## Satellite Data for Nowcasting: Estimating Cambodia’s GDP in Real Time Using Satellite Data in a Machine Learning Framework

## Source details

**Canonical URL:** [Satellite Data for Nowcasting: Estimating Cambodia’s GDP in Real Time Using Satellite Data in a Machine Learning Framework](https://www.imf.org/-/media/files/publications/selected-issues-papers/2026/english/sipea2026001.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/selected-issues-papers/2026/english/sipea2026001.pdf.md)
- [Structured JSON version](/-/media/files/publications/selected-issues-papers/2026/english/sipea2026001.pdf.json)

---

### Motivation – Why Satellite Data
- Cambodia has annual GDP data only, published with significant lag, limiting timely analysis and policy response.
- Satellite indicators are available near-real time and at granular levels, serving as proxies for sectoral economic activity (e.g., manufacturing, agriculture).
- In Cambodia, interpolated quarterly GDP growth rate (see Data and Methodology) is positively correlated with changes in Nighttime Light (NTL), Vegetation Health Index (VHI), and Precipitation (PCP).
- Machine learning models can exploit complex, non-linear interactions between satellite-derived indicators and traditional high-frequency indicators to produce monthly-updatable nowcasts.

### Data and Methodology
- Non-traditional (satellite-derived) indicators used:
  - Nighttime Lights (NTL)
  - Nitrogen Dioxide (NO₂)
  - Normalized Difference Vegetation Index (NDVI)
  - Enhanced Vegetation Index (EVI)
  - Vegetation Health Index (VHI)
  - Precipitation (PCP) (CHIRPS)
  - Agricultural Stress Index (ASI)
- Traditional high-frequency indicators used:
  - Exports and Imports
  - Broad Money
  - Exchange Rates
  - Consumer Price Index
  - Lending Rate
  - Credit
  - Tourist Arrivals
- Data sources include Google Earth Engine, NASA (NO₂), FAO - Agricultural Stress Index System (ASIS), and Climate Hazards Center (CHIRPS).
- For GDP (annual only for Cambodia), the authors apply an interpolated quarterly GDP series:
  - Quarterly GDP (year-on-year growth rates) constructed using Cambodia’s major trading partners aggregated with respective export weights. Major trading partners listed: Canada, China, Germany, Japan, Korea, Malaysia, Singapore, Thailand, United States, and Vietnam.
- Missing observations in traditional indicators are imputed based on historical patterns as needed.
- Machine learning approach:
  - Algorithm: Random Forest (Breiman, 2001).
  - Training/test split: 85 percent of historical data treated as “training”, 15 percent as “test” (chronological cross-validation with holdout sets).
  - Random forest prediction is the average of predictions from M decision trees (bagging); splits chosen to minimize root mean squared error (RMSE).
  - Model evaluation metrics include RMSE and Mean Absolute Error (MAE).
- Interpretation tools:
  - Shapley decomposition used to show contributions of variables to predictive power (not causal effects).

### Results and Interpretation
- Nowcast projections:
  - Nowcast GDP growth (year-on-year) of 5.7 percent in 2025q1 and 6.7 percent in 2025q2.
  - Specific nowcast for 2025q1 reported as 5.68 percent.
- Model fit and accuracy:
  - RMSE for the random forest model: 0.9.
  - Inclusion of satellite variables improved model accuracy by over 20 percent.
  - RMSE reduced from 1.2 to 0.9 when non-traditional indicators were added.
  - MAE reduced from 1.0 to 0.8 when non-traditional indicators were added.
- Variable importance and time patterns (Shapley contributions):
  - Over 2012–2025, NO₂ emissions have the largest influence among satellite indicators on model predictions (proxy for industrial activity).
  - In 2021–2025, Nighttime Lights (NTL) show a stronger influence, reflecting alignment with urbanization, tourism, household electricity access and consumption.
  - Vegetation indices contribute modestly but consistently, with spikes in 2017 and 2025q1, reflecting agriculture’s role and vulnerability to climate-driven volatility.
- Practical implications:
  - Satellite indicators meaningfully complement traditional indicators, particularly where traditional indicators are scarce or delayed.
  - Satellite data enable subnational, near-real-time insights that can reveal regional heterogeneities (e.g., increased NTL in Phnom Penh and Preah Sihanouk in 2025; lower NTL in Siem Reap and northeastern provinces in 2025, suggesting tourism or regional slowdowns).
  - Vegetation-related indices (e.g., VHI, ASI) indicate healthier vegetation in key cropland provinces in 2025q1–q2 compared with 2019 and 2024 in the authors’ calculations.

### Key Numerical Findings and Model Metrics
- Nowcasts:
  - 2025q1 GDP growth (y/y): 5.7 percent (reported also as 5.68 percent)
  - 2025q2 GDP growth (y/y): 6.7 percent
- Model performance:
  - RMSE (with satellite indicators): 0.9
  - RMSE (without satellite indicators): 1.2
  - MAE (with satellite indicators): 0.8
  - MAE (without satellite indicators): 1.0
  - Improvement from adding satellite variables: over 20 percent (RMSE and MAE metrics)

*Prepared by Iyke Maduako, Dharana Rijal, and Alberto Sanchez Rodelgo. SIP/2026/001 (Completed on November 5, 2025).*

---


_Source: https://www.imf.org/-/media/files/publications/selected-issues-papers/2026/english/sipea2026001.pdf_