This section outlines the classification procedure for the Guam High Resolution C-CAP. WorldView2 imagery used in producing this land cover product was also utilized in producing an associated impervious surfaces layer for the island. The mapping approach utilized a boundary summary, outlier change detection process and refinement procedure developed by Photo Science to leverage the 2005 high resolution C-CAP data set to create the 2011 high resolution C-CAP. A calibration visit was not conducted though NOAA had access to local resources for validation. Non impervious features were mapped using a 0.25 acre minimum mapping unit (MMU) and impervious features were mapped using a 0.1 acre MMU. Pre-processing steps: The WorldView2 mosaic utilized for this project was provided from Digital Globe as an orthorectified, georeferenced product. Multiple image primitives and indicies such as texture and NDVI were derived from the 8-band satellite data. The imagery was re-sampled from it's native 2m spatial resolution to 2.4m. The 2005 C-CAP classification and the base Quickbird imagery (used in the original mapping) were geometrically corrected to co-register to the WorldView2 data which has a higher geolocation accuracy.
Impervious Update: The 2005 impervious was over-laid with semi-transparency on the 2011 satellite data. It was panned at a scale of 1:3,500 and manually updated to match the 2011 data. Analysts zoomed in to a larger scale when necessary to perform edits. The product went through a QC procedure to ensure features were accurately captured.
Segmentation and Outlier Detection Process: The 2011 impervious was combined with the 2005 land cover to create a hybrid data set. Image segmentation, done in Trimble's eCognition software, was completed at multiple scales using the multispectral (2.4 m) imagery in order to group like spectral and textural objects within the imagery. For consistency, the associated hybrid data set was incorporated into the segmentation layer as a boundary delimiter. Segments contained image attributes and a label from the 2005 classification. These data were inputs to a custom multi-variate outlier detection tool that identified objects of potential change. These areas created the change mask.
Training Site Data: Training data was photo interpreted and collected within the change mask and was super-sampled from features outside of the mask. The referenced data set went through a QA/QC procedure to ensure it met accuracy standards.
Classification: Automated Classification - Image segments were classified using a decision tree classifier in See5 software based on the training data and image attributes. Automated Classification Refinement - Models are built to refine or reclassify land cover areas by utilizing the wealth of attribute information linked to each segment within eCognition. Classification Edits - As with any automated or semi-automated land cover classification there are often inconsistencies in the land cover map. The final step before map finalization was to remove inaccuracies through manual segment labeling as interpreted by an analyst. Map Finalization - Photo Science used independent reviewers comments to further refine the land cover map.
Attributes for this product are as follows: 0 Background 1 Unclassified 2 Impervious 3 4 5 Developed, Open Space 6 Cultivated Crops 7 Pasture/Hay 8 Grassland/Herbaceous 9 Deciduous Forest 10 Evergreen Forest 11 Mixed Forest 12 Scrub/Shrub 13 Palustrine Forested Wetland 14 Palustrine Scrub/Shrub Wetland 15 Palustrine Emergent Wetland 16 Estuarine Forested Wetland 17 Estuarine Scrub/Shrub Wetland 18 Estuarine Emergent Wetland 19 Unconsolidated Shore 20 Bare Land 21 Open Water 22 Palustrine Acquatic Bed