A classification-based assessment of the optimal spectral and spatial resolutions for Great Lakes coastal wetland imagery

摘要

We analyzed hyperspectral airborne imagery (CASI 2 with 46 contiguous VIS/NIR bands) that was acquired over a Lake Huron coastal wetland. To support detailed Great Lakes coastal wetland mapping, the optimal spatial resolution of imagery was determined to be less than 2 m. There was a 23% change in classification resiliency using the SAM classifier upon resampling the original 1-meter, 18-band imagery to 2-meter pixels, and further classifications with larger pixels (4 and 8 m) increased overall classification change to 35% and 50%, respectively. We performed a series of image classification experiments incorporating three independent band selection methodologies (derivative magnitude, fixed interval and derivative histogram), in order to explore the effects of spectral resampling on classification resiliency. This research verified that a minimum of seven, strategically located bands in the VIS-NIR wavelength region (425.4 nm, 514.9 nm, 560.1 nm, 685.5 nm, 731.5 nm, 812.3 nm and 916.7 nm) are necessary to maintain a classification resiliency above the 85% threshold. Significantly, these seven bands produced the highest classification resiliency using the fewest number of bands of any of the 63 band-reduction strategies that were tested. Analyzing only derivative magnitudes proved to be an unreliable tool to identify optimal bands. The fixed interval method was adversely influenced by the starting band location, making its implementation problematic. The combined use of derivative magnitude and frequency of occurrence appears to be the best method to determine the "optimal" bands for a wetland mapping hyperspectral application.