Nexrad mosaic construction: an AI skill for analysis

NEXRAD Mosaic Construction

Overview

Build a custom multi-radar composite from aligned, validated single-site data. Every output pixel must have a traceable source, coverage state, and quality decision. Preserve the distinction between a local analysis product and an official provider mosaic.

Use this skill after products from multiple radars have been resolved. It does not discover, download, or decode source files. Use nexrad-product-access for single-site access and nexrad-mosaic-access when the required result is an existing NOAA MRMS composite rather than a newly constructed analysis.

When to Use This Skill

Combine two or more radars to cover a region with overlapping or adjacent

fields.

Construct a consistent base-reflectivity, velocity, spectrum-width, or

dual-polarization composite.

Build a research mosaic on a common grid or for a cross-radar comparison.
Compare a custom mosaic with an official MRMS product.
Visualize source contributions and coverage gaps in a composite.

Do not use this skill to fetch an official MRMS product, select a single radar, or claim that a simple pixel average is a physically resolved regional field.

Define the Mosaic Contract

Record:

product and exact units from each source;
source sites, coordinates, volume start/end times, and sweeps;
target region, output grid, projection, and cell size;
time-matching and temporal interpolation policy;
native range limits, beam width, beam height, and vertical sampling;
quality-mask, clutter, attenuation, range-folding, and dealiasing policy;
overlap-resolution and quality-weighting method;
treatment of coastlines, terrain, blocked sectors, and missing gates;
output format, metadata, and provenance requirements.

Do not combine a Level II moment and a Level III display product merely because their names resemble one another. Their processing, resolution, and quality semantics must be verified equivalent or the difference must be represented explicitly.

Validate and Align Source Data

Before compositing:

verify each site's decoded metadata, time, product, units, and quality;
convert all source fields to one declared projection and common output grid;
align source times using a documented tolerance and interpolation policy;
preserve each source's native spatial support and quality fields;
reject a source that is too stale, out of range, or outside its valid product

contract;

label any resampling or interpolation rather than presenting the output as

a native-resolution observation.

When a common time cannot be formed without excessive interpolation, use a nearest time within the stated tolerance, retain the actual time difference, or mark the output unavailable. Do not average distant scans into one pseudo-time without showing the temporal support.

Define the Output Quantity and Quality

For reflectivity, retain dBZ or the product's native calibrated unit and use a quality-aware combination rule. Do not average dBZ as a linear physical quantity without a stated rationale. For velocity, preserve positive and negative radial-velocity conventions and handle dealiasing and folding before compositing. For dual-polarization moments, preserve their physical units and quality masks.

Keep these fields separate:

measured or decoded value;
quality or suitability score;
source site and source time;
beam-height or range metadata;
interpolation or resampling state;
coverage and no-data state.

The output should make it possible to identify which radar contributed each pixel, especially near overlap boundaries.

Select Candidate Data by Quality and Geometry

For each output cell, consider only sources that provide valid coverage at a compatible time and product level. A useful selection score can include:

lower beam height at the target location;
shorter range and smaller beam width;
fewer terrain or clutter artifacts;
better data quality or signal-to-noise status;
newer scan or smaller time difference;
lower attenuation or range-folding penalty;
product-specific suitability such as dual-pol quality.

The score is a modeling choice, not a universal truth. State its terms and weights, normalize only comparable components, and test sensitivity to the selection rule. Never hide the score behind an unexplained “best radar” rule.

Resolve Overlaps Deterministically

Use a deterministic rule for every target cell:

select the source with the highest declared quality/suitability score;
apply a documented tie-breaker such as lower beam height, smaller range,

newer valid time, or stable site order;

preserve the selected source ID and runner-up source; or
use a physics-aware fusion method only after validating it against

single-site fields and reference observations.

Do not average dBZ, radial velocity, or quality fields across radars by default. Blending is acceptable only with a declared product-specific rationale and sensitivity analysis. Never let a blocked, stale, or invalid source win merely because its numeric value is larger, larger in magnitude, or easier to interpolate.

Preserve Coverage and Quality Metadata

The mosaic should emit at least:

value field;
source-site field;
source-time or time-difference field;
beam-height or range-quality field when relevant;
quality or suitability field;
coverage and no-data mask;
product, units, grid, projection, and timestamp metadata.

No-data and masked gates must remain distinguishable from a physical zero or a valid low value. A blank-looking region is acceptable only if the coverage mask makes its reason visible. Do not fill blocked sectors with neighboring radar data without labeling the substitution.

