## wpiea2024204-print-pdf

## Source details

**Canonical URL:** [wpiea2024204-print-pdf](https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024204-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2024/english/wpiea2024204-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2024/english/wpiea2024204-print-pdf.pdf.json)

---

### Purpose and Scope
- Examines the potential use of open-source satellite and building footprint data to compile a spatially explicit housing (residential building) census and demonstrates its application in climate risk assessment.
- Introduces a technique to assign monetary value to residential buildings using readily available aggregate capital stock data.

### Motivation and Policy Relevance
- Housing is both investment and consumption, accounting for the largest share of national wealth and closely linked to financial system health and monetary transmission.
- Up-to-date, spatially explicit housing censuses are often lacking due to cost and capacity constraints, limiting policy formulation, implementation, and monitoring for issues such as climate change.
- Detailed, global-scale residential property data are crucial to quantify financial and economic exposure and vulnerability to climate hazards and to analyze transition risks and opportunities arising from low-carbon policy reforms.

### Data and Methods Overview
- Primary geospatial datasets and processing:
  - Google Open Buildings (latest version v3, released in May 2023): about 1.8 billion buildings across continental Africa, South Asia, South-East Asia, Latin America and the Caribbean; source imagery ~50 centimeters; Deep Learning models.
  - GHSL ANBH (average net building height) estimates for 2018 at 100m x 100m resolution (10m x 10m estimates also available; study used 100m x 100m).
  - GHSL built-up characteristics layer (GHS-BUILT-C) 2018 classification at 10 meter resolution (residential/non-residential).
  - Gridded Population of the World Version 4 (GPWv4) 2020 estimates at 30 arc-second resolution (~1 km).
  - Corrected VIIRS annual average nighttime light imageries (2020) at 15 arc-second resolution (~500 meters) from EOG.
  - Investment and Capital Stock Datasets (ICSD) compiled by IMF FAD covering ~170 countries from 1960 to 2019 for dwellings capital stock estimation.
- Key methodological steps:
  - Detect and refine building attributes: location, ground-level area, height (from GHSL ANBH), classification (GHSL), and confidence (Google Open Buildings).
  - Impute building heights where GHSL ANBH = 0 using the minimum non-zero ANBH for the country (Kenya minimum non-zero ANBH = 2.5 meters).
  - Estimate number of stories by dividing imputed building height by minimum non-zero height (2.5 meters for Kenya).
  - Compute total building area = ground-level area × number of stories (excluding below-ground floors).
  - Downscale aggregate dwellings capital stock to pixel and building square-meter levels using population and nighttime lights with alternative weighting schemes (LitPop, logLit+logPop, weighted LitPop with α and β parameters).

### Key Facts and Quantities (preserved verbatim)
- Construction and building operations contributed to 38%, or 13.1 gigatons, of global energy-related CO2 emissions in 2015 (United Nations Environment Programme, 2021).
- IMF data from 2019 estimated the total capital stock at $316,253 billion in international dollars.
- The built-up area accounts for approximately 83% of the variation in total capital stock.
- Spatial resolution and dataset specifics:
  - Early global models used a spatial resolution of 5 km.
  - Google Open Buildings uses high-resolution (~10 meters) satellite imagery.
  - Landsat imagery used previously has a 30-meter resolution.

### Descriptive Statistics — Kenya (preserved verbatim where provided)
- Total number of buildings detected by Google Open Buildings: 26,405,031.
- Aggregate ground-level building area: 1,324,699,574 square meters.
- Aggregate height-adjusted building areas: 1,367,850,580 square meters.
- Average ground-level building area: around 50 square meters.
- Average confidence score: 0.78.
- After filtering (outliers, <10 square meters, low-population pixels) number of buildings dropped from 26.41 million to 22.04 million; reported final number: 22,044,652.
- Table 1 highlights (Kenya):
  - Ground-level area (square meters): Mean 50.19; Median 30.54; Min 2.51; Max 49,630.01.
  - Confidence score: Mean 0.78; Median 0.78; Min 0.65; Max 0.99.
  - Buildings with confidence score (%):
    - [0.9, 1.0): 2.65
    - [0.8, 0.9): 36.14
    - [0.7, 0.8): 44.07
    - [0.65, 0.7): 17.07
  - Average net building height (meters): Mean 1.29; Median 0.00; Min 0.00; Max 41.52.
  - Proportion of zero building heights (%): 50.93
  - Built-up area classification (%):
    - Non-built-up 83.19
    - Residential 16.78
    - Non-residential 0.03
  - Estimated average # of building stories/floors*: Mean 1.02; Median 1.00; Min 1.00; Max 17.0
  - Population per pixel (2018)—GHSL (100m): Mean 4.42; Median 0.00; Min 0.00; Max 5,329.20.
    - Proportion of zero (%)—GHSL 51.4
  - Population (2020) per pixel—GPWv4: Mean 1,340.10; Median 502.4; Min 0.00; Max 143,427.70.
    - Proportion of zero (%)—GPWv4 0.00
  - Nighttime light radiance per pixel (nW/cm2/sr): Mean 1.54; Median 0.00; Min 0.00; Max 75.8.
    - Proportion of zero NTL 63.65

