data trove noaa atmospheric weather alerts

Overview

Source: /home/coolhand/html/datavis/data_trove/data/quirky/atmospheric_real.json

Saturn profiled 571 rows across 19 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

event categorical label

This column contains National Weather Service alert/advisory event type names, serving as a categorical label for meteorological warning events. With 35 unique values across 571 rows, it covers a meaningful range of weather phenomena. 'Small Craft Advisory' dominates at 26.5% (149 occurrences), while the top 4 event types together account for the majority of records — suggesting a skewed distribution toward marine and winter/wind events. Entropy ratio of 0.72 indicates moderate-to-high diversity, but the heavy concentration in a few categories is worth noting for class-imbalance handling in any classification task.

Treatment: One-hot encode or target-encode for modelling; be aware of class imbalance with 'Small Craft Advisory' at 26.5% and many tail categories.

headline categorical free_text

This column contains NWS (National Weather Service) alert headlines — structured text strings describing weather advisory type, issuance timestamp, expiry, and issuing office. Despite appearing categorical, the entropy ratio of 0.943 and 305 unique values out of 571 records signal near free-text behaviour, with a long-tail alert confirming most headlines appear only once or a handful of times. The most frequent value ('Small Craft Advisory issued February 17…') appears only 19 times (top_rate ≈ 3.4%), indicating very little repetition across the dataset.

Treatment: Parse structured subfields (alert type, issue time, expiry time, NWS office) via regex before modelling; do not use raw string as a categorical feature.

description categorical free_text

This column contains full-text NWS (National Weather Service) alert and forecast descriptions — multi-line, structured prose covering marine, fire weather, wind, and winter storm warnings across various US regions. With 527 unique values out of 571 rows and an entropy ratio of 0.9945, nearly every entry is distinct, making this essentially free text. The long-tail alert and the presence of duplicate entries (e.g., the same Southeast Alaska marine forecast appearing 4 times) suggest periodic reissue of templated advisories rather than purely unique records, which may indicate time-series duplication worth investigating.

Treatment: Tokenize and embed (e.g., TF-IDF or sentence transformer) before modelling; consider deduplicating or grouping by alert template for frequency analysis.

severity categorical label

This column is an ordinal severity classification with 5 levels — Moderate, Minor, Severe, Unknown, and Extreme — likely describing the intensity of incidents, events, or conditions. 'Moderate' dominates at 36.9% of records (208/571), and 'Extreme' is strikingly rare at only 5 occurrences, suggesting a heavily right-skewed real-world distribution where catastrophic events are uncommon. The 19 'Unknown' values represent a data-quality concern distinct from the 1.4% null rate, effectively adding a second form of missingness. Entropy ratio of 0.77 indicates a reasonably spread distribution, though the extreme imbalance at the tail warrants attention for any classification task.

Treatment: Encode as ordinal (Minor < Moderate < Severe < Extreme); treat 'Unknown' as a separate missing indicator before modelling.

certainty categorical label

This column encodes an analyst-assigned confidence level for some observation or classification, with four ordinal categories: Likely, Observed, Unknown, and Possible. The distribution is severely skewed: 'Likely' dominates at 89.3% of all records (503/571), while 'Observed', 'Unknown', and 'Possible' each account for only 18–24 records. The low entropy ratio of 0.328 confirms near-constant behaviour, and the 1.4% null rate is minor. The near-total dominance of a single category limits this column's discriminative power as a feature.

Treatment: Ordinal-encode with awareness of severe class imbalance; consider collapsing minority categories or using as a stratification variable rather than a predictive feature.

urgency categorical feature

This column is a categorical urgency classification, likely from an incident or request management system, with 5 distinct levels. It is severely dominated by 'Expected' (510 out of 571 rows, 90.6%), leaving the remaining 4 categories — 'Unknown', 'Future', 'Past', and 'Immediate' — collectively accounting for fewer than 10% of records. The very low entropy ratio of 0.27 confirms extreme class imbalance, which will limit this column's discriminative power in most models. The 'Immediate' category, presumably the most critical, appears only 6 times, making it near-invisible to any classifier trained on this distribution.

