Abstract:

With much higher spectral resolution, hyperspectral remote sensing data has higher potential capability to identify land cover types than traditional multispectral data. But under limited training sample size, increased dimension of remote sensing data means decreased samples/dimension ratio, which can lead to low classification accuracy if common statistic based pattern classification methods were used. One scene of EO-1 Hyperion hyperspectral data was acquired for the test site in north-east of China. In order to aid for ground true information collection, 2.5 m SPOT-5 PAN image was segmented into self-closure polygons. Detailed ground true data was surveyed according to the boundary of each polygon. Based on these ground true data, the hyperspectral data was used to validate the forest types identification accuracy of several advanced statistic classification methods. Finally, one classification scheme being able to effectively solve the small train sample problems for forest type classification using hyperspectral data was suggested. It was shown that forest type classification accuracy can be improved if advanced feature extraction method, much more effective second order statistic parameter estimation method, and context-sensitive samples classifier such as ECHO was applied.

Key words: EO-1 Hyperion, hyperspectral, statistic pattern recognition, forest type