Development of the New Hamsphire Land Cover Assessment was made possible by financial support from the Cooperative Institute for Coastal and Estuarine Environmental Technology CICEET), USDA Forest Service, NH Department of Resources and Economic Development, NH Department of Fish and Game, USDA Natural Resources Conservation Service, NH Space Grant, and UNH Cooperative Extension.
Please cite as "New Hampshire GRANIT. 2001. New Hampshire Land Cover Assessment. New Hampshire GRANIT, Durham, NH."
CLASS - Code PRODUCER'S ACC. USER'S ACC. Residential/Commercial/Industrial - 110 86.9% 88.3% Transportation - 140 100.0% 85.0% Row Crops - 211 94.6% 88.3% Hay/Pasture - 212 84.6% 91.7% Orchards - 221 97.4% 92.5% Beech/Oak - 412 68.1% 53.3% Paper Birch/ Aspen - 414 28.6% 28.6% Other Hardwood - 419 53.2% 70.0% White/Red Pine - 421 90.7% 81.7% Spruce/Fir - 422 93.8% 80.4% Hemlock - 423 95.1% 65.0% Pitch Pine - 424 100.0% 97.5% Mixed Forest - 430 39.7% 62.5% Alpine (Krumholz) - 440 100.0% 80.0% Water - 500 100.0% 100.0% Forested Wetland - 610 74.3% 86.7% Open Wetland - 620 88.2% 75.0% Tidal Wetland - 630 100.0% 100.0% Disturbed - 710 90.0% 90.0% Bedrock/ Veg. - 720 100.0% 100.0% Sand Dunes - 730 100.0% 100.0% Other Cleared - 790 82.4% 93.3% Tundra - 800 100.0% 100.0%When the classification is collapsed to the 17-class level, the overall accuracy is 88.4%, and the User's and Producer's Accuracies are as follows:
CLASS - Code PRODUCER'S ACC. USER'S ACC. Residential/Commercial/Industrial - 110 86.9% 88.3% Transportation - 140 100.0% 85.0% Crops/Pasture - 211-212 95.0% 95.8% Orchards - 221 97.4% 92.5% Deciduous Forest - 410-419 90.7% 94.8% Coniferous Forest - 420-429 97.3% 81.9% Mixed Forest - 430 39.7% 62.5% Alpine (Krumholz) - 440 100.0% 80.0% Water - 500 100.0% 100.0% Forested Wetland - 610 74.3% 86.7% Open Wetland - 620 88.2% 75.0% Tidal Wetland - 630 100.0% 100.0% Disturbed - 710 90.0% 90.0% Bedrock/ Veg. - 720 100.0% 100.0% Sand Dunes - 730 100.0% 100.0% Other Cleared - 790 82.4% 93.3% Tundra - 800 100.0% 100.0%So that users can interpret the data most effectively, rules were created to develop broader ("fuzzier") categories of "right" and "wrong" and to assess the accuracy using these fuzzy sets. We applied the linguistic scale developed by Woodcock and Gopal (2000):
(1) Absolutely wrong: This answer is absolutely unacceptable.
Very wrong.
(2) Understandable but wrong: Not a good answer. There is
something about the site that makes the answer understandable,
but there is clearly a better answer. This answer would pose a
problem for users of the map. Not right.
(3) Reasonable or acceptable answer: May not be the best possible
answer but it is acceptable; this answer does not pose a problem
to the user if it is seen on the map. Right.
(4) Good answer: Would be happy to find this answer on the map.
Very right.
(5) Absolutely right: No doubt about the match. Perfect.
Each accuracy assessment site was given a fuzzy rating (see fuzzyratings.pdf for definitions). The overall accuracy of the 23- class classification increases to 89.1% when the "good answers" are included as "right," and to 92.0% when "reasonable or acceptable answers" are included as well. Please see the project's final report for a full discussion of the accuracy assessment.
