Machine Learning Predictions of GDP at higher spatial and sectoral resolution

• Daniel D. Moran
  Senior Scientist, NILU
  Distinguished Professor, Tohoku University Graduate School of Environmental Studies
• Mohamed-Bachir Belaid
  Senior Scientist, NILU
• Francis Isidore Barre
  Scientist, NILU
  PhD candidate, Norwegian Science and Technology University (NTNU)
• Keiichiro Kanemoto
  Associate Professor, Tohoku University Graduate School of Environmental Studies

Gridded GDP → Neural Network

Gridded GDP estimates total GDP from each pixel

But it can systematically overpredict (blues) or underpredict (reds) true employment.

And has no information about sectoral composition.

An AI model can predict total employment and sectoral composition.

Context

Economic maps, locating which businesses operate where, are useful in many applications including:

• Constraining pollution sources / drivers of environmental impact
• Assessing vulnerability, e.g. to physical climate risk
• Performing regional economic impact assessments (disaster IO; building IRIOs)

The Knowledge Gap

A good map offers:

• Accessible without prohibitive costs
• Comprehensive covering more than fixed point sources
• Differentiates sectors, e.g. agriculture, industry, services
• Global, consistently covering many countries

The options available today are insufficient:

Proposed Solution

Train an ML model to predict employment per sector in a map tile based on the surrounding features.

The Task: Learn how features (e.g. OpenStreetMap features, night luminosity, etc.) in a tile and surrounding neighborhood predict employment composition in a tile. Use training data from Norway on employment per sector in each tile.

Tile Example

XGBoost Regression and Neural Network

XGBoost builds a multivariate regression model. In this illustrative 3-dimensional graph, imagine the X axis is luminosity, the Y axis is # of buildings in the tile, and the Z axis is total employment. XGBoost builds a model from many input features to predict total employment in the tile.

Image source

The neural network model accepts a vector of input features (e.g. # of schools, # of airports), and predicts a vector of output features (e.g. share of employment per sector).

Glosser.ca - Own work, Derivative of File:Artificial neural network.svg

Model Architecture

Steps:

1. Filter out empty tiles to focus on inhabited areas
2. Predict total employment per tile using an XGBoost regression model
3. Predict employment composition (sector shares, summing to 1) per tile using a neural network
4. Rescale to match national totals

For the prediction features currently we use OpenStreetMap and population. More features to explore, including: EO foundation model embeddings, building height data, and feature engineering to create derived features, e.g. distance to highways.

The model uses a wide spatial context window so predictions consider neighboring features (e.g. airports likely to have hotels nearby, etc.).

Training Data

Training data are from the Norwegian business register (processed & geocoded in ActMap).

1km grid data with employment per sector is also available for Japan. This data was withheld from training and used only for later validation.

Gridded GDP (resolution)

Dot size=employment, color=sector

High resolution GDP modelling for climate risk assessments with an application to coastal flooding. Barre, F.I; Bouman, E.A.; Simpson, M.J.R.; Borck, H.S.; Hertwich, E.; Moran, D. 2025, Environ. Res. Lett. 20 094045 10.1088/1748-9326/adf867

How accurate is the model?

Step 1: Predict total employment per tile

Current state of the art: Using gridded GDP to estimate employment per tile:

log1p R² = 0.84.

Adding OSM features to better predict total employment, using XGBoost

log1p R² = 0.97.

How accurate is the model?

Step 2: Predict employment composition per tile

Kullback-Leibler divergence (KL) is a measure of how well the system predicts true employment composition for a tile. Low scores are better: a tile scores zero when the model can perfectly predict the true composition.

In this histogram we see the distribution of KL scores, comparing the gridded GDP approach (grey bars) to the model (blue bars).

The distribution of model KL scores (blue bars) closer to 0 indicates the model makes better predictions than the gridded GDP approach.

How accurate is the model?

More important tiles are predicted more accurately.

Grid GDP predictive performance per sector, in Norway

Model predictive performance per sector, in Norway

Transfer learning in Japan

Japan also provides employment per sector at 1km resolution. These were not used for training but can be used to evaluate transfer learning.

A leftward shift in KL scores is also seen for the predictions for Japan. This indicates the neural network has learned generalizable patterns.

A key limitation is the relatively small training dataset (~10,000 tiles). Using a finer grid mesh would not change the number of observations per variable, the crucial factor. More training data from other countries, even if only partial, will help.

Neural Network vs. Baseline

We can also see if the model predicts employment composition better than the gridded GDP approach commonly used today.

Left: improvement in log1p R² over gridded GDP (larger is better).

Right: improvement in RMSE over gridded GDP (smaller is better).

Filled arrows show data for Norway; hollow arrows show data for Japan.

Next Steps