Urban Image Classification Using Multi-Angle Very-High Resolution Satellite Data.

摘要

The ability to automatically generate large-area land-use/land-cover (LU/LC) classification maps from very-high spatial resolution (VHR) satellite data is dependent on two capabilities: (1) the ability to create a data model able to accurately classify satellite data into the appropriate surface types and (2) the ability to apply this model to the multiple images necessary to create a large-area VHR mosaic. This research describes methods for improving these capabilities by leveraging the unique characteristics of VHR in-track and composite multi-angle data.;It is shown that new features can be extracted from both in-track and composite multi-angle data in order to improve classification performance. These features encode information extracted from the spatial and spectral variations of the multi-angle data, such as spectral fluctuation with view-angle and pixel height. This additional knowledge provides the capability to both improve image classification performance (29% in the demonstrated experiments) and include urban LU/LC classes, such as bridges, high-volume highways, and parking lots, that are normally difficult to identify in multispectral urban data.;Additionally, methods that apply a multispectral classification model across multiple images (model portability) are also explored using the simplifying test cases of in-track and composite multi-angle data. The in-track results show that the portability of a multispectral model can be improved from no portability (losing all classification capability when applying the model across the multi-angle images) to a 10% reduction in kappa coefficient across the sequence of in-track images when physically based image normalization techniques are appropriately applied. The additional noise of seasonality limits the portability performance in the composite multi-angle sequence to an approximate reduction in kappa coefficient of 20% in the best cases.