Treatment: One-hot encode but flag severe class imbalance; consider oversampling minority classes (especially 'Immediate' with n=6) or collapsing into binary 'Expected vs. Other' before modelling.

areaDesc categorical label

This column contains geographic area descriptions used in weather or emergency alerts, predominantly covering coastal and inland waterways of Southeast Alaska (e.g., 'Glacier Bay', 'Stephens Passage', 'Northern Lynn Canal') with some continental US zones also present (e.g., 'Greater Lake Tahoe Area', 'Sacramento Mountains'). With 441 unique values across only 571 rows and an entropy ratio of 0.98, cardinality is extremely high — nearly every row is a distinct area. The long-tail alert confirms that most areas appear only once or twice, with even the top value ('Glacier Bay') appearing just 9 times (1.6% of rows). The multi-zone concatenated entries (e.g., 'Greater Lake Tahoe Area; Greater Lake Tahoe Area') suggest some records bundle multiple zones into a single string, which may cause deduplication or parsing issues.

Treatment: Parse semicolon-delimited multi-zone entries into separate records, then use as a grouping/filter dimension rather than a model feature; too high-cardinality for direct encoding without aggregation.

sent unknown other

The column 'sent' has 571 non-null rows but was skipped by the profiler, leaving its type and value distribution entirely unknown. No stats, uniqueness counts, or distribution metrics are available. The name suggests a boolean flag (e.g., message/email sent status) or a timestamp, but this cannot be confirmed from the evidence. An analyst should inspect raw values directly before any downstream use.

Treatment: Inspect raw values to determine dtype, then re-profile before any modelling or filtering.

effective unknown other

The column 'effective' contains 571 non-null values but was skipped by the profiler, leaving its type and distribution entirely unknown. No stats, uniqueness count, or value samples are available, so its semantic role cannot be determined from this evidence alone. The name suggests a boolean flag (e.g., is-effective) or a date (effective date/start), but this is speculation beyond the evidence.

Treatment: Manually inspect raw values to determine type, then re-profile before any modelling use.

onset unknown timestamp

The column 'onset' likely records the timing or start of a clinical or epidemiological event (e.g., symptom onset date), but the profiler emitted a 'skipped' alert and returned no stats, leaving its true type and distribution entirely uncharacterised. With 571 non-null rows and zero null rate, data is present, but nothing about format, uniqueness, or value range can be confirmed from this evidence alone. An analyst should inspect raw values to determine whether it is a date string, numeric duration, or categorical stage before any downstream use.

Treatment: Inspect raw values to confirm type (date vs. numeric vs. categorical), then parse or encode accordingly before modelling.

expires unknown timestamp

The column 'expires' likely represents an expiration date or timestamp field, but the profiler skipped analysis entirely, yielding no stats, no uniqueness count, and a kind of 'unknown'. With 571 non-null rows and zero null rate, data is present but its structure or encoding prevented saturn from classifying it. No further distributional signals are available from the evidence.

Treatment: Inspect raw values to confirm encoding (ISO string, epoch int, or other format), parse to datetime, then use as a feature or filter boundary.

latitude numeric

longitude numeric

event_type categorical

date unknown timestamp

This column is named 'date' and contains 571 non-null values with a 0.0% null rate, suggesting it is a timestamp or date field. However, saturn skipped profiling it (kind: 'unknown', no stats, no uniqueness count), so no distribution, range, or format details are available. The absence of any computed statistics prevents assessment of cardinality, temporal range, or potential drift. Treat with caution until the parsing issue causing the skip is resolved.

Treatment: Investigate why saturn skipped this column, parse to a proper datetime type, then extract temporal features or use as an index.

state categorical

country categorical

magnitude categorical

source categorical