How SVRCLIP is constructed.

Overview

SVRCLIP contains radar-reflectivity sequences centered on tornado-report locations. The source is the Iowa State 0.01° composite-reflectivity archive, not the three-dimensional GridRad dataset. Each available two-dimensional composite_n0r field is regridded to the approximately 3.75-km WRF-BCC grid before report-centered products are extracted.

For each report, the report time is rounded to the nearest 15-minute radar time. The closest WRF-BCC grid cell to the report latitude and longitude becomes the fixed center of a 136 × 136 domain. Nine frames are extracted at 15-minute intervals from −60 through +60 minutes, including the valid-time frame.

Processing workflow

1. Discover available radar data. Source data are stored in hourly Zarr stores. An hourly store may contain one to four available 15-minute fields. Missing stores and missing internal times are not synthesized.
2. Regrid to WRF-BCC. Each available source field is mapped from the regular 0.01° latitude–longitude grid to the curvilinear WRF-BCC grid using xESMF nearest_s2d. Cells outside source-grid coverage are retained as NaN.
3. Select the report center. The nearest WRF-BCC grid cell to the report latitude and longitude is found using spherical distance. That cell is used for every frame in the sequence.
4. Select sequence times. Report time is rounded to the nearest quarter hour, and frames are requested from −60 through +60 minutes in 15-minute increments.
5. Extract a fixed domain. Each frame uses [center−68:center+68] along both horizontal dimensions. Near a grid boundary, locations outside the parent domain are padded with NaN.
6. Write numerical and display products. The continuous float32 sequence is preserved in NPZ format; rounded raw PNGs and mapped PNG/GIF products are produced separately.

Grid and geolocation

Raw arrays retain the WRF-BCC array orientation; they are not intentionally flipped for storage. Use the accompanying two-dimensional CLAT and CLONG arrays for geolocation rather than assuming a regular latitude–longitude orientation. Mapped PNG and GIF products use these coordinates directly.

Products for each report

Numerical representation

NPZ sequence

The compressed NPZ file is the recommended input for quantitative analysis and machine learning. Its reflectivity array has shape (9, 136, 136), dtype float32, and retains fractional dBZ values and NaN. It also contains exact timestamps, minute offsets, frame-availability flags, cropped CLAT/CLONG, report coordinates, selected grid-cell coordinates, and full-grid center indexes.

Raw PNG encoding

Raw PNG pixels are unsigned 8-bit values. Finite reflectivity is rounded to the nearest integer and clipped to 0–254; the value 255 is reserved for missing data. Do not interpret 255 as reflectivity. Because PNGs discard fractional precision, use NPZ files for scientific calculations.

Mapped PNG and GIF visualization

• cmweather HomeyerRainbow colors;
• discrete 5-dBZ bins from 0 through 80 dBZ;
• reflectivity below 5 dBZ displayed as transparent/white;
• state boundaries, national borders, and coastlines;
• longitude labels on the bottom and latitude labels on the left;
• a thin dashed 100-km geodesic circle around the selected center pixel; and
• a thin cross at that center pixel.

The mapped files are visualization products. They include axes, geographic features, annotations, and a colorbar and should not be used as numerical model inputs when the raw products are available.

Metadata schema

Catalog and annotations

interpolated_svrimg_manifest.csv is the cumulative product manifest. The website builds a local SQLite catalog from that manifest for fast filtering; the catalog is an index and is not the authoritative numerical radar dataset.

cd /var/www/nimbus/svrclip
sudo -u www-data php bin/build_catalog.php --reset

The annotation database is separate from the catalog. Each saved review contains one required primary storm mode, one required quality class, and any number of optional meteorological/context and subjective/reviewer tags. See Labels & ML for the current class and tag definitions.

Recommended machine-learning use

• Use the NPZ reflectivity array as the primary numerical input.
• Convert raw-PNG value 255 to missing data before normalization.
• Retain and inspect frame-availability flags.
• Keep human labels, consensus labels, and model-generated labels distinguishable.
• Use time-separated or event-separated data splits to limit leakage from closely related storms.
• Do not derive quantitative reflectivity from the mapped PNG or GIF products.

Known limitations

• Nearest-neighbor regridding is deterministic and preserves sharp structure better than smoothing interpolation, but it does not guarantee preservation of every small source-grid reflectivity maximum.
• The WRF-BCC domain extends beyond the source radar rectangle; out-of-coverage cells are NaN.
• Report time is rounded to the nearest 15-minute radar time, creating a possible temporal difference of up to 7.5 minutes.
• Centering on the nearest grid cell introduces a location difference of up to roughly half a grid cell.
• A report-centered crop does not establish that every echo in the image caused or is associated with the report.
• SPC report locations and times have their own observational and reporting uncertainties.

Citation

For the report-centered radar-image methodology and historical SVRIMG context, cite Haberlie, A. M., W. S. Ashley, and M. Karpinski (2020), “Mean storms: Composites of radar reflectivity images during two decades of severe thunderstorm events,” International Journal of Climatology. That paper does not document the newer Iowa State composite-reflectivity source, WRF-BCC regridding, or nine-frame SVRCLIP sequence workflow; those differences should be described explicitly in derived work.

Users should additionally acknowledge or cite the Iowa State radar archive and the severe-report source used for their downloaded subset, following the source providers’ current guidance.