Chesapeake Land Cover dataset - Learning on the prior extension

This dataset extends the "Chesapeake Land Cover" dataset at https://lila.science/datasets/chesapeakelandcover with an additional layer of data containing a prior of observing a 1m Chesapeake Conservancy land cover class label given an NLCD label. Specifically, for each tile in the original "Chesapeake Land Cover" dataset, this dataset contains another tile (named with a "_prior_from_cooccurrences_101_31_no_osm_no_buildings.tif" suffix) containing the prior probabilities of observing a four class version of the land cover classes at a 1m resolution. The prior probability is given by the normalized co-occurrence matrix between NLCD classes and the Chesapeake 1m land cover labels in each state with additional spatial smoothing (using a gaussian filter with a standard deviation of 31 pixels and a cutoff of 101 pixels) to reduce the block artifacts caused by the relatively low-resolution of the NLCD labels (30m) compared to the LC labels (1m). Note: the prior is a 4-class mapping of the 6-class labels. In the 4-class version of the dataset the "barren land", "impervious (other)", and "impervious (road)" classes are combined.

This dataset also includes the per-state co-occurrence matrices. Each of these is a matrix, \(C\), with size 7 x 17, where an entry \(C_{ij}\) gives the normalized count of land cover class \(i\) for given NLCD label \(j\). The class indices \(i\) and \(j\) correspond to the indices used in the Chesapeake Conservancy land cover label and NLCD layers.

The dataset is packaged in a way such that it can simply be unzipped over an existing copy of the "Chesapeake Land Cover" (i.e. follows the same directory structure). E.g., given a directory that contains the original dataset "cvpr_chesapeake_landcover/", and the file "cvpr_chesapeake_landcover_prior_extension.zip", run `unzip cvpr_chesapeake_landcover_prior_extension.zip` to create the merged dataset.