### Indicator Framework Proposed
- Three main dimensions of residential properties exposure indicators:
  i) Quantity indicators: square meters of residential properties per geographic unit (square km, grid cell, administrative boundary, etc.).
  ii) Value indicators: value of residential properties per geographic unit.
  iii) Density indicators: square meters of residential properties per person; national share of residential properties per geographic unit; national share of residential properties value per geographic unit.
- Indicators align with criteria: relevance, methodological soundness, measurability, clarity.

### Downscaling Approaches and Estimated Parameters (Kenya)
- Unweighted LitPop downscaling formula:
  - V_ci = V_c × (Lit_ci · Pop_ci) / Σ_i (Lit_ci · Pop_ci) with rules when Lit_ci or Pop_ci = 0.
- Log-transformed alternative:
  - logLit · logPop = log(Lit_ci + 1) · log(Pop_ci + 1).
- Weighted downscaling:
  - Ṽ_ci = V_c × (Lit_ci^α · Pop_ci^β) / Σ_i (Lit_ci^α · Pop_ci^β), with α + β = 1 (allow α and β to differ for urban and rural).
- Calibration via log-transformed Cobb-Douglas regression using Kenya National Housing Survey (2012/2013; sample size 8,884 households):
  - α = 0.9; β = 0.1
  - α_u = 0.95; β_u = 0.05
  - α_r = 0.88; β_r = 0.12
  - Interpretation: nighttime light has higher weight than population in determining residential property values in Kenya; higher weight in urban areas than rural.

### Results: Residential Property Value and Area Statistics (Table 2 highlights)
- Building area (sqr meters): Mean 46 (Std. Dev. 39); Min 10; Max 2,128.
- No. of buildings: 22,044,652.
- Unweighted downscaling:
  - Value per sqr meter: LitPop $102 ($723) $0 $14,170
  - Value per sqr meter: logLit+logPop $197 ($973) $2 $97,244
  - Building value: LitPop $6,678 ($58,041) $0 $8,209,556
  - Building Value: logLit+logPop $6,676 ($19,109) $27 $1,658,925
- Weighted downscaling:
  - Value per sqr meter: LitPop (pooled) $130 ($372) $0 $659,723
  - Value per sqr meter: LitPop (urban/rural) $152 ($373) $0 $385,948
  - Value per sqr meter: logLit+logPop (pooled) $207 ($1,141) $2 $114,297
  - Value per sqr meter: logLit+logPop (urban/rural) $206 ($1,121) $2 $117,546
  - Building value: weighted LitPop (pooled) $6,678 ($18,246) $0 $22,794,341
  - Building value: weighted LitPop (urban/rural) $6,677 ($12,611) $6 $12,552,431
  - Building value: weighted logLit+logPop (pooled) $6,676 ($22,067) $31 $1,668,971
  - Building value: weighted logLit+logPop (urban/rural) $6,676 ($21,741) $31 $1,639,328
- Note: Values are in constant 2017 international dollars using the GFCF deflators and purchasing power parities taken from the OECD and PWT depending on data availability.
- Observation: logLit+logPop approach reduces variation (CV) compared to LitPop, addressing extreme-value inflation.