Image Type Path-Row Bands Date Georeferencing/ Terrain Correction performed by:Ancillary data comprised numerous holdings from the GRANIT archive (the NH statewide GIS), including watershed boundaries, panchromatic Digital Orthophotoquads (DOQs), Digital Raster Graphics (DRGs), USGS Digital Line Graphs (DLGs) for hydrography, NH Department of Transportation road centerlines, Digital Elevation Models (DEMs), SPOT panchromatic (10 meter resolution) images, protected lands, and US Fish and Wildlife Service National Wetlands Inventory (NWI) maps.Landsat 5 TM 12-30 1-7 8-Sep-90 CSRC Landsat 5 TM 12-30 1-7 14-May-94 USGS Landsat 5 TM 12-30 1-7 24-Oct-95 CSRC Landsat 5 TM 12-30 1-7 22-Jul-96 USGS Landsat 5 TM 13-29 1-7 13-May-91 USGS Landsat 5 TM 13-29 1-5, 7 6-Oct-92 USGS Landsat 5 TM 13-29 1-7 12-Oct-94 USGS Landsat 7 ETM+ 13-29 1-8 31-Aug-99 ImageLinks, Inc. Landsat 5 TM 13-30 1-5, 7 6-Oct-92 USGS Landsat 5 TM 13-30 1-7 28-Oct-94 USGS Landsat 5 TM 13-30 1-7 14-Apr-98 USGS Landsat 7 ETM+ 13-30 1-8 31-Aug-99 ImageLinks, Inc.
The first product generated for each region was a generalized data set. Archived data from previous projects were used to create representative signatures, and a supervised, maximum likelihood classification was applied to each image subset. These classifications grouped each image subset into five broad categories: deciduous forest, coniferous forest, mixed forest, agricultural/cleared, and wetlands/water. The resulting classes were visually evaluated using DOQ's, other ancillary data sets, and local knowledge. Acceptable classes were carried through to the final data set, while unacceptable classes were used to mask various image band combinations and/or band transformations. This was followed by unsupervised classifications using the ISODATA cluster routine. At each iteration, the generalized classes were evaluated, and either archived for incorporation in the interim generalized product or retained for additional processing. As many as four supervised and unsupervised classification iterations using various image date/band derivatives were run on the resulting data sets. Finally, each of the general classifications was recoded to reflect the appropriate land cover value and mosaicked to generate the full, region-wide generalized land cover data set.
Class-specific classifications were accomplished through a series of image subsets, masks, and classification iterations to produce the final product. Each class-specific procedure was initialized by creating a layer-stack of various bands/band derivations. These were selected in part by applying the ERDAS Imagine signature separability tool to the layer stack and using the Transformed Divergence measure. Once bands were selected, the image composite was masked to retain pixels of interest (e.g., the forest-specific classification retained forested classes from the generalized land cover). This was followed by an iterative process of classifications using a combination of techniques (similar to that of the general classification) to derive the final data for that class.
The series of specific classifications typically began with a supervised classification, using both archived training sites and training sites collected for this project. Over 1,400 new data points were collected to supplement 1,200 archived sites from previous projects. A large number of non-forested sites were available from pre-existing sources, such as DOQ's, DRG's, NWI, and local knowledge. Forested sites, as well as some wetland and agricultural sites, required field sampling. Field crews navigated to each site using a Trimble Pro-XRS GPS receiver obtaining real time corrections, and at each forested location conducted two to four 10 BAF prism tallies to quantify the canopy composition.
In the southeast, the three forested classes (coniferous, deciduous, and mixed) from the generalized land cover were each processed independently, while in the north and southwest regions, the three classes were processed together because it was determined that there was no appreciable improvement in classification quality by separating the three.
As with the general classification, there was a series of iterative classifications from which acceptable results were saved to a final data layer and unacceptable results were used to mask subsequent data sets. For the forested classes, 14 iterations were needed to achieve an acceptable data layer. A total of 2,794 training signatures were used in these classifications (though in some cases the same training site was used to produce signatures for multiple images). For the cleared sites, 542 signatures were used in 12 classifications, and 126 signatures were used in 6 wetlands classifications. Our use of NWI data and the ISODATA clustering routine reduced the number of signatures needed to classify wetlands.
Some ancillary data were applied in this process as well: NWI data were used as a mask in the North Country to help distinguish many forested wetlands from spruce/fir forests; orchards were screen digitized from DRG's and DOQ's; and other data sets such as DRG's and DOQ's were used to determine the reliability of classes. Elevation data from USGS digital elevation models were used to change forest classes based on certain thresholds. Beech/Oak above 2,500 feet and Other Hardwoods above 3,000 feet were converted to Paper Birch/Aspen; White/Red Pine above 1,500 feet and Hemlock above 2,400 feet were converted to Spruce/Fir; and any forested class above 4,200 feet was converted to Alpine (Krumholz).
