TH scores update: OVERVIEW

[matamadio]

continues after #948

TH Score Logic and Workflow Summary

• Global data update #948
• TH scores update: FLOODS #950
• TH scores update: TROPICAL CYCLONES #954
• TH scores update: LANDSLIDE #955
• TH scores update: EARTHQUAKE #953
• TH score update: EXTREME HEAT #958
• TH score update: VOLCANO #957
• TH scores update: WILDFIRE #956
• TH scores update: TSUNAMI #960
• TH scores update: WATER SCARCITY #961

Output folder

https://worldbankgroup-my.sharepoint.com/:f:/g/personal/mamadio_worldbank_org1/EuEuuNHaPFhNpkU9ztUL5TwBNCz1q4Tq-vot47rtklfNyA?e=JoWMgI

Download for WB: https://worldbankgroup-my.sharepoint.com/:u:/g/personal/mamadio_worldbank_org1/IQD39K9_tjlBSaEhTr3oMc5vAXsun78keK_0ZoPONi-dJuE?e=KTMLuh

Value Threshold

• Definition: Minimum hazard intensity to count pixels as "affected"
• Application: Used by ALL hazards

Area Threshold %

• Definition: Minimum percentage of admin unit that must be affected to trigger scoring
• Application: Used by ALL hazards
• Range: 0-100%

Scoring Philosophy

Consistency Across Hazards

All hazards produce scores on a -1 to 3 scale:

• Score -1: Not affected (geographically outside hazard zone) - Cyclone, Volcano, WIldfire only
• Score 0: No significant hazard (within hazard zone but below thresholds)
• Score 1: Low hazard
• Score 2: Medium hazard
• Score 3: High hazard

Note on Score -1 vs Score 0:

• Score -1: The hazard is geographically not applicable (e.g., landlocked countries for cyclones, areas far from volcanoes, non-vegetated areas for wildfire)
• Score 0: The hazard could occur but intensity/exposure is below thresholds (e.g., low VEI within range, minimal cyclone winds)

Method implementation

Two analytical notebooks have been built to produce updated scores based on recent dataset and procedure as defined below.

These notebook could be later released as open-source TH utility for anyone that wants to replicate TH approach on custom boundaries, different years, diffetent RPs or using different thresholds - which some users have previously mentioned.

Output Files

GeoPackage (.gpkg)

• Contains all spatial data and attributes
• Full geometry and statistics preserved

Excel (.xlsx)

• Tabular data without geometry
• Easy analysis and reporting

Naming Convention

{COUNTRY}_{HAZARD}_hazard_{UNIT_LEVEL}_vt{VALUE_THRESHOLD}_at{AREA_THRESHOLD}_{TIMESTAMP}

Examples:

• GLB_Earthquake_hazard_ADM2_vt80_at1.0_20250127_143022.gpkg
• THA_Cyclone_hazard_ADM2_vt36_at1.0_20250127_143022.xlsx
• GLB_Extreme Heat_hazard_ADM2_vt25-28-32_at1.0_20250127_143022.gpkg

Special Cases:

• Extreme Heat: vt{RP100}-{RP20}-{RP5} (e.g., vt25-28-32)
• Multiple countries: Country codes joined with underscore
• Global: GLB prefix

[matamadio]

Except for Drought (WIP), all scores have been merged in one file:

• TH_scores_ADM2_2025.gpkg >> New boundaries (2025), new datasets, new calculation approach

Original TH scores are in:

• TH_scores_ADM2_original.gpkg

[matamadio]

With the same rules, scores have been produced for the selection of 2945 urban areas.

TH_scores_urban.gpkg

[matamadio]

ThinkHazard Scores Comparison Report

Original (2019) vs. 2025 Update

Data Overview

• Original Dataset: 38,229 administrative regions (ADM2 level)
• 2025 Dataset: 39,370 administrative regions (ADM2 level)
• High-Quality Matches: 33,207 regions (90%+ matching confidence)
• Unmatched Original: 5,022 regions
• Unmatched 2025: 6,268 regions

Technical Notes

• Hazard ranks: -1 = N/A, 0 = Very Low, 1 = Low, 2 = Medium, 3 = High
• Original data used text codes (VLO, LOW, MED, HIG), converted to numeric for easier comparison (delta)
• Matching threshold: 90% similarity required for inclusion in analysis
• Medium-quality matches (70-90%) available in separate sheet for manual review

Matching Method

Fuzzy string matching using hierarchical location names (Country | Admin1 | Admin2).
Only high-quality matches (≥90% similarity) were used for hazard comparison.

---

Key Findings

1. Landslide (LS)

• Regions Compared: 21,919
• Agreement: 47.7% (10,459 unchanged)
• Net Trend: Downgrades (9,141 downgrades vs 2,319 upgrades)
• Average Change: -0.73 ranks
• Notable Shifts:
  • High → Medium: 2,831 regions
  • Low → Very Low: 2,125 regions
  • Many high-risk regions were downgraded

2. Earthquake (EQ)

• Regions Compared: 18,367
• Agreement: 58.4% (10,726 unchanged)
• Net Trend: Balanced (3,662 upgrades vs 3,979 downgrades)
• Average Change: -0.12 ranks
• Notable Shifts:
  • Medium → High: 3,094 regions (significant!)
  • High risk regions increased from 2,436 to 5,237

3. Tropical Cyclone (TC)

• Regions Compared: 6,107
• Agreement: 11.5% (705 unchanged) ⚠️ LOW AGREEMENT
• Net Trend: Strong Downgrades (5,315 downgrades vs 87 upgrades)
• Average Change: -1.82 ranks
• Notable Shifts:
  • High → Low: 2,674 regions
  • High risk regions dropped from 4,164 to 306 (major reassessment)