### Validation with Housing Survey
- Validation dataset: 2012-2013 Kenyan National Housing Survey (44 counties; sample size 8,884).
- Pearson correlations between log house value per square meter and log downscaled values:
  - Unweighted: 0.3
  - Weighted (pooled): 0.27
  - Weighted (urban/rural): 0.14
- Interpretation: unweighted downscaled dwellings capital stock values have stronger correlation with survey values than weighted downscaling; note seven-year difference between survey (2012/2013) and capital stock values (2020) may affect correlations.

### Application — Residential Exposure to Riverine Flood Risks (Kenya) — Key Findings
- Total estimated residential built area: 915.1 million square meters that are estimated to be worth 147.2 billion in 2017 constant international dollars (PPP adjusted).
- Exposure to riverine flood:
  - More than 3.5 million square meters are exposed to 10-year return period riverine flood.
  - More than 4.1 million square meters are exposed to 20-year return period riverine flood.
  - Corresponding estimated monetary values:
    - 10-year return period exposure: 915.5 million in constant 2017 international dollars.
    - 20-year return period exposure: 1 billion in constant 2017 international dollars.
- Comparative macro context:
  - Corresponding real GDP data from the same source was 228 billion in constant 2017 international dollars.
- Spatial and value patterns:
  - Quantity (area) and estimated monetary value of residential properties within flood risk zones decrease with severity as measured by depth in meters.
  - The estimated total area or stock of residential properties within flood risk zones declines as flood depth increases; same pattern holds for total estimated monetary value.
  - Presence of buildings in severe flood-risk areas is attributed to factors such as urban residential land shortage or poverty.
- Flood hazard data:
  - EC-JRC Global Flood Hazard Map (30 arcseconds ≈ 1 km) showing water depth for return periods.
  - Return period interpretation:
    - 10-year flood: 1-in-10 (10%) chance in any one year.
    - 20-year flood: 1-in-20 (5%) chance in any one year.

### Limitations, Data Issues, and Risks
- Building height data:
  - GHSL ANBH has zero values for a significant proportion of pixels where buildings were detected; about 51 percent of GHSL ANBH pixels have zero values.
  - Zero-height pixels with detected buildings are imputed with the minimum non-zero ANBH for the country (Kenya minimum non-zero ANBH = 2.5 meters).
  - Reliance on such imputations raises questions about GHSL height accuracy.
- Building classification and population estimates:
  - More than 83 percent of the pixels within which buildings are detected are misclassified as non-built-up area; analysis removes only pixels classified as non-residential and keeps detected buildings located in pixels with a human settlement.
  - GHSL population estimates are zero for some pixels where buildings are detected while GPWv4 provides non-zero estimates.
- Capital stock data gaps:
  - Absence of dwelling-specific fixed capital stock for many countries; study uses average shares by income group to infer dwelling share:
    - LIDCs: 0.748
    - EMs: 0.362
    - AEs: 0.171
- Algorithmic bias and model error risks:
  - Machine learning-based layers (Google building footprints, GHSL classifications, GHSL heights, EC-JRC population) can propagate biases from training data, human labeling, terrain/topology differences, and statistical limitations.

### Policy Implications and Potential Uses
- Spatially explicit housing exposure data can close data gaps for countries lacking detailed housing statistics, improving assessment of financial markets, housing supply and demand, and climate impacts on housing stock.
- The dataset can feed into IMF macroeconomic models such as the Debt-Investment-Growth-Natural-Disasters (DIGNAD) model.
- Addresses a recognized data gap under the G20 Data Gap Initiative phase III (DGI-3) recommendation five by providing a cost-effective approach using publicly available satellite and geospatial datasets.
- Enables targeted risk management and policy design for residential buildings as climate actions and demand for granular exposure/vulnerability data grow.

*IMF WORKING PAPER — Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment*

### References .............................................................................................................

### wpiea2024204-print-pdf - References

### Purpose and Scope
- Examines the potential use of open-source satellite and building footprint data to compile a spatially explicit housing (residential building) census and demonstrates its application in climate risk assessment.
- Introduces a technique to assign monetary value to residential buildings using readily available aggregate capital stock data.

### Motivation and Policy Relevance
- Housing is both investment and consumption, accounting for the largest share of national wealth and closely linked to financial system health and monetary transmission.
- Up-to-date, spatially explicit housing censuses are often lacking due to cost and capacity constraints, limiting policy formulation, implementation, and monitoring for issues such as climate change.
- Detailed, global-scale residential property data are crucial to quantify financial and economic exposure and vulnerability to climate hazards and to analyze transition risks and opportunities arising from low-carbon policy reforms.

