NLCD 1992-2001 Retrofit Change Product
What is the NLCD 1992/2001 Retrofit Land Cover Change Product?
Although one of the guiding principles of the NLCD 2001 design was to maintain as much compatibility with NLCD 1992 as possible, there were enough differences in the classifications to confound any direct comparison of the two datasets. Taking into consideration inherent differences in source image seasonality and georegistration, mapping methodologies, classification accuracy, and map legends, it became clear that a direct post classification comparison of NLCD 1992 and NLCD 2001 would be subject to unavoidable error.
Recognizing both the temptation and limitation of direct comparison of NLCD 1992 with NLCD 2001, the USGS NLCD design team initiated a research effort to devise an optimal way to compare the products. To generate the product, a multistage processing method used both NLCD products in procedures that included (1) reclassification of both eras of land cover using a decision tree classifier at Anderson Level I, (2) filtering intermediate results with confidence parameters, (3) determining changed versus nonchanged pixels, and (4) labeling the final change product using a "from-to" change classification code. A screenshot of the change product is provided below:
Unchanged Pixel Legend Example
What are the class definitions for the NLCD 1992/2001 Retrofit Change Product?
| Modified | Anderson Level 1 Class Codes and Descriptions |
|---|---|
| 0 | No Data |
| 1 | Open Water - All areas of open water, generally with less than 25% vegetation or soil cover. Pixels coded to a value of 1 have not changed between 1992 and 2001. |
| 2 | Urban - Includes developed open spaces with a mixture of some constructed materials, but mostly vegetation in the form of lawn grasses such as large-lot single-family housing units, parks, golf courses, and vegetation planted in developed settings for recreation, erosion control, or aesthetic purposes. Also included are lands of low, medium, and high intensity with a mixture of constructed materials and vegetation, such as single-family housing units, multifamily housing units, and areas of retail, commercial, and industrial uses. Pixels coded to a value of 2 have not changed between 1992 and 2001. |
| 3 | Barren - Barren areas of bedrock, desert pavement, scarps, talus, slides, volcanic material, glacial debris, sand dunes, strip mines, gravel pits, and other accumulations of earthen material. Generally, vegetation accounts for less than 15% of total cover. Pixels coded to a value of 3 have not changed between 1992 and 2001. |
| 4 | Forest - Areas dominated by trees generally taller than 5 meters, and greater than 20% of total vegetation cover. Includes deciduous forest, evergreen forest, and mixed forest. Pixels coded to a value of 4 have not changed between 1992 and 2001. |
| 5 | Grassland/Shrub - Includes grassland areas dominated by gramminoid or herbaceous vegetation and shrub/scrub areas dominated by shrubs less than 5 meters tall with shrub canopy typically greater than 20% of total vegetation, including true shrubs, young trees in an early successional stage, or trees stunted due to harsh environmental conditions. Management techniques that associate soil, water, and forage-vegetation resources are more suitable for rangeland management than are practices generally used in managing pastureland. Some rangelands have been or may be seeded to introduced or domesticated plant species. Includes those areas in the Eastern United States that commonly are called brushlands. Pixels coded to a value of 5 have not changed between 1992 and 2001 (Anderson, et al., 1976). |
| 6 | Agriculture - including cultivated crops and pasture/hay – Cultivated crops are described as areas used for the production of annual crops, such as corn, soybeans, vegetables, tobacco, and cotton, and also perennial woody crops such as orchards and vineyards. This class also includes all actively tilled land. Pasture/Hay is described as grasses, legumes, or grass-legume mixtures planted for livestock grazing or the production of seed or hay crops, typically on a perennial cycle. Pixels coded to a value of 6 have not changed between 1992 and 2001. |
