The National Park Service contracted with GRS to perform a comprehensive land-cover classification of Lassen Volcanic National Park (LAVO) in northern California. The National Park Service had dual goals in this project: the first goal was to create two detailed vegetation maps for LAVO - one based on photo-interpretation and the other on image classification; the second goal was to compare traditional aerial photo‐interpreted map data set with the map data set developed using the Discrete Classification Mapping Methodology (DCMM) developed by GRS.
Initial project efforts by GRS entailed the procurement, review, and processing of the satellite imagery that covered the entire project area. Part of the initial effort involved processing the imagery with GRS’s illumination correction software to correct differential illumination due to slope and aspect within the imagery. The imagery was then stratified and sample site locations were chosen based on their size and homogeneity to assure full coverage of the diversity of land cover types within the park.
Field data collection efforts started in July 2006 and ran through October, 2007. Field data were collected at 654 locations within Lassen Volcanic National Park and within a five-mile buffer of the park using GRS’s line-point transect sampling methodology. GRS field crews sampled vegetation characteristics primarily using the line-point transect methodology and the GRS Densitometer in combination with the field-data collection software TransIn. In addition,FireMon sampling protocols were integrated into GRS’s sampling methodology to include sampling of fine and coarse woody debris encountered during the vegetation sampling efforts. Collected data was input into PDA’s while at the site so that it was verified as it was being entered. GPS locations and digital photos of all field sites were collected and recorded to provide documentation of what was encountered at each site.
GRS the applied the Discrete Classification Mapping Methodology to classify the satellite imagery. This methodology resulted in a highly detailed vegetation and landcover map that includes the assignment of categorical map values, such as National Vegetation Classification System names, as well as cover estimates for individual species that comprise the different mapped stands. Map data were developed in both a raster and vector format with no editing of any polygon boundaries. These classification efforts were started in April 2007 and finished in June 2008. Over 450 sites were used as training sites in the image classification training set. Confusion was evaluated and training site fidelity was checked. Final class maps were developed in early July, 2008. GRS’s pixel aggregation process GRS_aggregate was run on the final class maps to form polygons with a Minimum Mapping Unit size limit of 0.5 hectares. Draft image classification and photo-interpretation data sets were developed in late July 2008 and formed the basis for the Accuracy Assessment design and sample site selection.
Final maps were developed by the end of August, 2008. The DCMM map data sets were developed in both raster (unaggregated) and stand (polygon/vector) formats. DCMM map data sets included the typical NVCS Alliance and Association information, but in addition included individual species-specific/landscape feature cover estimates that enabled the mapping of the extent and magnitude of individual species/landscape features, rather than just color-coded type maps.
The Accuracy Assessment was designed by GRS using Stratified Random Sampling within the NVCS type classes. A total of 29 map classes (NVCS alliances and super alliances) were selected for testing with a minimum sample size of thirty samples per class. A set of candidate sites was randomly generated and AA sites were selected for each AA class using List Sampling. All sample selection and evaluation was performed specific to the individual map data sets and was performed independently of the other map data set. Due to unforeseen differences in the two maps, a total of over 1200 sample sites were determined to be necessary for the Accuracy Assessment because some sites could only be used to sample one map or the other and not both maps. To decrease this effort, minimum sample size per class was lowered to 20 or 25 samples per class. The revised AA sample included 910 sites. Over 565 randomly selected sites were visited in the 2008 field season before snowfall forced an early end to the field season. The remaining sites were visited in the 2009 field season. Target IDs were loaded into GPS devices and sites were located by field crews. Field crews were not aware of what map they were sampling or what the map type information was for any particular site. Crews developed cover estimates and used the LAVO Preliminary NVCS Type Key to derive type names for each AA site. Alternate calls were developed when field crews noted that the cover estimates were estimated at cover thresholds of the key. GPS locations and digital photos of all field sites were collected and recorded to provide documentation of what was encountered at each site.
All Accuracy Assessment data have been scored and used to develop contingency tables testing Generalized Alliances, Detailed Alliances/Associations, Cover Classes, and Size Classes. Results of the Accuracy Assessment will be included in the Project Report. In addition, accuracy by map type was included in the map data sets thereby enabling the visual display of map accuracy as shown the report Figure 14.
In addition, a statistical analysis was undertaken that showed that there was a relationship of probability of error to polygon size in the DCMM map data set; no such relationship existed in the APMM map data set.
The DRAFT versions of these map data sets have been released to LAVO staff as of May, 2014. The final report to be published by USGS has been delayed but is now scheduled to be completed by September, 2014.