This study aims to develop random forest models using high-resolution data to predict the spatial distribution of local GDP. The target variable is cell GDP share, defined as the proportion of GDP attributed to a specific cell relative to its higher administrative unit (e.g., province, state, or country). Predictor variables are also expressed as shares, including cell population, vegetation, nighttime light emissions, urban areas, cropland, forest, water, snow & ice, ruggedness, national GDP per capita, and CO₂ emissions from manufacturing, heavy industry, and transportation.

Training data spans countries across North America, South America, Europe, Africa, and Asia. The model is trained to capture the relationship between predictor shares and cell GDP shares, and is applied globally to estimate local cell GDP shares.

Refer to the following and our paper for detailed information on the construction of cell GDP for training data and a series of workflow: