urban forest development has been recognized as an important component of socialist ecological civilization construction. Urban forest classification is of great significance to make full use of the urban forest resources, and to adequately bring their ecological, social and economic benefits. Based on the triple-phase high resolution GF-2 remote sensing images coupled with field survey data, the major forest vegetation including bamboo forest, shrubs, evergreen coniferous forest and deciduous broad-leaved forest in Beijing Olympic forest park was classified by implementing two kinds of classification schemes in the current work. Specifically, in consideration of vegetation's phenologies, the multi-temporal spectral and textual characteristics of remotely sensed images were derived to construct spectral difference rate and texture difference rate models, followed by the extraction of different vegetation distribution based on hierarchical classification algorithm. On the other hand, machine learning based classifiers were also implemented to map urban forest vegetation. Firstly, pixel-based classification methods including the minimum distance classification, the maximum likelihood classification, the SVM classification, the decision tree classification and the random forest classification were applied to map urban forest vegetation, after feature optimization of spectral and textual characteristics by the maximum correlation and minimum redundancy algorithm. Meanwhile, based on object-oriented classification technology, applying KNN classification, decision tree classification, SVM classification and random forest classification to map urban forest vegetation was also done. Through comparing the accuracy, operability and efficiency of the two types of classification, the most suitable method for urban forest classification was determined. At the same time, the ability of GF-2 satellite images to map urban forest vegetation was evaluated. Finally, urban forest classification map was generated from the optimal classification algorithm to provide a decisive support for sustainable management of urban forests.The results showed that the classification method that takes vegetation phenologies into account was concise and effective when mapping urban forest vegetation, which illustrates that the changing rate of spectral features and textural features can act as an important key to distinguish urban forest vegetation. Through a comprehensive evaluation of the ten classification methods used in the study, the object-oriented random forest method was the best one for fast urban forest classification, with enough accuracy guarantee; under the high precision requirement, the pixel-based SVM method is the best way to obtain urban forest information, with an overall accuracy of 95.00% in the current work. In addition, the GF-2 satellite images' temporal-spatial resolution can satisfy the need of mapping urban forest, demonstrating great potential in urban forest resources survey applications.