Rooftop and structures semantic segmentation

Led an end-to-end computer vision project that segments rooftops and new constructions from aerial imagery, temporal data and LiDAR, for an insurance client.

Role: led the data science contribution at Sparkgeo, from research and design through client communication and results. Client details are limited by an NDA.

The problem

An insurance-industry client needed roof and structure characteristics extracted from high-resolution aerial imagery at scale. The targets were rooftops and new constructions, including additions to existing buildings. Because new construction is a change over time, the work also used temporal data alongside single-date imagery.

What I did

I owned the project end to end:

• Research and approach: reviewed the options, chose the modelling approach, and scoped the work as issues so the team and client could follow progress.
• Model: trained and deployed an R-CNN instance-segmentation model to detect and segment new constructions, using imagery and temporal data.
• Multi-modal check: added a LiDAR-based roof-segment comparison that cross-checks and augments the image model with elevation information.
• Client communication: ran the client conversations, wrote the slides, and reported results.
• Handoff: the output feeds the client’s internal models.

Result

F1 above 0.77 on this specialized task, delivered into the client’s internal models.