Skip to content

Environmental Risk Signals

Phase 3 adds a first-pass Shark Encounter Risk score. This is encounter-risk estimation, not attack prediction.

The score summarizes environmental conditions that may increase the chance of overlap among sharks, prey, and people. It is not a safety guarantee, beach closure recommendation, or substitute for local lifeguard, weather, wildlife, or emergency guidance.

AI1SAD intentionally starts with deterministic, explainable rules. The order is:

  1. Deterministic and explainable warning model.
  2. Bayesian/statistical layer after source freshness and validation are stable.
  3. ML ensemble only after the public data contracts, labels, uncertainty, and review process are mature.

Risk Bands

Score Band
0 to 24.99 low
25 to 49.99 moderate
50 to 74.99 elevated
75 to 100 high

Score Split

AI1SAD separates three scores:

  • warning_score: environmental/live-condition risk from weather, ocean, biological, vessel, and exposure signals.
  • surveillance_priority_score: where safety or drone teams should look first.
  • activity_hazard_score: risk introduced by what the human is doing in context.

Do not call any of these attack probability. The activity hazard layer can be high while the environmental warning score remains low if live environmental signals are missing or quiet.

Examples of activity context:

  • Spearfishing
  • Diving with catch
  • Fishing near reef/dropoff
  • Swimming near bait/prey activity

For Western Australia, spearfishing on reef/dropoff habitat with known or suspected white shark suitability strongly raises surveillance priority. It does not automatically raise the environmental warning_score unless live environmental signals also support that.

Rule-Based Factors

Factor Max Points Current Rule
Historical incident density 20 Scales from 0 to 20 based on nearby public incident count, capped at 20 incidents.
Recent rainfall/runoff 15 10+ mm adds 5, 25+ mm adds 10, 50+ mm adds 15.
River mouth distance 15 <=1 km adds 15, <=5 km adds 10, <=10 km adds 5.
Sea surface temperature / seasonal suitability 10 20-28 C adds 10; shoulder temperatures add 5; seasonal months can add a small boost.
Fishing activity 10 User/provider signal from 0 to 1 scaled to 10 points.
Baitfish / prey indicator 10 User/provider signal from 0 to 1 scaled to 10 points.
Water visibility 10 <1 m adds 10, <3 m adds 6, <5 m adds 3.
Human water activity 10 User/provider signal from 0 to 1 scaled to 10 points.

Regional Profiles

Phase 3A adds regional_risk_profiles. The API detects the nearest public regional profile and applies region-specific calendars and multipliers to produce:

  • score: base environmental rule score
  • band: base score band
  • warning_score: regional-profile-adjusted warning score
  • warning_band: adjusted warning band
  • confidence: first-pass confidence value based on whether a regional profile matched
  • dominant_contributing_factors: largest factor contributions after regional adjustment

Current-condition warning responses also return dominant_factors, where every factor includes a numeric contribution value. Missing or stale provider data is listed in missing_data_sources and reduces confidence; it is never silently treated as normal.

Seeded profiles:

  • Florida
  • Hawaii
  • New South Wales, Australia
  • Western Australia
  • California
  • South Africa
  • Red Sea

The model does not use one global summer definition. Local summer/high-exposure months are stored per profile.

Special handling:

  • Hawaii: October applies a Sharktober multiplier.
  • New South Wales and Western Australia: January-February apply Southern Hemisphere high-exposure/high-attention handling.
  • Florida: weekend exposure and non-summer tourist/beach exposure can increase the human-exposure warning score.

Assumptions

  • Historical incident density is a coarse proxy for overlap among sharks, humans, habitat, and reporting.
  • Rainfall is treated as a runoff and water-clarity proxy until direct runoff/turbidity feeds are added.
  • River mouths, inlets, estuaries, and outflows may concentrate nutrients, prey, turbidity, and predator movement.
  • Sea surface temperature is only a coarse seasonal suitability signal.
  • Fishing and baitfish indicators are treated as local attractant/prey signals.
  • Human water activity changes encounter opportunity, not shark behavior by itself.
  • Regional profiles are coarse public-warning profiles; they should be refined with local expert review and official regional data.

Limitations

  • This model does not predict attacks.
  • It does not know whether beaches are open, guarded, or under local warning.
  • It does not include real-time shark tracking, drone sightings, lifeguard reports, or official advisories yet.
  • It depends on the quality of supplied environmental signals.
  • Historical incident data has reporting bias and does not measure exposure-adjusted risk.

Future Data Sources

Potential future inputs:

  • NOAA sea surface temperature and coastal observations
  • Rainfall and river discharge feeds
  • Turbidity, water clarity, and sediment plume products
  • Tide, moon, swell, and wind data
  • Fishing pier, boat ramp, and fishing-pressure proxies
  • Baitfish reports, bird-feeding observations, and ecological surveys
  • Lifeguard advisories and beach closure feeds
  • Local official shark sighting reports

External provider API keys and config must stay in .env or deployment secrets only.