Abstract:

【Objective】 In order to provide a reliable foundation for the sustainable development of Nanjing City, this study compared different endmember extraction methods and mixed pixel decomposition algorithms to calculate the coverage of urban impervious surfaces (UISs) in Nanjing in 2018, and evaluated the accuracy of each method. 【Method】 Using Landsat 8 OLI images, an improved spatial pixel purity index (SPPI), based on the pixel purity index (PPI), was first proposed to refine or purify the endmembers for four land cover types including vegetation, bare soil, high albedo UISs and low albedo UISs. Next, based on the two extracted suites of endmembers derived from PPI and SPPI, a linear spectral mixed model, a mixture tuned matched filtering, a bilinear mixed spectral model and a BP neural network were implemented to extract UISs in Nanjing. The extracted UIS abundance was validated using visual interpretation of high spatial resolution Google Earth images for 2018. 【Result】 The results showed that SPPI could improve the identification accuracy of endmembers effectively and reduce the computational load by combining the spectral information from multi-spectral bands with the spatial heterogeneity of the panchromatic band. The UIS extraction accuracy of the SPPI-based BP network algorithm was the highest, at 90.45%, while the PPI-based linear spectral mixed model was the lowest at 80.62%. Overall, the BP network model outperformed the other three models in UIS extraction, regardless of whether PPI-based or SPPI-based endmembers were used. 【Conclusion】 The proposed method of using the panchromatic band to assist with extracting endmembers effectively integrated spatial heterogeneity to compensate for inadequate spectral information in the multispectral bands. The study provides methodological references to improve or develop means for endmember extraction from medium/high-resolution multispectral images. Additionally, integration of the SPPI and BP network model can be popularized in the UIS extraction domain.

Key words: urban impervious surface, spatial pixel purity index(SPPI), mixture tuned matched filtering(MTMF), bilinear mixed spectral model(BMM), BP neural network(BPNN), linear mixing spectral model(LMM)