Summary of the Two-Stage Process for Global Rooftop Area Estimation

Stage 1: Top-down approach using deep learning

• Goal: Quantify rooftop area in selected representative regions.
• Method:
  • Used SegFormer, a cutting-edge Vision Transformer-based deep learning model.
  • Pretrained on publicly available building identification datasets (~2,500 km² across diverse regions, spatial resolutions 0.1 m to 3 m).
  • Fine-tuned using high-resolution Google Earth imagery (~1.2 m resolution), which is cloud-free and harmonized from multiple satellite/airborne platforms.
• Sample selection:
  • 1,724 cities were chosen based on geographical and environmental representativeness using a K-means clustering of natural and human environmental features and a spatial sampling scheme optimized by simulated annealing.
• Output: Quantified rooftop area per city/region.

Stage 2: Bottom-up approach using random forest ensembles

• Goal: Extrapolate rooftop area to global scale.
• Method:
  • Collected multi-source geospatial variables at 1 km² grid scale: built-up proportion, night-time light intensity, road length, population, tree cover, terrain elevation & slope, geographic coordinates, etc.
  • Aggregated rooftop areas from top-down stage to these grid cells.
  • Developed regression and classification random forest ensembles to model nonlinear relationships between geospatial variables and rooftop area.
  • Excluded grids with no high-resolution imagery; total 8.5 million grid samples used.
• Postprocessing: Used a water map to allocate zero rooftop area to grids fully covered by water.

Model Evaluation

Evaluating the deep learning model (top-down):

• Created a global representative test set: 386 one-km² plots across countries; manually labelled rooftop areas.
• 2,951 image patches processed for validation.
• Performance:
  • True positive rate (rooftop correctly identified): 76%
  • False positive rate (non-rooftop misclassified as rooftop): 2.7%
  • Compared favorably with state-of-the-art building footprint datasets (MBF: 61.6% TPR, 4% FPR; GBF: 66.5% TPR, 3.8% FPR).
• Strong correlation between predicted and actual rooftop area:
  • r² = 0.93
  • Slope = 1.04
• Performance varied by macroregion:
  • Economically developed regions: r² > 0.95
  • Less developed regions: r² ~ 0.9

Evaluating the random forest model (bottom-up):

• Selected 16,000 independent grid samples (800 per macroregion).
• Quantified rooftop area using high-resolution imagery and compared to random forest predictions.
• Performance:
  • Overall r² = 0.89, slope = 0.87 (slight underestimation)
• Lower accuracy for some regions:
  • Pacific Islands: r² = 0.61, bias error = 55%
  • Western Asia: r² = 0.67, bias error = 24%
• Residual analysis showed greater errors in grids with larger rooftop areas.
• Residuals roughly normally distributed, mostly within ±5,000 m².

Important Notes

• The bias error formula normalizes the absolute error by the observed rooftop area sum:

[
\text{bias} = \frac{\left|\sum{N} \left(Y{\text{obs}} – Y{\text{pred}}\right)\right|}{\left|\sum{N} Y_{\text{obs}}\right|}
]

where (Y{\text{obs}}) is observed rooftop area and (Y{\text{pred}}) predicted rooftop area.

Summary conclusion

• The integration of a deep learning model for building rooftop detection with random forest regression using multiple geospatial predictors enables accurate estimation of rooftop areas globally.
• While the model performs best in well-represented, economically developed regions, some limitations exist for under-sampled regions such as small island states and parts of Asia.
• Overall, the two-stage framework provides a scalable, data-driven method for global rooftop area estimation which can support various applications including urban planning, renewable energy potential assessment, and sustainability efforts.

If you want, I can help with a more detailed explanation of any stage, discussion about the methodology, or assist in interpreting the results further!