4. Volcano (VO)

• Regions Compared: 1,316
• Agreement: 55.5% (731 unchanged)
• Net Trend: Downgrades (542 downgrades vs 43 upgrades)
• Average Change: -1.16 ranks
• Notable Shifts:
  • High risk regions decreased from 1,045 to 681

5. Extreme Heat (EH)

• Regions Compared: 29,628
• Agreement: 88.4% (26,185 unchanged) ✓ HIGH AGREEMENT
• Net Trend: Slight downgrades (3,023 downgrades vs 420 upgrades)
• Average Change: -0.79 ranks
• Most Stable: This hazard showed the highest consistency

6. Tsunami (TS)

• Regions Compared: 4,729
• Agreement: 96.0% (4,540 unchanged) ✓ VERY HIGH AGREEMENT
• Net Trend: Slight downgrades (169 downgrades vs 20 upgrades)
• Average Change: -0.81 ranks
• Most Stable: Minimal changes overall

7. Wildfire (WF)

• Regions Compared: 31,742
• Agreement: 51.3% (16,269 unchanged)
• Net Trend: Strong Downgrades (15,071 downgrades vs 402 upgrades)
• Average Change: -1.64 ranks
• Notable Shifts:
  • High → Very Low: 2,502 regions
  • High risk regions decreased from 22,744 to 14,978

---

Hazards Not Yet Updated (2025)

The following hazards appear in the original dataset but not yet calculated for 2025:

• FL - River Flood
• CF - Coastal Flood
• DG - Drought

---

Unmatched Regions Analysis

Top Countries with Unmatched Regions (Original)

1. Turkey: 926 regions
2. Mexico: 505 regions
3. Chad: 345 regions
4. Venezuela: 317 regions
5. Denmark: 271 regions

Top Countries with Unmatched Regions (2025)

1. Mexico: 1,086 regions
2. Turkey: 926 regions
3. Latvia: 587 regions
4. Venezuela: 317 regions
5. Kenya: 290 regions

Note: Unmatched regions are primarily due to:

• Boundary updates and administrative reorganization
• Name changes (e.g., "Turkey" → "Türkiye", "Vietnam" → "Viet Nam")
• Mergers/splits of administrative units
• Data quality improvements

---

Overall Interpretation

Major Changes:

1. Tropical Cyclone - Dramatic reassessment with most high-risk areas downgraded
2. Wildfire - Significant downward revision in risk levels
3. Earthquake - Notable increase in high-risk classifications
4. Landslide - Moderate downward revision

Stable Hazards:

1. Tsunami - 96% agreement, very consistent
2. Extreme Heat - 88% agreement, stable methodology

Reasons behind score changes:

• Updated hazard data and approach (Major changes for TC, LS, WF)
• Improved spatial resolution (LS)
• Methodological refinements
• New boundary datasets

---

Files Generated

https://worldbankgroup-my.sharepoint.com/:f:/g/personal/mamadio_worldbank_org1/EuEuuNHaPFhNpkU9ztUL5TwBNCz1q4Tq-vot47rtklfNyA?e=JoWMgI

1. TH_scores_comparison_report.xlsx - Full comparison report with:

   • Summary statistics
   • Detailed region-by-region comparison
   • Match quality scores
   • Unmatched regions
   • Change matrices for each hazard
   • README sheet with interpretation guide

2. compare_hazard_scores_v2.py - Python script for reproducing analysis

[matamadio]

Implementation of scores via DDH API

Once scores are finalised, those will be uploaded on WB datacatalog which now has API capabilities. TH will harvest the scores from there and, ideally, it should update automatically when the scores file is updated.

This possibly remove the need to maintain an API service from TH and just redirect to the DDH service - to be discussed @pzwsk

[matamadio]

Hazardsets summary

@arnaud-morvan below is the text to be shown below the minimap for each hazard
@stufraser1 please check if phrased correctly. Thought it would be good to link them through RDL when possible. Some needs to be added.

• FL/PL/CF_Hazard_source: based on the global River/Pluvial/Coastal flood hazard maps produced by Fathom (v3). Data are not publicly disclosed due to license terms.

• DR_Hazard_source: based on the global water stress (baseline) hazard maps produced by Aqueduct (v4).

• LS_Hazard_source: based on the global Landslide susceptibility index published by UNEP/GIRI (2025). Available for download at Risk Data Library.

• EQ_Hazard_source: based on the global Earthquake hazard maps published by GAR (2017).

• TC_Hazard_source: based on the global STORM (v4) wind hazard maps published by Bloemendaal N. (2023). Available for download at Risk Data Library.

• VO_Hazard_source: based on the global Volcanic eruption database published by NOAA (2025).

• EH_Hazard_source: based on the global extreme heat maps published by GFDRR-VITO (2025). Available for download at Risk Data Library.

• TS_Hazard_source: based on the Global Tsunami Model published by GTM network (2017).

• WF_Hazard_source: based on the global wildfire hazard maps published by CEMS (2020).

[matamadio]

@arnaud-morvan please evaluate alternative options for implementing this, i.e. file import VS admin UI page.

Mockup proposal of expected outcome:

[arnaud-morvan]

As discussed today, it seems acceptable to have those values hardcoded in source code for each hazardtype.

[matamadio]

As discussed today, it seems acceptable to have those values hardcoded in source code for each hazardtype.

Yes - as long as there is a quick action to re-deploy website on changes.

[matamadio]

Scores GPKG - Download for WB: https://worldbankgroup-my.sharepoint.com/:u:/g/personal/mamadio_worldbank_org1/IQD39K9_tjlBSaEhTr3oMc5vAXsun78keK_0ZoPONi-dJuE?e=KTMLuh