### Key Facts and Quantities (preserved verbatim)
- Construction and building operations contributed to 38%, or 13.1 gigatons, of global energy-related CO2 emissions in 2015 (United Nations Environment Programme, 2021).
- IMF data from 2019 estimated the total capital stock at $316,253 billion in international dollars.
- The built-up area accounts for approximately 83% of the variation in total capital stock (see Figure 1 in Appendix A).
- Spatial resolution and dataset specifics mentioned:
  - Early global models used a spatial resolution of 5 km.
  - Google Open Buildings uses high-resolution (~10 meters) satellite imagery.
  - Landsat imagery used previously has a 30-meter resolution.
  - Figures presented include: Spatial Distribution of Residential Properties Indicators for Kenya, by Grid Cell at a Resolution Level of 5; Spatial Distribution of Residential Properties Indicators for Kenya, by Admin Level 3; Flood Hazard Map of 10-years and 20-years Return Periods, Kenya; Quantity and Value of Residential Properties Exposed to Riverine Flood Risks, Kenya.
  - Appendix figures: A.1 Built Up Area and Capital Stock; A.2 Density of Log Estimated Property Values; A.3 Illustration of Hierarchical Grid Cells and Level 3 Admin Boundaries, Kenya; A.4 Key Satellite Datasets and Sources Used for this Study – Extracted from Nairobi Area, Kenya.

### Data and Methods Overview
- Uses open-source building footprint data (Google Open Buildings) and Sentinel-2 imagery to detect and classify buildings, refine building attributes (ground-level area, height, classification as residential or non-residential).
- Proposes a census approach leveraged by buildings’ fixed characteristics: location, construction year, size, shape, height, construction materials, and purpose.
- Suggests assigning monetary value to residential buildings by combining the spatially explicit building census with aggregate capital stock data.

### Contributions to Literature and International Initiatives
- Extends earlier global exposure databases (e.g., De Bono and Mora, 2014) by employing higher-resolution imagery and detailed building footprints.
- Builds on and complements related methodologies (Eberenz et al., 2020; Doan et al., 2023) and aligns with the Data Gap Initiative phase III (DGI-3) recommendations for climate data and statistics.
- Supports national statistical offices with limited resources by presenting a cost-effective alternative to compile granular structure and residential building data using publicly available satellite and geospatial datasets.

### Applications and Policy Implications
- Enables climate risk assessment by producing spatially explicit exposure layers for residential properties, which are essential for:
  - Quantifying physical exposure to hazards (e.g., riverine flood risks with 10-years and 20-years return periods).
  - Informing transition-risk analysis tied to energy-efficiency regulations, retrofitting requirements, and low-carbon policy measures.
  - Supporting financial stability analysis where housing markets are key transmission channels for monetary policy.

### Structure of the Document (as provided)
- Section 2: discusses data sources used to compile exposure layers and provides summary statistics.
- Section 3: discusses methods and proposed indicators.
- Appendix and Figures/Tables include descriptive statistics, spatial distributions, hazard maps, and additional figures (e.g., A.1–A.4) and tables (e.g., Table 1, Table 2, Table A.1).

*IMF Working Paper — Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment (excerpt: Introduction, figures/tables/appendix listings and references).*

### Section 4 discusses the analytical results and limitations of the proposed indicators, and section 5 concludes

### Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment (sections II excerpt)

### Data and Description
- Processed satellite and geospatial datasets are used to compute residential properties exposure layers, including:
  - building footprints, estimated building height and classification,
  - gridded population data,
  - urbanization, nighttime light, and dwellings capital stock.
- The primary case study demonstrated is Kenya.
- Table A.1 and Figure A.4 are referenced for data descriptions and imageries (not reproduced here).

### Google’s Open Buildings
- Primary dataset: Google’s Open Buildings dataset (latest version v3, released in May 2023).
- Coverage: about 1.8 billion buildings across continental Africa, South Asia, South-East Asia, Latin America and the Caribbean.
- Source imagery and method: high-resolution (~50 centimeters) daytime satellite imagery and Deep Learning models (Sirko et al., 2021).
- Data fields: building polygon, latitude and longitude of the centroid, area in meters, and confidence levels for each detected building.
- Licenses: Creative Commons Attribution (CC BY-4.0) and Open Data Commons Open Database License (ODbL) v1.0.
- Limitation: does not include building heights and other attributes needed to compute proposed residential exposure indicators.
- Microsoft’s Building Footprint noted as another open-source building data covering about 1.2B buildings; Microsoft also has 174M building height estimates providing 3D polygons for some areas.

