Cascade Bridge by Nature Cipher

We forecast the satellite signals of drought before satellites can see them through ML cascade bridges + ECMWF SEAS5 deliver 1–3 month drought warnings for Kenya's pastoral drylands.

Comment 1

Project Overview

  • The Constraint: Seasonal drought forecasts rely on CHIRPS rainfall and MODIS vegetation, which are only available for past months. Satellites cannot photograph the future.
  • The Cascade Bridge Solution:
    • Trains XGBoost regressors to synthesize satellite indicators from atmospheric variables.
    • Drives a drought classifier (built on 51 satellite-derived features) up to three months ahead using ECMWF SEAS5 seasonal forecasts.
    • Distills the satellite record during training, eliminating the need for a live satellite feed at forecast time.
  • Intended Users: Organizations acting before a food-security emergency:
    • ICPAC and national meteorological agencies (issuing regional outlooks).
    • County drought committees in Kenya's ASALs (pre-positioning water trucking and fodder).
    • Anticipatory-action funds (releasing money on a trigger).
    • Note: This adds a forecast layer to existing monitoring systems like ICPAC's Drought Watch or HUSIKA.
  • Coverage Area: 11 counties across 3 ASAL clusters:
    • Turkana, Marsabit, Samburu, Isiolo, Wajir, Mandera, Garissa, Kitui, Makueni, Machakos, and Tharaka-Nithi.
  • Transparency & Ethics:
    • Forecasts are explicitly labeled as experimental.
    • The dashboard publishes base rates, baselines, and confidence intervals alongside every metric.
    • Principle: An early warning system that overstates its own confidence is worse than none.

Solution Details

  • Core Mechanism: Overcomes the lack of future satellite data by chaining four XGBoost regressors: text SEAS5 (51 ensemble members, atmosphere only) Bridge 1: ERA5 vars -> CHIRPS rainfall proxy Bridge 2: + CHIRPS proxy -> MODIS NDVI proxy Bridge 3: + NDVI proxy -> LST day / night proxies -> 51 features -> XGBoost classifier -> P(drought)
  • Pipeline Processing:
    • Physical Ordering: Each bridge consumes the previous one's output (rainfall drives greenness → greenness moderates surface temperature).
    • Derived Indices: SPI (gamma-fitted), SPEI (log-logistic), VCI, TCI, and VHI.
    • Adjustments:
      • Bias correction maps the SEAS5 anomaly onto the ERA5 climatology.
      • PET uses the Hargreaves method.
      • Soil moisture relies on a disclosed persistence assumption (as SEAS5 does not provide it).
  • Decision Logic:
    • Default Threshold: Set at 0.25 (rather than 0.5), encoding the judgment that a missed drought costs roughly three times as much as a false alarm.
    • Customizability: A dashboard slider allows different stakeholders to read probabilities at their own risk tolerance. Changing the threshold changes the decision, never the forecast.
  • LLM Advisory Layer:
    • Uses Groq to transform the forecast into an advisory format.
    • Features grounded Q&A and generates county bulletins exportable to PDF.
    • Data Integrity: The model writes the prose, but all tabular numbers are rendered directly from the forecast JSON to ensure a mis-stated figure cannot reach the data.
  • Technology Stack:
    • Backend & Data: Python 3.11, XGBoost 2.0.3 (CPU), xarray/cfgrib, pandas, scipy.
    • Infrastructure: AWS S3/EC2, boto3.
    • Frontend: Static dashboard with MapLibre GL, Cloudflare Pages + Pages Functions.
    • Inference: Groq.
  • Baseline Comparison:
    • Publishes the ERA5-only baseline (the classifier without the bridges) to transparently show whether the bridges add predictive value over their own inputs.

Built With

Share this project:

Updates

Submission history