Mapping Gross Domestic Product Distribution at 1 km Resolution across Thailand Using the Random Forest Area-to-Area Regression Kriging Model

Yan Jin, Yong Ge, Haoyu Fan, Zeshuo Li, Yaojie LiuISPRS Int. J. Geo Inf 2023 · DOI

TL;DR

This paper introduces a novel hybrid model (RFATARK) that combines Random Forest and Area-to-Area Kriging to map Gross Domestic Product (GDP) at a high 1 km resolution across Thailand. This approach significantly improves the accuracy and precision of GDP distribution mapping, offering a more detailed understanding of regional economic disparities.

What Problem Does This Paper Address?

Accurately mapping the spatial distribution of Gross Domestic Product (GDP) is crucial for understanding regional economic dynamics, industrial distribution, and urbanization. Traditional statistical GDP data, aggregated at administrative unit levels, often obscure fine-scale spatial differences. While nighttime light (NTL) data has been a popular proxy for economic activity, it has limitations, especially when trying to differentiate between sectors or capture nuanced economic patterns. Existing downscaling methods often struggle with issues like the modifiable areal unit problem (MAUP) and may overlook the distinct characteristics of different economic sectors. This paper aims to address these limitations by developing a more accurate and detailed gridded GDP map, specifically by differentiating between agricultural and non-agricultural GDP and employing a robust hybrid downscaling methodology.

Methods

The core of this research is the Random Forest Area-to-Area Regression Kriging (RFATARK) model. This hybrid approach combines two powerful techniques:

  • Random Forest (RF): Used for the "trend module," RF is a machine learning algorithm that excels at capturing complex, non-linear relationships between various predictor variables and GDP. It's known for its high predictive accuracy and ability to handle large datasets.
  • Area-to-Area Kriging (ATAK): This geostatistical technique is employed for the "residual module." ATAK is crucial for redistributing the errors (residuals) predicted by the RF model from coarser administrative units to finer grid cells (1 km resolution). This addresses the MAUP by ensuring that the fine-scale predictions are consistent with the coarser statistical data.

The study also explored different combinations of auxiliary variables, including unmerged/merged and classified/unclassified treatments of OpenStreetMap (OSM) data, to assess their impact on prediction accuracy.

Datasets

The study utilized a comprehensive set of datasets, categorized as follows:

  • Remote Sensing Data:
    • Nighttime Lights (NTL): NPP/VIIRS annual composite data (15 arc-seconds resolution).
  • Vegetation Index: NDVI, EVI, NIRv derived from MODIS (250 m resolution).
  • Land Surface Temperature (LST): MODIS LST data (1000 m resolution).
  • Land Use/Land Cover (LULC): FROM-GLC data (30 m resolution).
  • Geographic Information Data:
    • Digital Elevation Model (DEM): ASTER/GDEM data (30 m resolution).
  • Boundary Information: Provincial boundaries.
  • Road Network, Water Bodies, and Points of Interest (POIs): Sourced from OpenStreetMap (OSM).
  • Socioeconomic Statistical Data:
    • GDP and Census Data: Provincial-level data from the Thailand Economic and Social Development Council.
  • Gridded GDP Data: Existing gridded GDP product (Kummu et al., 2018).
  • Population Data: Worldpop gridded population counts (100 m resolution).

No new benchmark datasets were introduced; rather, existing publicly available datasets were synthesized and processed.

Key Insights and Ideas

  • Sectoral GDP Mapping is Crucial: Directly modeling total GDP can be less accurate than modeling agricultural and non-agricultural GDP separately, as their drivers and spatial patterns differ significantly.
  • Hybrid Models Outperform: Combining machine learning (RF) for trend prediction with geostatistical methods (ATAK) for residual redistribution significantly improves GDP downscaling accuracy, effectively addressing MAUP.
  • Auxiliary Variable Selection Matters: While more variables can be beneficial, the way they are processed and integrated (e.g., classification and weighted merging of OSM data) is more critical than sheer quantity. The "Merged and Classified" (MC) group of variables consistently yielded the best results.
  • Residual Redistribution is Key: Models incorporating residual redistribution (like RFATARK) showed substantially lower RMSE and MAE compared to those without, highlighting the importance of accounting for model errors at finer scales.
  • NTL Alone is Insufficient: While NTL is a useful proxy, integrating it with a diverse range of geoscientific and socioeconomic data provides a much richer and more accurate picture of economic activity.

Key Results

  • The proposed RFATARK model achieved the highest accuracy among all tested methods, with an average RMSE of 699.348 and an average MAE of 256.458 (in 10⁷ USD) at the province level.
  • The RFATARK model, specifically using the MC group of auxiliary variables, achieved the highest accuracy with an RMSE of 402.082, MAE of 174.195, and R² of 0.998.
  • The RFATARK model demonstrated superior performance compared to other models like SVR, MLR, SVATARK, and MLATARK, particularly in capturing GDP distribution in economically developed regions.
  • The spatialized GDP maps provided a significantly finer-grained perspective than provincial-level statistical maps, enabling a more detailed examination of economic activities.

Limitations and Future Work

  • Data Granularity: The study was limited by the availability of detailed statistical GDP data at finer administrative levels (e.g., county or street level). Using provincial-level data for modeling and correction introduced scale effects that could impact accuracy. Future work should aim to obtain statistical data at multiple scales.
  • Pixel-Based Truncation: For socio-economic variables, the 1 km pixel-based partitioning might inadvertently truncate production factors, potentially affecting accuracy. Future research could explore using analysis units with greater spatial coherence.
  • Model Interpretability: While RFATARK is highly accurate, like many machine learning models, its "black box" nature can limit interpretability.
  • Future Model Integration: The authors suggest that other advanced machine learning models, such as XGBoost, could be integrated with the ATAK residual prediction method in future applications, following similar steps to the RFATARK approach.
  • Auxiliary Variable Refinement: The study noted that while the MC group of variables performed best, further exploration of auxiliary variables, potentially including those excluded due to data constraints, could further enhance predictive accuracy.

Summary updated

BibTeX
@article{jin2023mapping,
  author = {Yan Jin and Yong Ge and Haoyu Fan and Zeshuo Li and Yaojie Liu and Yan Jia},
  title = {{Mapping Gross Domestic Product Distribution at 1 km Resolution across Thailand Using the Random Forest Area-to-Area Regression Kriging Model}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2023},
  doi = {10.3390/ijgi12120481},
}
Figure