### Additional building attributes: Height and Classification
- Building height and classification derived from European Commission Joint Research Center (EC-JRC) Global Human Settlement Layer (GHSL).
- Building height estimates:
  - Methods by Pesaresi et al (2021) and Pesaresi and Politis (2023).
  - Utilize Digital Elevation Models (DEMs) from ALOS Global Digital Surface Model (AW3D30) and Shuttle Radar Topography Mission, plus shadow markers from Sentinel-2 image composites for 2018.
  - Study uses average net building height (ANBH) estimates for 2018 at 100m x 100m resolution.
  - Note: building height estimates also available at 10m x 10m resolution; this study used the 100m x 100m layer.
- Building functional classification:
  - GHSL built-up characteristics layer (GHS-BUILT-C) classifies structures into residential and non-residential at 10 meter resolution.
  - Study uses 2018 classification data.

### Gridded Population
- Dataset: Gridded Population of the World Version 4 (GPWv4) from NASA SEDAC.
- Advantages: population densities calibrated to UN World Population Prospects (UN WPP) country totals.
- Resolution and period: 30 arc-second resolution (~1 km at the equator); covers 2000 to 2020 in five-year intervals.
- This study uses the 2020 global population estimates.
- Note: GHSL gridded population data (100m and 1km resolutions) is also available for cross-examination.

### Nighttime Light
- Dataset: corrected VIIRS annual average nighttime light imageries from the Earth Observation Group (EOG) at Colorado Mines.
- Purpose: proxy measure of human and economic activities and used to downscale country-level dwellings capital stock values.
- Data characteristics:
  - Corrected monthly and annual VIIRS imageries available since 2014 at 15 arc-second resolution (around 500 meters at the Equator).
  - Study uses annual nighttime light composite for 2020 to correspond with GPWv4 data.
- Processing: raw satellite images processed for extraneous artifacts and biases (sunlit, moonlit, cloudy, other light contaminants); cloud and lunar-BRDF-corrected composites used.

### Dwellings Capital Stock
- Source: Investment and Capital Stock Datasets (ICSD) compiled by IMF Fiscal Affairs Department (FAD) covering ~170 countries from 1960 to 2019.
- Use: most recent private fixed capital stock data from ICSD to estimate dwellings capital stock value at the square meter level.
- Approach to dwelling share:
  - Lack of dwelling-specific fixed capital stock for many countries addressed by using average share of dwelling in fixed capital formation from countries with available data.
  - Averages calculated by country income group: Low Income and Developing Countries (LIDCs), Emerging Markets (EMs), and Advanced Economies (AEs).
  - Corresponding average shares of dwellings in real gross fixed capital stock for 2015–2019:
    - LIDCs: 0.748
    - EMs: 0.362
    - AEs: 0.171
- Downscaling procedure:
  - Downscale values to pixel level based on population and nighttime lights, with pixel resolution determined by the higher resolution of either nighttime light or gridded population.
  - Pixel-level average values further downscaled to square meter levels using share of the building area within a pixel as downscaling factor.
- Methods described further in section III (not reproduced here).

### Descriptive Statistics (Kenya case study)
- Total number of buildings detected by Google Open Buildings: 26,405,031.
- Average ground-level building area: around 50 square meters.
- Aggregate ground-level building area: 1,324,699,574 square meters.
- Aggregate height-adjusted building areas: 1,367,850,580 square meters.
- Average confidence score: 0.78 (minimum and maximum values referenced but not provided in the excerpt).

