DL-FRONT MERRA-2 vectorized weather fronts over North America, 1980-2018 (netCDF format)

DL-FRONT is a Deep Learning Neural Network (DLNN) that was trained to detect weather fronts using spatial grids of near-surface atmospheric variables. The dataset is composed of netCDF-4 files. Each file contains one year of hourly geospatial data grids describing the locations of four types of weather fronts—cold front, warm front, stationary front, and occluded front, over the time span 1980-2018. This dataset is the product of processing data from the National Aeronautics and Space Administration (NASA) Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2). DL-FRONT processed MERRA-2 hourly data grids of instantaneous measures of air pressure reduced to mean sea level, air temperature at 2 meters, specific humidity at 2 meters, and wind velocity at 10 meters over the time span 1980 - 2018 to produce this dataset. The original MERRA-2 data were resampled at 1 degree resolution over the spatial range 31W - 171W x 10N - 77N using bicubic interpolation. At each hourly time step the network produced a set of spatial grids with the same resolution and spatial range as the input, one for each of the five categories mentioned above. Each cell in a spatial grid for a given category records the network-assigned probability (from 0.0 to 1.0) that the cell is in a weather front boundary region of that category (or, for the "no front" category, the probability that the cell is not in any weather front boundary region). Each weather front probability map was then processed to obtain polyline skeletons of the weather front boundary regions found by DL-FRONT. These vector representations of the fronts were then written to JSON files—one file for each hour. These front polylines were then rasterized into geospatial data grids and stored by year into netCDF-4 files that conform to the Climate and Forecast Metadata Conventions. The front data in each file is stored in a netCDF variable with dimensions (time, front type, y, x), where x and y are geospatial dimensions. There is a 2D geospatial data grid for each time step for each of the 4 front types—cold, warm, stationary, and occluded. There are two large groupings of the netCDF files. One group uses a data grid based on the North American Regional Reanalysis (NARR) grid, which is a Lambert Conformal Conic projection coordinate reference system (CRS) centered over North America. The NARR grid is quite close the the spatial range of data displayed on the WPC workstations used to perform surface analysis and identify front locations. The native NARR grid has grid cells which are 32 km on each side. Our grid covers the same extents with cells that are 96 km on each side. The other group uses a 1° latitude/longitude data grid centered over North America with extents 171W – 31W / 10N – 77 N. The files in this group are identified by the name MERRA2, because they were used with data from the NASA MERRA-2 dataset, which uses a latitude/longitude data grid. There are a number of files within each group. The files all follow the naming convention merra2_[masked]_&lt;grid&gt;_&lt;n&gt;wide_&lt;year&gt;.nc, where [masked] indicates that the presence of the word <em>masked</em> is optional and &lt;grid&gt; is either <em>merra2-1deg</em> or <em>narr-96km</em>. The the sequence &lt;n&gt;wide indicates the width with which the fronts were drawn, and &lt;year&gt; is the year for the data stored in the file. The files marked as masked had a mask applied to the data grids that corresponded to the envelope of the geospatial region where there are, on average, 40 or more front crossing of any type per year, as determined using the Coded Surface Bulletin dataset. The &lt;n&gt;wide portion of the file names takes two forms—<em>1wide</em> and <em>3wide</em>. The fronts in the <em>1wide</em> files were rasterized by drawing the front polylines with a width of one grid cell. The fronts in the <em>3wide</em> files were rasterized by drawing the front polylines with a width of 3 grid cells. Within each grid group, there are four subsets of files: merra2_masked_&lt;grid&gt;_1wide_&lt;year&gt;.nc merra2_masked_&lt;grid&gt;_3wide_&lt;year&gt;.nc merra2_&lt;grid&gt;_1wide_&lt;year&gt;.nc merra2_&lt;grid&gt;_3wide_&lt;year&gt;.nc