Solar Potential Assessment using Multi-Class Buildings Segmentation from Aerial Images

Hasan Nasrallah, Abed Ellatif Samhat, Ghaleb Faour, Yilei Shi, Ali J. Ghandour

Abstract

Lebanon faces a severe energy crisis, yet lacks high-resolution data for solar energy planning. This paper presents a comprehensive building footprint and solar rooftop potential map for the country. By utilizing deep learning-based instance segmentation and a morphology-aware panel placement algorithm, the study estimates Lebanon's total solar potential at 28.1 TWh/year—double its 2019 consumption.

Introduction

Since 2019, Lebanon has grappled with frequent blackouts and rising energy costs. While solar energy offers a sustainable solution, effective policy requires accurate national-level data. Traditional "area-based" assessments often overestimate potential by ignoring roof morphology—the specific shape and orientation that dictate how many rectangular solar panels can actually fit. This research bridges this gap by using high-resolution satellite imagery to identify individual buildings and simulate realistic panel layouts.

Methodology

The researchers developed a two-stage pipeline to transform raw satellite pixels into energy estimates.

Instance Segmentation

A UNet-like architecture was employed for semantic segmentation. To address the challenge of touching buildings in dense urban areas, the model introduced a third class—building boundaries—and utilized a watershed post-processing step. After evaluating various backbones, EfficientNet-B3 was selected for its optimal balance of accuracy (F-score of 84.3%) and low variance. The model was trained on the newly created Tyre Urban Map (TUM) dataset, which includes 10,000 annotated building objects across diverse Lebanese landscapes.

Morphology-Aware Panel Placement

Moving beyond simple area calculations, the team implemented a greedy algorithm to "pack" standard 1x1.98m PV panels onto extracted footprints. The algorithm identifies the roof's longest axis and aligns panels accordingly to maximize coverage while ensuring no overlap with building edges.

The morphology-aware panel placement algorithm maximizes solar panel density by aligning with the roof's main axis and removing overlaps.

Results and Key Insights

The study provides the first detailed breakdown of solar potential across Lebanon's 26 districts.

Segmentation Performance

The segmentation model demonstrated high precision in identifying complex building shapes across urban and rural settings. The inclusion of the boundary class was critical, improving the F-score by over 10% by preventing multiple houses from merging into single "blobs."

Comparison of original satellite imagery, ground truth annotations, predicted masks, and individual building instances.

National Solar Potential

Lebanon’s total solar rooftop potential is estimated at 28.1 TWh/year (at a 50% utilization factor). Key findings include:

  • The "5% Rule": An average Lebanese rooftop can meet the yearly electricity needs of a single-family residence using only 5% of its surface area.
  • Density vs. Area: While the Baalbeck district has the highest total potential due to its size, Beirut has the highest average potential per building due to its vertical urban density.
  • Energy Security: Utilizing half of a typical apartment building's roof could provide energy security for up to eight households.

Limitations and Future Work

While groundbreaking, the current model relies on several simplifying assumptions. It assumes all roofs are flat and does not account for shadows from neighboring structures or trees. Additionally, small rooftop obstacles like water tanks are not yet detected. Future research aims to incorporate 3D reconstruction from stereo satellite imagery or LiDAR to refine these energy models further.

Summary updated

BibTeX
@article{nasrallah2022solar,
  author = {Hasan Nasrallah and Abed Ellatif Samhat and Ghaleb Faour and Yilei Shi and Ali J. Ghandour},
  title = {{Solar Potential Assessment using Multi-Class Buildings Segmentation from Aerial Images}},
  journal = {IEEE JSTARS},
  year = {2022},
  doi = {10.1109/jstars.2022.3181446},
}
Figure