*IMF Working Paper — sections II excerpt (Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment)*

### 0.65 and 0.99, respectively, and more than 83 percent of detected buildings have confidence scores of 0.7 and

### wpiea2024204-print-pdf - 0.65 and 0.99, respectively, and more than 83 percent of detected buildings have confidence scores of 0.7 and

### Data sources, preprocessing, and key assumptions
- Building detections augmented with GHSL ANBH layer (pixel size 100m x 100m); for each detected building the average building height is extracted.
- Assumption: buildings within a 100m x 100m GHSL ANBH pixel have the same height.
- About 51 percent of GHSL ANBH pixels have zero values; zero-height pixels with detected buildings are imputed with the minimum non-zero ANBH for the country.
  - For Kenya the minimum non-zero ANBH is 2.5 meters.
- Number of stories/floors per detected building approximated by dividing imputed building heights by the minimum non-zero height value (2.5 meters for Kenya).
- Ground-level area multiplied by number of stories to compute estimated total building area, excluding below-ground floors.
- Built-up area classification has three classes: non-built-up, residential, non-residential.
  - More than 83 percent of the pixels within which buildings are detected are misclassified as non-built-up area.
  - Analysis removes only pixels classified as non-residential and keeps detected buildings located in pixels with a human settlement.
- Confidence score information used for robustness checks where height, population, and building classification are zero or null.
- Filters applied:
  - Exclude outliers using rule |xi − x̃|/sd(x) > 3 (median x̃; sd standard deviation).
  - Exclude buildings smaller than 10 square meters.
  - Exclude buildings within GPWv4 pixels (1 square km) with fewer than four people (average household size in Kenya).
- After filtering, number of buildings dropped from 26.41 million to 22.04 million.

### Descriptive statistics for Kenya (Table 1 highlights)
- Ground-level area (square meters): Mean 50.19; Median 30.54; Min 2.51; Max 49,630.01.
- Confidence score: Mean 0.78; Median 0.78; Min 0.65; Max 0.99.
- Buildings with confidence score (%):
  - [0.9, 1.0): 2.65
  - [0.8, 0.9): 36.14
  - [0.7, 0.8): 44.07
  - [0.65, 0.7): 17.07
- Average net building height (meters): Mean 1.29; Median 0.00; Min 0.00; Max 41.52.
- Proportion of zero building heights (%): 50.93
- Built-up area classification (%):
  - Non-built-up 83.19
  - Residential 16.78
  - Non-residential 0.03
- Estimated average # of building stories/floors*: Mean 1.02; Median 1.00; Min 1.00; Max 17.0
  - Note: computed after imputing zero height with minimum non-zero average building height of 2.5 meters for Kenya.
- Population per pixel (2018)—GHSL (100m): Mean 4.42; Median 0.00; Min 0.00; Max 5,329.20.
  - Proportion of zero (%)—GHSL 51.4
- Population (2020) per pixel—GPWv4: Mean 1,340.10; Median 502.4; Min 0.00; Max 143,427.70.
  - Proportion of zero (%)—GPWv4 0.00
- Nighttime light radiance per pixel (nW/cm2/sr): Mean 1.54; Median 0.00; Min 0.00; Max 75.8.
  - Proportion of zero NTL 63.65
- Number of buildings (after filtering reported later): 22,044,652

### Indicator framework proposed
- Three main dimensions of residential properties exposure indicators:
  i) Quantity indicators: square meters of residential properties per geographic unit (square km, grid cell, administrative boundary, etc.).
  ii) Value indicators: value of residential properties per geographic unit.
  iii) Density indicators: square meters of residential properties per person; national share of residential properties per geographic unit; national share of residential properties value per geographic unit.
- Indicators align with recommended criteria: relevance, methodological soundness, measurability, clarity.

### Downscaling aggregate dwellings capital stock values
- Unweighted LitPop downscaling (equation (1)):
  - V_ci = V_c × (Lit_ci · Pop_ci) / Σ_i (Lit_ci · Pop_ci), with rules for Lit_ci · Pop_ci when Lit_ci or Pop_ci are zero.
  - Sensitivity to extreme values noted.
- Log-transformed alternative: logLit · logPop = log(Lit_ci + 1) · log(Pop_ci + 1) to reduce influence of extreme values.
- Residential property value per square meter within pixel i:
  - V_ci_sqr.m = V_ci / Σ_j a_ci,j, where a_ci,j is area of residential building j and n_j number of residential buildings within pixel i.

