Research For The Validation System And Consistency Of Hyperspectral Mineral Maps Using AVIRIS Scenes Of Cuprite,Nevada,USA

Hyperspectral remote sensing is one of the frequently used tools and techniques for geologic exploration of mineral resources.The product of a hyperspectral study often is a mineral map whose accuracy needs to be validated.The existing accuracy assessment system for hyperspectral products is unstandardized and relies on ground truth data.To fill this niche,this research proposed an idea that the reproducibility and/or consistency of mineral-maps derived from multitemporal hyperspectral images over the same area may be correlated with the accuracy of their classification results.This idea depends on the fact that the mineral distribution barely changed during a short period of time.Therefore,the aim of this research was to measure the reproducibility of hyperspectral mineral maps,investigate assumed factors(e.g.vegetation cover,different spectral endmembers,and classification threshold)that may disturb reproducibility and estimate the relationship between reproducibility and accuracy of hyperspectral classification.In this study,three multitemporal AVIRIS scenes acquired in May 2006,September 2008 and October 2010 respectively and over the Cuprite,Nevada,USA were used for hyperspectral mineral mapping.These three scenes were subsetted then classified by Spectral Angle Mapper(SAM)with the spectral endmembers extracted from the images using both various threshold and the unique threshold.Mask tool was used to mask out the pixels that have vastly different values(more than ±100 for reflectance)between the 3 pairs of images(May image and Sep image,May image and Oct image,as well as Sep image and Oct image).This process resulted in similarized images then these similarized images were classified by the various threshold set of SAM and the extracted endmembers to create new mineral maps.The Normalized Difference Vegetation Index(NDVI)was used to analyse the vegetation cover in the study area.Universal endmembers were created by averaging the three set of extracted endmember libraries and these were used to create minerals maps from the original images.Subsequently,all the classification results were compared to investigate reproducibility and the factors that affect it.Eventually,the reproduced classification results were compared with the ground truth to assess the accuracy of classification.The classification results created by the extracted endmembers with the various thresholds shown reproducibility around 60%.However,the classification results created with the extracted endmembers and the unique threshold shown poor classification.The similarity of identified classes between mineral maps derived from the similarized May image and similarized Sep image is around 23%,between mineral maps derived from the similarized May image and similarized Oct image is around 50%,between mineral maps derived from the similarized Sep image and similarized Oct image is around 48%.On the contrary,the similarity of identified classes between mineral maps derived from the original images is around 22%.The reproducibility of mineral maps that created with the same set of the universal spectral endmembers is around 85%.The comparisons between the producer’s accuracy of mineral maps that were created with the universal endmembers and reproducibility of the same set of maps shown a similar variation for every single mineral.Higher ratios of reproducibility indicated that differences of the original hyperspectral scenes and the differences amongst different endmember libraries influenced the reproducibility of hyperspectral mineral maps.The presence of similar variations in trend between the reproducibility and producer’s accuracy shown the potential of using reproducibility to assess the hyperspectral classification accuracy.