| 7 | Wetlands - including woody wetlands and herbaceous wetlands – Areas where forest or shrubland vegetation accounts for greater than 20 percent of vegetative cover and the soil or substrate is periodically saturated with or covered with water. This class also includes areas where perennial herbaceous vegetation accounts for greater than 80 percent of vegetative cover and the soil or substrate is periodically saturated with or covered with water. Pixels coded to a value of 7 have not changed between 1992 and 2001. |
| 8 | Ice/Snow - All areas characterized by a perennial cover of ice and/or snow, generally greater than 25% of total cover. Pixels coded to a value of 8 have not changed between 1992 and 2001. |
| 12 | Open Water to Urban |
| 13 | Open Water to Barren |
| 14 | Open Water to Forest |
| 15 | Open Water to Grassland/Shrub |
| 16 | Open Water to Agriculture |
| 17 | Open Water to Wetlands |
| 18 | Open Water to Ice/Snow |
| 21 | Urban to Open Water |
| 23 | Urban to Barren |
| 24 | Urban to Forest |
| 25 | Urban to Grassland/Shrub |
| 26 | Urban to Agriculture |
| 27 | Urban to Wetlands |
| 28 | Urban to Ice/Snow |
| 31 | Barren to Open Water |
| 32 | Barren to Urban |
| 34 | Barren to Forest |
| 35 | Barren to Grassland/Shrub |
| 36 | Barren to Agriculture |
| 37 | Barren to Wetlands |
| 38 | Barren to Ice/Snow |
| 41 | Forest to Open Water |
| 42 | Forest to Urban |
| 43 | Forest to Barren |
| 45 | Forest to Grassland/Shrub |
| 46 | Forest to Agriculture |
| 47 | Forest to Wetlands |
| 48 | Forest to Ice/Snow |
| 51 | Grassland/Shrub to Open Water |
| 52 | Grassland/Shrub to Urban |
| 53 | Grassland/Shrub to Barren |
| 54 | Grassland/Shrub to Forest |
| 56 | Grassland/Shrub to Agriculture |
| 57 | Grassland/Shrub to Wetlands |
| 58 | Grassland/Shrub to Ice/Snow |
| 61 | Agriculture to Open Water |
| 62 | Agriculture to Urban |
| 63 | Agriculture to Barren |
| 64 | Agriculture to Forest |
| 65 | Agriculture to Grassland/Shrub |
| 67 | Agriculture to Wetlands |
| 68 | Agriculture to Ice/Snow |
| 71 | Wetlands to Open Water |
| 72 | Wetlands to Urban |
| 73 | Wetlands to Barren |
| 74 | Wetlands to Forest |
| 75 | Wetlands to Grassland/Shrub |
| 76 | Wetlands to Agriculture |
| 81 | Ice/snow to Open Water |
| 82 | Ice/snow to Urban |
| 83 | Ice/snow to Barren |
| 84 | Ice/snow to Forest |
| 85 | Ice/snow to Grassland/Shrub |
| 86 | Ice/snow to Agriculture |
| 87 | Ice/snow to Wetland |
How was the NLCD 1992/2001 Retrofit Change Product created?
(Excerpt from FGDC Metadata)
The change product database, stratified by NLCD 2001 mapping zones, is a system of processing modules that contain the modeling components, input variables, and output results from the change detection processing sequence. To create intermediate and final product thematic layers, decision tree classification modeling is used. The specific decision tree program employed is called See5, which implements a gain ratio criterion in tree development and pruning (Quinlan, 1993). See5 also implemented several advanced features that can aid and improve land cover classification, including boosting and cross-validation. Boosting is a technique for improving classification accuracy, while cross-validation can provide a certain level of estimation regarding the land cover classification quality. In addition, See5 can generate a confidence estimate for each classified pixel and record the associated classification logic in a text file.
In the first processing module, class differences between NLCD 1992 and NLCD 2001 are addressed by collapsing the original land cover products to an Anderson Level I (AL1) classification where the coarser classes help to absorb some of the disagreements present between the native NLCD products at Anderson Level II (AL2). Classification method differences are eliminated by generating new land cover maps for both 1992 and 2001 at Anderson Level 1, using a decision tree classification method, employing a combination of Landsat imagery, image derivatives, a digital elevation model, and topographic derivatives as independent variables. Training points for the new land cover maps are randomly collected and restricted to areas where both the 1992 and 2001 classifications agree at AL1. The new AL1 classifications and their respective confidence maps are used as input for the next module.