### Weighted downscaling and calibration
- Weighted downscaling (equation (3)):
  - Ṽ_ci = V_c × (Lit_ci^α · Pop_ci^β) / Σ_i (Lit_ci^α · Pop_ci^β), with α + β = 1.
  - Allow α and β to differ for urban (α_u + β_u = 1) and rural (α_r + β_r = 1) areas.
- Calibration via log-transformed Cobb-Douglas regression (equation (4)):
  - log(h_ij) = α log(Lit_ij) + β log(Pop_ij) + ε_ij, with α + β = 1.
- Estimated parameters using 2012/2013 Kenya National Housing Survey (sample size 8,884 households):
  - α = 0.9; β = 0.1
  - α_u = 0.95; β_u = 0.05
  - α_r = 0.88; β_r = 0.12
  - Interpretation: nighttime light has higher weight than population in determining residential property values in Kenya; higher weight in urban areas than rural.

### Aggregation approaches
- Two aggregation approaches demonstrated:
  - Equally spaced grid cells using H3 hierarchical indexing (example: hex resolution 5 ≈ 252.9 square km).
  - Level 3 administrative boundaries (lowest available administrative level for Kenya).
- Trade-offs noted: grid cells flexible for aggregation scale; administrative boundaries align with policy and demographic units.

### Results: summary statistics of computed residential property indicators (Table 2 highlights)
- Building area (sqr meters): Mean 46 (Std. Dev. 39); Min 10; Max 2,128.
- No. of buildings: 22,044,652.
- Unweighted downscaling:
  - Value per sqr meter: LitPop $102 ($723) $0 $14,170
  - Value per sqr meter: logLit+logPop $197 ($973) $2 $97,244
  - Building value: LitPop $6,678 ($58,041) $0 $8,209,556
  - Building Value: logLit+logPop $6,676 ($19,109) $27 $1,658,925
- Weighted downscaling:
  - Value per sqr meter: LitPop (pooled) $130 ($372) $0 $659,723
  - Value per sqr meter: LitPop (urban/rural) $152 ($373) $0 $385,948
  - Value per sqr meter: logLit+logPop (pooled) $207 ($1,141) $2 $114,297
  - Value per sqr meter: logLit+logPop (urban/rural) $206 ($1,121) $2 $117,546
  - Building value: weighted LitPop (pooled) $6,678 ($18,246) $0 $22,794,341
  - Building value: weighted LitPop (urban/rural) $6,677 ($12,611) $6 $12,552,431
  - Building value: weighted logLit+logPop (pooled) $6,676 ($22,067) $31 $1,668,971
  - Building value: weighted logLit+logPop (urban/rural) $6,676 ($21,741) $31 $1,639,328
- Note: Values are in constant 2017 international dollars using the GFCF deflators and purchasing power parities taken from the OECD and PWT depending on data availability.
- Observation: logLit+logPop approach reduces variation (CV) compared to LitPop, addressing extreme-value inflation.

### Validation with housing survey
- Validation data: 2012-2013 Kenyan National Housing Survey (44 counties; sample size 8,884).
- Downscaled values (unweighted and weighted) spatially joined to H3 polygons (resolution 7) and compared to survey house values.
- Pearson correlations between log house value per square meter and log downscaled values:
  - Unweighted: 0.3
  - Weighted (pooled): 0.27
  - Weighted (urban/rural): 0.14
- Interpretation: unweighted downscaled dwellings capital stock values have stronger correlation with survey values than weighted downscaling that assigns higher weights to nighttime light.
- Limitation noted: seven-year difference between housing survey (2012/2013) and capital stock values (2020) may affect correlations.

### Application: exposure to riverine flood risk (Kenya illustration)
- Flood hazard layer used: EC-JRC Global Flood Hazard Map (30 arcseconds ≈ 1 km) showing water depth for return periods.
- Return period interpretation:
  - 10-year flood: 1-in-10 (10%) chance in any one year.
  - 20-year flood: 1-in-20 (5%) chance in any one year.
- Method: combine residential properties layer with flood hazard map to calculate aggregate residential properties and dollar values at risk.
- For illustration, unweighted downscaled capital stock values used.
- (Text ends at "Our calculation shows that the total area of residential properties in 2019 Kenya were" — the numeric totals and specific exposure dollar values for the flood-risk illustration are not included in the supplied content.)

