An open, radar-native convolutional detector of tornadic rotation signatures

00 · Abstract

We present TIMA Radar, an open-source, real-time system that scores gridded weather-radar fields for the spatial signature of tornadic rotation. The detector is a compact convolutional neural network (≈242k parameters) that consumes three co-registered Multi-Radar Multi-Sensor (MRMS) products — composite reflectivity and azimuthal (rotational) shear in the 0–2 km and 3–6 km layers — sampled as a 64×64 grid box centered on a query point.

Positive examples are drawn from significant (EF2+) tornado reports in the NCEI Storm Events database, backed by a training corpus of over 16,000 indexed tornado events and 1.64 TB of raw NEXRAD radar data. Negatives include both clear-air scenes and a deliberately mined set of strongly rotating but non-tornadic storms. The production model attains an area under the ROC curve of 0.944 and a precision–recall AUC of 0.803 on held-out data, with high positive recall (0.901) and complete rejection of clear-air alarms.

01 · Input Sampling Geometry

Each sample is a 64×64 grid box spanning ±0.25° of latitude and longitude (roughly 5 km grid spacing at mid-latitudes), centered on an arbitrary query coordinate. Three MRMS products are stacked directly as separate input tensor channels:

• MergedReflectivityQCComposite — quality-controlled composite reflectivity mapping core storm structure and precipitation intensity.
• MergedAzShear_0-2kmAGL — low-level azimuthal shear tracking near-ground velocity rotation.
• MergedAzShear_3-6kmAGL — mid-level azimuthal shear revealing mesocyclone depth.

Missing or below-threshold values (sentinel values less than −900) are explicitly zero-filled during tensor preprocessing. This forces the network to evaluate absent rotation as a zero-strength signal rather than as an anomalous input.

02 · Neural Architecture

The network ("SmallCNN") is intentionally compact — roughly 242,000 parameters. It consists of stacked 2D convolutional blocks with batch normalization and spatial pooling, terminating in a concise fully connected head that emits a single logit.

This tight parameter footprint is matched to the information-dense but low-resolution nature of MRMS products. It regularizes heavily against spatial overfitting while enabling sub-100 ms real-time evaluation across dense grid domains on CPU-class hardware — a deliberate design constraint, so the model can be deployed on essentially any machine, including edge devices with no cloud connectivity. The output logit passes through a standard sigmoid, yielding a signature score bounded in [0, 1].

03 · Leak-Safe Training Protocol

Initial development configurations yielded an ROC-AUC of 0.9947 — a figure we diagnosed as a catastrophic artifact of temporal data leakage. Grid boxes from the same convective storm day were shared across training and validation partitions, letting the network memorize localized context rather than learn general kinematic structure.

To enforce scientific honesty, data partitioning was refactored into a strict leak-safe split: all samples from a given storm day live in exactly one partition, never both.

# Conceptual strategy: leak-safe storm-day segregation
def partition_storm_manifest(manifest_df):
    # All samples from one storm day stay in a single partition
    storm_days = manifest_df['storm_day_epoch'].unique()
    train_days, val_days = split_by_date_blocks(storm_days, ratio=0.8)

    train = manifest_df[manifest_df['storm_day_epoch'].isin(train_days)]
    val   = manifest_df[manifest_df['storm_day_epoch'].isin(val_days)]
    return train, val

Under this segregated baseline, performance settled into an operationally honest profile, which is maintained across all deployment verification passes.

04 · Evaluation & Results

Validation uses a location-blind grid-scanning protocol that denies the classifier any advance knowledge of true event coordinates. On completely held-out, leak-safe validation data, the clear-air-augmented production system achieves:

Because severe convective datasets suffer acute class imbalance, precision–recall AUC is our headline summary metric; ROC curves can flatter a binary model under heavy structural imbalance, and we report both for transparency.

05 · Limitations & False-Alarm Analysis

Aggregate metrics obscure localized failure modes. Decomposing the empirical false-alarm rate (FAR) across distinct storm configurations shows exactly where model errors reside:

The remaining hard-negative FAR of ≈0.27 represents an intrinsic constraint of single-time-step radar data: an intense cyclonic velocity couplet aloft can appear identical on raw radar structure whether or not it is producing a surface tornado. Discriminating these cases requires near-storm thermodynamic context and temporal sequence information — the direct targets of the roadmap below.

06 · Pipeline Roadmap

To push past the single-snapshot radar ceiling, three architectural enhancements are planned, each attacking the hard-negative frontier from a different direction:

• Phase 01 — Dual-Polarization Debris Fusion: integrate NEXRAD Level II correlation-coefficient (CC) channels. Co-locating localized CC drops with tightening velocity couplets detects active tornadic debris signatures (TDS), directly confirming a tornado in progress and neutralizing non-tornadic false alarms.
• Phase 02 — Near-Storm Environmental Late Fusion: train a secondary network branch conditioning radar scores on HRRR atmospheric soundings. Ingesting CAPE, storm-relative helicity (SRH), and LCL-height fields lets the model down-weight intense rotation occurring in thermodynamic environments hostile to tornadogenesis.
• Phase 03 — Temporal Recurrent Modeling: transition from single-volume static inputs to a short-sequence convolutional-recurrent model tracking recent radar history — capturing the tightening trends that precede tornadogenesis and converting detection into lead time.