Several post processing refinements were applied to the provisional land cover data in the ESRI Grid environment. NH Department of Transportation road data (resident in the GRANIT data base, 2001) were "burned in" to the land cover data set, effectively overwriting any coincident class. Also, DLG hydrography data were used to update double banked river, lake, and pond edges. Finally, several filters were applied to remove speckling and produce minimum map units of one acre. In order to maintain the integrity of linear features, filtering was preceded by the REGIONGROUP command, such that the majority filter applied would only operate on groups of pixels smaller than approximately one acre (five pixels). This filter was followed by a second REGIONGROUP and contiguous pixels in sets less than five were finally NIBBLED to eliminate those pixels that were not eliminated by the majority filter.
A total of 975 sites were evaluated for the accuracy assessment. More than 600 of these were field visited, and others were evaluated using ancillary data such as NWI maps, DOQ's, and TM imagery. All sites classified as forest or agriculture, and most classified as wetland, were field visited using Trimble Pro-XRS GPS units receiving real time differential correction. At forested sites, field crews recorded stand information and conducted up to five 10 BAF prism tallies to quantify stand composition.
As with the classification itself, the accuracy assessment was conducted separately for each of the three geographic regions. In each region, we attempted to sample 30 sites per land cover class, but in some cases we were unable to do so because of limited area covered by the class, post data-collection re-classification, or for other reasons. Conversely, some classes were over-sampled, because of post data-collection re-classification or because we decided to merge subclasses. In order to limit distortion due to disparate sample sizes among classes, we randomly selected 20 sites from each class in each region to tabulate in the error matrices. This yielded a total of 60 sites per class for the full state (though some classes, particularly those like Tundra that are regionally focused, still have fewer sample sites).
Error matrices were generated for the Level 3 (23 class), Level 2 (16 class) and Level 1 (7 class) classifications, and user's and producer's accuracy were calculated. Additionally, the Level 3 classification was assessed using fuzzy set rules. See the Attribute Accuracy report above and the project's final report for more information.
New Hampshire Land Cover Assessment Data Key Developed 110 Residential, commercial, or industrial 140 Transportation Active agricultural land 211 Row crops 212 Hay/rotation/permanent pasture 221 Fruit orchards Forested 412 Beech/oak 414 Paper birch/aspen 419 Other hardwoods 421 White/red pine 422 Spruce/fir 423 Hemlock 424 Pitch pine 430 Mixed forest 440 Alpine (Krumholz) Water 500 Open water Wetlands 610 Forested wetlands 620 Non-forested wetlands 630 Tidal wetlands Barren Land 710 Disturbed 720 Bedrock/vegetated 730 Sand dunes 790 Cleared/other open Tundra 800 Tundra
Beech/oak stands (412) are deciduous stands comprising at least 30% beech and oak. Paper birch/aspen stands (414) are deciduous stands comprising at least 20% paper birch and aspen. Other deciduous stands (419) are deciduous stands not meeting either the beech/oak or paper birch/aspen criteria.
White/red pine stands (421) are coniferous stands in which white and red pine constitute a plurality of the coniferous basal area. Spruce/fir stands (422) are coniferous stands in which spruce and fir constitute a plurality of the coniferous basal area. Hemlock stands (423) are coniferous stands in which hemlock constitutes a plurality of the coniferous basal area. Pitch pine stands (424) are coniferous stands in which pitch pine constitutes a plurality of the coniferous basal area.
Other class definitions are as follows:
Developed (110) - built-up areas. (Note that this class was coded as
100 in early releases of the data.)
Active agriculture (200) - hay fields, row crops, plowed fields, etc.
Water (500) - lakes, ponds, some rivers or any other open water feature.
Wetlands (600) - areas dominated by wetland characteristics defined by
the U. S. Fish and Wildlife Service National Wetlands Inventory.
Basically hydric soils, hydrophytic vegetation and the hydrologic
conditions that result in water at or near the surface for extended
periods of the growing season.
Disturbed (710) - gravel pits, quarries or other areas where the earth
and vegetation have been altered or exposed.
Bedrock/vegetated (720) - exposed bedrock or ledge (usually in the
mountains) that may have some forms of stunted vegetation growing in
cracks or lichens growing on the surface rock.
Sand dunes (730) - areas along the seacoast that are dominated by sand.
Cleared/other open (790) - clear cut forest, old agricultural fields
that are reverting to forest, etc.
Tundra (800) - areas dominated by short vegetation that occurs above
tree line in the White Mountains (only mapped on Mt Washington). (Note
that this class was coded as 810 in early releases of the data.)