In the second processing module, the simplest form of change is identified on a per pixel basis where both of the new AL1 classifications do not agree. The pixel values correspond to a "from-to" classification code. For example, a pixel changing from Forest (4) to Agriculture (6) would have a 46 change pixel value. This intermediate change product is filtered with the confidence map from each new classification and only those pixels that were classified at a 70% confidence level in both the 1992 and 2001 eras are retained as "confident" change in a binary mask. The mask is further refined using an 8-neighbor clump and 5-pixel sieve process retaining only those contiguous areas of 5 pixels (~1 acre) or more, regardless of the type of change. The mask is then combined with the per pixel change image to create a more confident, readable form of change. The final output from this module is used for training data by random sampling in the next steps.
Processing sequences in the third module provide a refinement closely linked to spectral change. Three independent variables are extracted for input in the decision tree models. The first independent variable is a spatio-temporal image derived from intersected footprints of Landsat image date indices for both 1992 and 2001. A change/no-change continuous estimate is generated using a regression tree algorithm called Cubist (RuleQuest Research). The third independent variable is a three-layer special reflectance derivative created from differences in "Green" Normalized Vegetation Index (Gitelson, et al., 1996), Normalized Difference Water Index (Gao, 1996), and Specific Leaf Area Vegetation Index (Lymburner, et al., 2000) between the 1992 and 2001 MRLC datasets. Training data points used to create the change/no-change estimate are revised to include "from-to" class values and a refinement of the initial change image, based directly on spectral values, rather than classified maps is generated.
To maintain consistency with the 2001 NLCD land cover product, a postprocessing recode sequence is implemented to modify unchanged pixels and pixels with a "to" class equivalent to the NLCD 2001 cross-walk back to the original NLCD 2001 AL1 class code.
How do the land cover classes cross-walk from Anderson Level II to Anderson Level I?
The best information for Anderson Level II and Anderson Level I classes and how they relate to each other is contained in the original article:
Anderson, J., E. Hardy, J. Roach, and R. Witmer. 1976. A Land Use and Land Cover Classification System for Use with Remote Sensor Data, Geological Survey Professional Paper 964. 28p. (http://landcover.usgs.gov/pdf/anderson.pdf)
The table below summarizes how NLCD 1992 Land Cover and NLCD 2001 Land Cover were cross-walked to a modified Anderson Level I for the NLCD Retrofit Change Product.
NLCD 1992-2001 Anderson Level I Cross-walk Table.
This table illustrates Anderson Level I and II land cover classification codes and descriptions. Class descriptions and codes for NLCD 1992 and NLCD 2001 were derived from components of Anderson Level I and II (Anderson, et al., 1976) and National Oceanic and Atmospheric Administration (NOAA) Coastal Change and Analysis Program (C-CAP) land cover classes (NOAA, 2005). To facilitate comparison, classes for NLCD 1992 and NLCD 2001 are cross-walked to the modified Anderson Level I scheme.
Why is the NLCD 1992/2001 Retrofit Change Product at Anderson Level I?
The change product classes are described in a modified Anderson Level I scheme to eliminate the most incompatible legend differences between the two products at Anderson Level II. In an Anderson Level I cross-walk, some classes contain different land cover types. For example, in 1992, the urban classes include a class called Urban and Recreational Grasses as a subgroup in the Herbaceous/Planted-Cultivated land cover class. The most closely related 2001 equivalent is Developed Open Space that includes not only parks and golf courses but also subdivision areas with large open lots that contain up to 20% impervious surface area. Another example is the Barren class. In the 1992 NLCD product, a transitional Barren class includes mechanically disturbed and forest clear-cut areas that are not included in the 2001 Barren class definition.
