AeroScan

Project: AeroScan

Inspiration

Large industrial and commercial buildings represent a massive opportunity for efficiency improvements, but identifying them at scale has always been a manual, time-consuming process. AeroScan was born from the need to automate the discovery of high-impact real estate. By combining satellite intelligence with custom machine learning, we’ve created a pipeline that can scan vast territories to find the "hidden giants" of the industrial world and assess their infrastructure in seconds.

What it Does

AeroScan is an automated geospatial intelligence platform. It scans geographical regions to identify large-scale buildings (100k+ sq. ft.) and uses computer vision to detect specific critical infrastructure—specifically cooling towers. By identifying these towers, AeroScan provides immediate insights into a building's HVAC complexity, water usage, and potential for energy optimization.

How We Built It

Our focus for this build was the United States, leveraging a robust stack of geospatial and AI tools:

Geospatial Data: We utilized the Overture Maps dataset and the Google Earth Engine to access and analyze massive amounts of building footprint data.
Scale: The analysis of 100k+ sq. ft. structures was made possible by querying these massive Microsoft datasets, allowing us to filter for high-priority targets.
Custom ML Model: We developed and trained a YOLOv5 segmentation model specifically to detect whether a building has a cooling tower on its roof or nearby.
AI Insights: We integrated Gemini to provide contextual insights and narrative data based on the detected infrastructure.
Reporting: We built an automated PDF generation engine to turn raw satellite data and ML detections into professional, actionable reports.
Calculations: The system performs automated area and footprint calculations to verify building size and cooling capacity requirements.

Challenges We Ran Into

Usable Data Sourcing: Finding high-quality, labeled datasets for cooling towers was significantly difficult. We had to source and curate specific imagery to ensure the model could distinguish a cooling tower from other rooftop equipment like fans or skylights.
Model Training: Fine-tuning the ML model to maintain high accuracy across different geographical terrains and lighting conditions required extensive iteration and data cleaning.

Accomplishments That We're Proud Of

Global Scalability: We successfully built a pipeline capable of detecting buildings across the world with footprints exceeding 100,000 sq. ft.
Automated Infrastructure Detection: Achieving high-precision detection of cooling towers, turning a visual search task into a data-driven one.
Seamless Reporting: Successfully bridging the gap between raw ML output and a user-ready PDF report.

What We Learned

ML Pipeline Management: We gained deep experience in creating ML models from scratch, specifically the importance of data cleaning and augmentation in achieving a deployable model.
Complex Document Logic: We learned the intricacies of automated PDF generation, ensuring that spatial data and AI insights are formatted correctly for stakeholders.
Big Data Handling: Working with the Overture dataset taught us how to efficiently query and filter millions of data points to find specific architectural criteria.

What's Next for AeroScan

Financial Integration: We plan to integrate deeper financial calculations, such as estimated retrofitting costs, energy savings ROI, and property tax implications.
Expanded Infrastructure Detection: Training the model to identify other assets like solar arrays, substations, and loading docks.
Live Monitoring: Moving from static analysis to time-series monitoring to track industrial expansion over time.