Statement |
This map was created by the United States Department of Agriculture (USDA), who classified multi-part Landsat ETM+ imagery scenes that were pan-sharpened for the year 2000. Cloudy parts of the reference dates for each scene were replaced by image data from cloud-free parts of other image dates. Scene parts for non-reference dates had been radiometrically matched to reference dates for each scene with regression tree models, and cloud-filled scenes were radiometrically matched to each other with histogram matching. Ancillary data was combined with satellite imagery to create an island-wide predictor variable dataset for image classification. Topographic variables, derived from USGS 30 m digital elevation models (DEM) included elevation, slope, aspect, curvature and facet. Climatic variables included mean annual precipitation and temperature. Vector derived variables included distance to roads, distance to streams (ghuts) and distance to coastlines. Finally, band indices and ratios were created from the Landsat image bands to produce the normalized difference vegetation index (NDVI) and the 4/5 band ratio. The ancillary predictor data was spatially co-registered with the cloud free image mosaics for the two time periods and stacked with the native Landsat reflectance bands 1-5 and 7 at a spectral resolution of 15 meters, resulting in a 18 band image for each classification time period. Field reference surveys were performed during a reconnaissance visit in 2004 to collect training and reference data. The island wide field surveys and expert consultation with regional vegetation experts enabled us to discern land cover attributes and forest formation from 1 m IKONOS panchromatic sharpened imagery and 1m color DOQQs. Land-cover and forest formation were then identified in the satellite imagery based on a comparison to areas interpreted in reference imagery and the spatial location of field reference survey data. Forest formation was identifiable in both the reference imagery and the Landsat imagery by a number of factors including spectral components such as color, contrast, tone and texture as well as spatial indicators including geology and elevation. Difficulties distinguishing forest formation were encountered in transitional areas including semi-deciduous and seasonal evergreen forest. Field survey data collected meeting this criteria proved useful in the identification of uncertain forest transitions. The decision tree classifier See5 (www.rulequest.com) was used for landcover and forest formation classification of Landsat ETM+ imagery for the year 2000. Training data polygons for each land-cover and land-use class were converted to an ASCII point file providing spatial coordinates, class name and assigned class number. The classification was performed with a 10 trial adaptive boosting option to improve the overall accuracy by combining many decision trees into a single combined classifier. This process provides more predicted data to analyze before determining each final class (www.rulequest.com). The default global pruning option was also used to reduce the likelihood of over fitting the tree to the training data. The pruning process removes predicted parts of the decision tree with relatively high error rates and makes decisions as to the final class assignment. Manual edit confused classes using 1 meter IKONOS multispectral imagery and 1 meter DOQQ reference data. Manually digitize areas where remaining cloud and cloud shadow artifacts exist in image mosaics. A stratified random sample generated approximately 50 validation points for each landcover class. The 1 m IKONOS imagery (dates ranging from 2000 to 2003) and DOQQ (2004) data were used to verify the accuracy assessment points for each class in the classified images. Using high resolution imagery to validate the classification eliminated restrictions on point locations, permitting us to include points that would otherwise not be accessible because of topography, remoteness, or property ownership. An error matrix was produced for each mapped class that estimates the overall percentage of correctly classified pixels, producer and user's accuracies, and the kappa coefficient. For publication to EMODnet, SAERI reprojected published geoTIFF file to EPSG:4326 - WGS 84 coordinate system, polygonised the file to an ESRI Shapefile, and processed the structure to fit specific data exchange format. |