An Empirical Study on Learning Rooftop Solar Potential from Satellite Imagery Directly Without Explicit Geometry
Organizations: Skyline High School, Ann Arbor, MI 48103
Abstract
Rooftop solar potential estimation often relies on LiDAR, digital surface models (DSMs), or pipelines that explicitly reconstruct roof geometry before estimating solar potential. We propose Direct Rooftop Solar Prediction (DRSP), an end-to-end formulation that maps satellite imagery and coarse regional sunlight information directly to application-level solar outputs. DRSP neither requires explicit geometric representations as input nor predicts roof orientation, slope, height, DSMs, or other geometric quantities as intermediate outputs, and therefore requires no geometric intermediate supervision. To enable large-scale study of this formulation, we develop a programmatic data-curation pipeline that converts per-panel information from the Google Solar API into usable-rooftop masks and spatial solar-energy targets, yielding two geographically distinct datasets containing approximately 60K rooftops from the U.S. Midwest and 87K from Texas. We demonstrate the feasibility of DRSP and systematically characterize the effects of geographic training distribution, model depth, image resolution, regional sunlight conditioning, and heatmap regression objectives. Overall, our results suggest that useful rooftop-scale solar information can be learned directly from 2D imagery without explicit geometric intermediate representations, offering a simpler path toward large-scale solar assessment. Our code and dataset are available at https://github.com/joshualee124/direct_rooftop_solar_prediction.
Figures & tables
| MIDWEST Test Set | TEXAS Test Set | |||||||
| DRSP-Net Training Data | Usable Rooftop Segmentation MIOU | Energy Heatmap Pixel-wise MAE | Max Panel Count MAE | Total Energy Generation MAE | Usable Rooftop Segmentation MIOU | Energy Heatmap Pixel-wise MAE | Max Panel Count MAE | Total Energy Generation MAE |
| BOTH | 0.67 | 0.045 | 17.45 | 8410.46 | 0.77 | 0.034 | 18.73 | 10936.67 |
| MIDWEST only | 0.66 | 0.043 | 18.30 | 8962.27 | 0.66 | 0.062 | 37.50 | 20939.09 |
| TEXAS only | 0.53 | 0.078 | 35.37 | 17447.87 | 0.76 | 0.035 | 19.71 | 11147.66 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| County/City | Dataset | Data Split | Rooftop Counts |
|---|---|---|---|
| Flint, MI | MIDWEST | Train | 1129 |
| Washtenaw, MI | MIDWEST | Train | 792 |
| Wayne, MI | MIDWEST | Train | 5500 |
| Marion, IN | MIDWEST | Train | 4555 |
| Milwaukee, WI | MIDWEST | Train | 455 |
| Cuyahoga, OH | MIDWEST | Train | 16754 |
Remote Sensing Image Understanding
Remote sensing image understanding uses AI to interpret satellite and aerial imagery for objects, land cover, change, hazards, and spatial relationships.
See all 217 papers