Process_Description:
NLCD 2016 Land Cover Process Step
National Land Cover Database (NLCD) land cover data is fundamentally based on the analysis of Landsat data. Approximately 3,608 Landsat scenes are required for NLCD 2016 (this total does not include additional scenes needed for cloud/shadow fill). Using automated scripts, Landsat scenes were selected for seven target years: 2001, 2003, 2006, 2008, 2011, 2013, and 2016. All scenes with cloud cover less than 20 percent were downloaded and one cloud-free, leaf-on image was selected for each path/row in each year. In addition, one leaf-off image was added for 2016 only.
If no cloud-free images were available for the target year, then an alternate cloud-free image was selected either one year before or one year after the target. If there was still significant cloud cover, then up to three fill images were selected and applied in a process to remove clouds, cloud shadow, smoke, or other artifacts.
Other input datasets include: NLCD 2001, 2006, and 2011; 3D Elevation Program (3DEP) digital elevation data; Coastal Change Analysis Program (C-CAP) land cover; Cropland Data Layer (CDL); National Wetlands Inventory (NWI); Soil Survey Geographic (SSURGO) Database; and State Soil Geographic (STATSGO2) Database. SSURGO (with STATSGO2 to fill in gaps) was the basis for a hydric soils data layer used in training data assembly.
NLCD 2016 is produced by modeling land cover change over seven intervals between 2001 and 2016, with consistent change trajectories built into the process. The first set of models in this process are for multi-spectral change detection. The Multi-Index Integrated Change Analysis (MIICA) model outputs a change map between two dates of imagery. Five spectral indices are also calculated, and a disturbance map is produced by the Vegetation Change Tracker (VCT) software. The MIICA outputs, the five spectral indices, and the 1986 to 2016 disturbance map are the inputs to the training dataset assembly stage.
A set of models was developed to assemble a training dataset for each land cover class for each of the seven target years. The training dataset models were built with Landsat images and derived indices, spectral change products, trajectory analysis, and ancillary data: NLCD 2001, 2006, and 2011; C-CAP land cover; CDL; NWI; a cultivated cropland 2008 to 2016 dataset; and a hydric soils dataset. Image segmentation was performed on the Landsat scenes, and the resulting image objects were used to mitigate noise in the training data. The final output of this stage is training data for each of the seven target years, for input into the initial land cover classification stage.
For each of the seven target years of Landsat data, two percent of all available training data per path/row was drawn from the data as training samples, and one percent was drawn as validation samples. The C5 decision tree classification software was run on the training samples to generate a set of rules, and the decision rules were applied to generate a land cover classification for each of the seven target years.
The C5 software was run with four sets of independent variables: the 1986 to 2016 disturbance year data derived from VCT; the set of Landsat images; compactness indices from image segmentation; and a DEM and its derivatives. The classifier was run twice, once with all land cover classes processed and the 1986 to 2016 disturbance year data included, and again with two classes - Urban and Water - omitted from the classification and the disturbance year data not included in processing.
The two classifications were integrated with ancillary data and the segmentation polygons to produce seven initial land cover maps.
A post-classification refinement process was developed to correct classification errors in each target year, check for consistency of land cover labels over time, and improve spatial coherence of land cover distribution. Refinement was conducted class-by-class in hierarchical order: (1) Water, (2) Wetlands, (3) Forest and forest transition, (4) Permanent snow, (5) Agricultural lands, and (6) Persistent shrubland and herbaceous. Models were developed for refinement of each class and each type of confusion. For example, confusion between coniferous forest and water, both spectrally "dark" could be corrected by reclassifying water to coniferous forest where slope was greater than 2 percent. Confusion between forest and cropland could be mitigated with CDL data, and so forth.
The final integration step resolved class label issues pertinent to local environments (such as coastal areas), and, for land cover classes other than Water and Developed, ensured that all pixels in a segmentation object were in the same class. Pixel-based and object-based land cover labels were checked for differences, which were reconciled by a rule-based model. Water and Developed classes kept pixel values intact even in areas that were smaller than segmentation objects. Change trajectories for each class were checked for consistency through the seven target years.