Flood-Threat Situation Note — Colombia

 

Fast Mode Run | Reliability-Calibrated Multi-Prior Log-Odds Fusion | Posterior Surface + K-Means Operational Classes

---

Executive Summary

This situation note presents the results of a fast-mode flood-threat assessment for Colombia using a calibrated, multi-prior log-odds fusion framework. The model integrates short-term meteorological forcing with historically informed spatial predisposition derived from event density and embedding similarity.

The resulting posterior probability surface identifies a concentrated zone of elevated flood-impact likelihood in north-western Colombia, particularly along the Antioquia–Chocó–Córdoba corridor and the Andean–Pacific transition. When converted into operational risk classes using K-means clustering, this signal resolves into a compact high-risk core surrounded by broader mid-level zones of concern.

The methodology improves upon earlier embedding-only Bayesian-style fusion approaches by introducing reliability calibration and prior diversification, thereby reducing overconfidence and improving interpretability for operational use.

---

1. Methodological Framework

This model follows a structured analytical design consistent with the JANUS framework: dual-perspective reasoning, explicit evidentiary integration, calibration against observed data, and transparent interpretation of uncertainty.

1.1 Near-Term Hazard Signal (Likelihood Proxy)

The hazard component represents short-term hydrometeorological forcing. It is constructed from:

• GFS 0.25° precipitation accumulation over a forecast window,

• Adjusted by a CHIRPS climatological baseline for equivalent calendar days,

• Modulated by landscape susceptibility proxies:

  • SRTM-derived slope,

  • JRC Global Surface Water dynamics,

  • Nighttime lights (impervious/built environment proxy),

  • Recent forest loss,

  • Recent SMAP soil moisture.

This produces a hazard_index, which reflects not simply rainfall magnitude, but rainfall relative to seasonal expectation and contextualized by land-surface vulnerability.

The hazard surface functions as a likelihood proxy. It does not represent a physically explicit flood simulation; rather, it captures conditions under which impactful flooding becomes statistically more plausible.

---

1.2 Calibrated Multi-Prior Structure

Instead of relying on a single vulnerability layer, the model integrates multiple priors:

1. Flood Event Density Prior
   A smoothed kernel density surface derived from historical flood event locations.

2. Embedding Signature Priors

   • A flood-specific embedding signature (mean vector at historical flood pixels),

   • A broader all-hazard embedding signature.

For each pixel, cosine similarity to these signatures is computed. Higher similarity implies resemblance to historically impacted locations.

These priors are not averaged in probability space. They are combined in log-odds space, allowing each to contribute additive evidence.

Reliability Calibration

Similarity is not inherently probabilistic. Therefore, a calibration table is constructed using sampled positive and negative points. Raw prior outputs are binned and mapped to empirical event frequencies, producing a reliability-corrected prior probability.

This step reduces the common overconfidence bias of embedding-based similarity models and aligns predicted predisposition with observed frequencies.

---

1.3 Log-Odds Fusion (Posterior Construction)

The final posterior probability is constructed as:

$$\text{logit}(p_{\text{post}}) = \text{logit}(p_{\text{prior,cal}}) + w \cdot \text{logit}(p_{\text{like}})$$

Where:

• $p_{\text{prior,cal}}$ is the calibrated multi-prior,

• $p_{\text{like}}$ is the hazard-derived likelihood proxy,

• $w$ is a tunable weight controlling forecast influence.

This formulation ensures interpretability: the posterior increases when both predisposition and forecast forcing align. The weight parameter explicitly governs the balance between historical predisposition and near-term atmospheric forcing.

This is Bayesian in structure, though the likelihood remains a proxy rather than a fully specified generative distribution. For clarity and publication integrity, this framework is best described as:

Reliability-Calibrated Multi-Prior Log-Odds Fusion (CMPLF).

---

2. Results — Colombia (Fast Mode)

2.1 Posterior Probability Surface

The posterior surface reveals a dominant hotspot in north-western Colombia, spanning:

• Antioquia,

• Chocó,

• Córdoba,

• Extending toward the Andean–Pacific transition zone.

Elsewhere, signals are present but comparatively weak and spatially fragmented. The posterior is generally low across large portions of the country, indicating that short-term forcing and historical predisposition are not broadly aligned outside the NW corridor.

Interpretation: if flood impacts materialize within the forecast window, the highest relative probability of impactful conditions is concentrated in this NW region.

---

2.2 K-Means Operational Risk Classes

To support decision-making, the continuous posterior is classified using K-means clustering.

Cluster centers (sorted):

• 0.0009

• 0.0350

• 0.1050

• 0.2143

• 0.6342

Threshold midpoints:

• 0.0180

• 0.0700

• 0.1597

• 0.4242

The distribution is strongly right-skewed. Most pixels fall below 0.1, while a small subset reaches high values (>0.4).

Operational interpretation:

• Classes 0–1 (< ~0.018): Background conditions.

• Class 2 (~0.07–0.16): Watch zones; moderate convergence of predisposition or forcing.

• Class 3 (~0.16–0.42): Elevated concern; alignment between historical and forecast signals.

• Class 4 (> ~0.42): Strongest convergence; compact, high-confidence risk core.

The highest class forms a spatially coherent cluster in NW Colombia. Mid-level classes form a halo around this core and appear in scattered patches elsewhere.

---

3. Methodological Significance

Relative to earlier embedding-only Bayesian-style fusion:

• Prior diversification reduces brittleness and overfitting to a single spatial “signature.”

• Reliability calibration transforms similarity into empirically anchored probability.

• Log-odds fusion preserves additive evidentiary structure and interpretability.

• Weight control becomes a meaningful structural parameter rather than a cosmetic adjustment.

If calibration samples remain representative and stable across runs, this approach is more operationally robust than uncalibrated embedding priors.

---

4. Operational Considerations

1. Right-Skewed Posterior
   The strong skew implies that the highest class represents genuine statistical outliers. This supports targeted intervention but requires careful threshold consistency for cross-country comparison.

2. Sensitivity to Scaling
   Changes in hazard normalization or calibration density can shift mid-class boundaries. For comparability across runs, percentile scaling parameters and calibration methodology must remain fixed.

3. Interpretation Boundaries
   The posterior represents relative probability under model structure, not confirmed inundation. It indicates where forecast forcing and historical predisposition converge.

---

5. Situation Assessment (This Run)

Headline:
North-western Colombia exhibits the strongest modeled convergence between historical flood predisposition and short-term meteorological forcing in the current forecast window.

Primary Area of Concern:
The Antioquia–Chocó–Córdoba corridor and adjacent Andean–Pacific transition zone.

Secondary Signals:
Fragmented mid-level zones in parts of the central Andes and eastern foothills, lacking coherent regional clustering.

Strategic Implication:
Monitoring and preparedness resources should prioritize the NW corridor, where calibrated prior evidence and forecast forcing are jointly maximized.

---

This assessment reflects a reliability-calibrated, multi-prior log-odds fusion framework designed for operational early warning support. The model explicitly integrates historical evidence, spatial embedding similarity, and short-term meteorological forcing, while preserving interpretability and calibration transparency.