settingsMethodology

How a forecast is produced

From elder-elicited rules to a calibrated ensemble.

timeline

Pipeline

  1. 1

    Elicitation

    Wind, moon, cloud and felt-temperature cues gathered from fishing elders.

  2. 2

    Weather ingest

    Open-Meteo current + daily for the chosen lat/lon.

  3. 3

    Fuzzy inference

    87 Mamdani rules → Good / Normal / Risky / Bad.

  4. 4

    ML ensemble

    Random Forest + fuzzy posterior, calibrated cross-fit.

  5. 5

    Threshold decision

    Per-class thresholds via tune_thresholds.

  6. 6

    Logged

    Inputs, probabilities and model version stored for review.

spa

Indigenous

Inputs

  • airWind names (Genya, Kus, Nyabukoba, Tarai…)
  • dark_modeMoon phase and position
  • cloudCloud cover and formation
  • thermostatFelt body temperature

Strengths

  • check_circleLocalised, place-aware reasoning
  • check_circleGenerations of observational data
  • check_circleRuns offline
satellite_alt

Scientific

Sources

  • cloudOpen-Meteo current + daily
  • device_thermostatDaily min/max temperature
  • rainyPrecipitation sums
  • airMaximum wind speed

Strengths

  • check_circleQuantitative, comparable
  • check_circleBroad-area coverage
  • check_circleCalibrated probabilities
shuffle

Fusion

Data integration

Ensemble + calibration

RF probabilities blend with the fuzzy posterior, then the joint distribution is calibrated cross-fit on a held-out fold.

  • balanceSample-weighted training
  • scheduleTemporal 80/20 split
  • tunePer-class thresholds on macro-F1
  • databaseEvery forecast logged
lightbulb

Why hybrid wins

Indigenous rules carry pattern recognition satellites miss; calibrated ML covers cases a single rule fires too late on.