Handle 3D and Vertical Products

A two-dimensional mosaic must not silently mix elevation sweeps with incompatible heights. For reflectivity, state the selected elevation policy and range-dependent beam height. For VIL, echo top, or another 3D-derived product, use the provider's product definition or a separately validated 3D algorithm.

For a vertical composite, retain per-source sweep and beam-height metadata, define the vertical target or column, and document the common vertical coordinate. Do not call a lowest-sweep composite a column-integrated quantity or compare it with a true VIL product without explaining the difference.

Validate the Mosaic Scientifically

Inspect:

coverage, holes, overlap seams, and domain edges;
discontinuities at site boundaries;
implausible values or source-selection flips;
range, terrain, and beam-height artifacts;
agreement with each source in non-overlap regions;
comparison with an independent analysis or observation where available;
sensitivity to quality weights, time tolerance, and product choice.

An aesthetically smooth mosaic can hide source disagreements. Preserve the unresolved or selected-source map for scientific review. A local composite is an analysis with assumptions, not ground truth.

Coordinate with Analysis and Visualization

Use radar-satellite-analysis to interpret storm structure. A separate feature-tracking skill may be used after the mosaic when available, but it is not required for mosaic construction. Use nexrad-radar-visualization for site products or decoded gridded products. For a constructed mosaic, plot the value alongside source, quality, and coverage layers and label it as a local analysis.

For an official NOAA composite, retrieve it with nexrad-mosaic-access and preserve its product semantics instead of rebuilding it. A custom mosaic can be compared with that official product, but the two must not be labeled as identical.

Examples

To combine two overlapping Level II reflectivity volumes, verify both sites, sweeps, times, units, and quality; reproject them to a declared grid; then select a source per cell using a documented rule based on beam height, range, time difference, and quality. Preserve source and coverage arrays beside the reflectivity field, and inspect seams and each radar's non-overlap area before interpreting the result.

For an MRMS comparison, retrieve the exact official product through nexrad-mosaic-access, align its valid-time and grid semantics explicitly, and report differences as product/algorithm differences rather than treating either grid as ground truth.

Output Contract

Return a custom mosaic with a clear audit trail:

source radar sites, product or moment, units, volume times, and quality;
output domain, grid, projection, cell size, and time policy;
source-selection or fusion rule and its parameters;
value, source, quality, beam-height/range, and coverage layers;
rejected sources and cells with reasons;
validation results and sensitivity to major construction choices;
scientific limitations and distinction from official MRMS products; and
provenance and checksums for inputs, code, configuration, and outputs.

Do not publish only a colored mosaic. The source, quality, and coverage layers are part of the result whenever overlap or heterogeneous radar support matters.

Verification Checklist

Every source product, site, unit, and time is verified before compositing.
The common time, grid, projection, and range policy are explicit.
Level II and Level III products are not treated as interchangeable.
Quality, terrain, beam height, range folding, and de-aliasing checks are

applied.

Overlap selection is deterministic, auditable, and sensitivity-tested.
No-data, blocked sectors, and source substitutions remain visible.
Source, quality, and coverage metadata accompany the value field.
Custom output is labeled as a local analysis, not an official MRMS product.

Security & Safety Notes

Use public or authorized radar data only.
Do not expose cloud credentials, signed URLs, private bucket names, or

sensitive station metadata in outputs or logs.

Treat metadata, source IDs, and product names as untrusted input; validate

paths and labels before rendering.

Bound site count, volume size, output extent, interpolation, and worker count.
Preserve NOAA, NEXRAD, Unidata, and other provider attribution and license

terms.

Common Pitfalls

Seams appeared at radar boundaries: Source selection or quality weights

were not declared. Preserve source and quality layers and test tie-breaks.

dBZ was averaged across sites: A logarithmic reflectivity quantity was

treated as a linear scalar. Use a product-specific rule and sensitivity.

A blocked sector looked like no rain: No-data or quality-masked gates were

converted to zero. Preserve the coverage mask.

A lower sweep was compared with a 3D product: The vertical meaning was

collapsed. State the elevation and product definition.

The composite changed between runs: Site order, time tolerance, or random

sampling determined the winner. Make selection deterministic and record its parameters.

A custom mosaic was called official: The local product was confused with

NOAA MRMS. Keep source and algorithm identity in the output metadata.

Limitations

Heterogeneous range, beam width, terrain, calibration, and product processing

limit cross-site comparability.

A selection-based mosaic may create discontinuities or discard useful

information from a second radar.

No pixel is an independent observation; quality and source fields are

correlated in space and time.

Custom mosaics require local validation and are not an official warning or

truth product.