|
Human activity has changed land covers on the earth surface significantly over decades, especially in urban areas. Mapping the urban typical land covers is critical for the analysis of environment and sustainable urban management. With the availability of various remote sensed data, urban typical land covers extraction at different scale has attracted growing attention. Linear spectral unmixing is a widely-employed technique in mapping urban land covers, which can estimate urban component abundance at sub-pixel scale, popular especially in medium spatial resolution images. However, it suffers from the endmember selection, including types and numbers of endmembers, due to the intra-class spectral variability and the inter-class spectral similarity in the feature space constructed of original spectral bands. In this paper, we propose a novel technique for mapping urban typical land covers under linear spectral unmixing model in GaoFen-5 data at 20 m resolution. A main contribution of our new technique is that we develop a new spectral analysis technique which uses the spectral trend between every two bands instead of arbitrary spectral values, demonstrated effective in spectral representation of different urban land covers. The feature space constructed based on this spectral analysis model is discriminative, which makes endmember selection effective and efficient. To validate the superiority of our new spectral analysis technique, the linear spectral unmixing was also conducted based on the original spectral bands. Experimental results illustrated that the newly spectral analysis technique show promising performance for mapping urban typical land covers in terms of RMSE and SE.
|