*IMF WORKING PAPER Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment*

### 915.1 million square meters that are estimated to be worth 147.2 billion in 2017 constant international dollars

### Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment

### Residential exposure to riverine flood risks (Kenya) — key findings
- Total estimated residential built area: 915.1 million square meters that are estimated to be worth 147.2 billion in 2017 constant international dollars (PPP adjusted).
- Exposure to riverine flood:
  - More than 3.5 million square meters are exposed to 10-year return period riverine flood.
  - More than 4.1 million square meters are exposed to 20-year return period riverine flood.
  - Corresponding estimated monetary values:
    - 10-year return period exposure: 915.5 million in constant 2017 international dollars.
    - 20-year return period exposure: 1 billion in constant 2017 international dollars.
- Comparative macro context:
  - Corresponding real GDP data from the same source was 228 billion in constant 2017 international dollars.
- Spatial and value patterns:
  - Quantity (area) and estimated monetary value of residential properties within flood risk zones decrease with severity as measured by depth in meters.
  - The estimated total area or stock of residential properties within flood risk zones declines as flood depth increases; the same pattern holds for total estimated monetary value.
  - Presence of buildings in severe flood-risk areas is attributed to factors such as urban residential land shortage or poverty.

### Methodological approach and illustrative outputs
- Data construction approach:
  - A census approach using high-precision building footprint data from Google Open Buildings.
  - Monetary value imputed to building level by downscaling capital stock using gridded population and nighttime light as weighting factors.
- Illustrative visualizations referenced:
  - Figure 4 panels (a)-(d) show quantity and estimated values of residential properties within riverine flood risks of 10-year and 20-year return periods; vertical axes for panels (a) and (c) represent quantity in square meters, whereas panels (b) and (d) represent estimated values in international dollars (2017 constant prices); horizontal axes represent flood depth in meters.
  - Log-scale axes used for area and value visualizations (axes labels and scales described in the source).

### Limitations and data issues
- Building height data:
  - GHSL's estimated average net building heights are zero for a significant proportion of pixels where buildings were detected.
  - Imputation applied: zero heights were imputed with a minimum non-zero height, approximately 2.5 meters for Kenya.
  - Reliance on such ad-hoc imputations raises questions about GHSL height accuracy.
- Building classification and population estimates:
  - GHSL building classification sometimes inaccurately labels pixels as neither residential nor non-residential, despite Google Open Building detection.
  - Pixels with detected buildings were reclassified as residential under the assumption that most misclassified structures are in non-urban areas and unlikely to be commercial.
  - GHSL population estimates are zero for pixels where buildings are detected, while GPWv4 provides non-zero estimates.
- Capital stock data gaps:
  - Absence of detailed information on the dwelling component of gross capital stock for many countries is a major limitation.
- Algorithmic bias and model error risks:
  - Layers generated using machine learning models can propagate biases and errors from statistical limitations, incomplete data, human labeling bias, or regional terrain/topology variations.
  - Google building footprints, GHSL building classifications, estimated building heights, and EC-JRC population data each apply machine learning models with differing performance levels, potentially exacerbating inaccuracies.

### Policy relevance and potential uses
- Data utility:
  - Spatially explicit housing exposure data can close data gaps for countries lacking detailed housing statistics, enabling better assessment of financial markets, housing supply and demand, and climate impacts on housing stock.
  - The dataset can feed into IMF macroeconomic models such as the Debt-Investment-Growth-Natural-Disasters (DIGNAD) model.
- Contribution to global initiatives:
  - The lack of spatially explicit and comparable housing exposure data is a recognized data gap under the G20 Data Gap Initiative phase III (DGI-3) recommendation five; this research aims to contribute to literature assessing the economic impact of climate change.
- Practical implication:
  - As climate actions advance and demand for granular exposure and vulnerability data grows, the described approach illustrates how open-source satellite data can inform policy and targeted risk management for residential buildings.

### Additional data and notes (selected)
- Imputation detail: minimum non-zero height used for Kenya imputation is approximately 2.5 meters.
- Visualization notes:
  - Monetary values are in constant 2017 international dollars (PPP adjusted).
  - Panel axis conventions: panels (a) and (c) vertical axes are quantity in square meters; panels (b) and (d) vertical axes are estimated values in international dollars (2017 constant prices); horizontal axes are flood depth in meters.

*Source: IMF WORKING PAPER Satellite-Based Census of Residential Buildings: Application for Climate Risk Assessment*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024204-print-pdf.pdf_