AI Roof Damage Detection from Aerial Imagery: NavGen Case Study

Most roofing leads are cold. NavGen flips that model by using aerial imagery and computer vision to spot damaged roofs before anyone knocks on a door. This case study covers the full production system: how the training set was labelled, the three-stage U-Net and CNN pipeline behind the detections, the AI infrastructure inference architecture that scores a property in under two seconds, and the commercial outcome for roofing companies that replaced blind canvassing with pre-qualified, damage-verified lead lists.

The Problem

Roofing companies spend enormous sums on lead generation. The traditional approach — door-to-door canvassing, purchased lead lists, digital ads — generates leads with conversion rates typically below 5%. Most homeowners they contact don't actually need roof work. The waste is staggering.

NavGen's founder came to us with a hypothesis: if you could identify which roofs actually have damage before making contact, you'd dramatically improve conversion rates and ROI. The question was whether computer vision could accurately detect roof damage from aerial imagery at a scale that makes business sense.

Our Approach

Phase 1: Data Collection and Labelling

The first challenge was assembling a training dataset. We partnered with aerial imagery providers to access high-resolution overhead photographs covering millions of properties. Our labelling team, working with experienced roofing inspectors, annotated thousands of images identifying:

• Hail damage (pitting, denting, granule loss)
• Wind damage (lifted or missing shingles)
• Moss and algae growth
• Structural sagging or deterioration
• Age-related wear patterns
• Storm debris damage

Getting the labelling right was critical. We went through three rounds of labelling guidelines before the inter-annotator agreement reached acceptable levels. This is the unglamorous part of AI that determines everything downstream.

Phase 2: Model Architecture

We built a multi-stage pipeline rather than a single monolithic model:

• Stage 1 — Roof segmentation: Identifying and isolating the roof area from the rest of the aerial image using a U-Net-based segmentation model.
• Stage 2 — Damage classification: A fine-tuned convolutional neural network classifying the segmented roof region into damage categories and severity levels.
• Stage 3 — Confidence scoring: An ensemble approach combining multiple model predictions to generate a confidence score, reducing false positives.

The multi-stage approach allowed us to optimise each component independently and dramatically improved accuracy over a single end-to-end model.

Phase 3: Scale and Infrastructure

Processing millions of property images requires serious infrastructure. We built on AWS with:

• GPU-accelerated inference using EC2 instances with batch processing pipelines
• S3 for imagery storage with intelligent tiering for cost optimisation
• PostgreSQL for property data and analysis results
• Redis for caching territory scan results
• Docker containers orchestrated with ECS for consistent deployments

We optimised the inference pipeline to process a single property in under 2 seconds, making territory-wide scans of thousands of properties feasible within hours.

The Results

The system now enables roofing companies to:

• Scan entire territories and identify properties with visible roof damage
• Receive tiered lead lists (Essential, Intelligence, and Advanced) based on confidence scores
• Access verified homeowner contact information for pre-qualified leads
• Track campaign performance through an ROI analytics dashboard

The key outcome: roofing companies using NavGen close significantly more deals because they're only contacting homeowners who actually need roof work. The lead quality improvement eliminated wasted outreach and dramatically improved sales team efficiency.

What changed commercially

• Canvassing effort moves to verified damage. Instead of knocking every street in a postcode, crews work a ranked list where every address already has a damage signal attached, so the same headcount covers a far smaller, far warmer set of doors.
• Storm response becomes same-week. After a hail or wind event, a territory can be re-scanned and re-scored in hours rather than waiting on manual inspection scheduling — the difference between arriving first and arriving after three competitors.
• Lead spend is reallocated, not increased. Budget previously burned on purchased lists and broad digital ads shifts into outreach on addresses with a confirmed damage classification.
• Estimators stop travelling to dead calls. Confidence tiers let sales leaders decide which properties justify a site visit and which get a lower-cost touch first.

Because the pipeline is tuned for precision rather than recall, the numbers that matter to an owner are contact-to-inspection and inspection-to-contract rates, not raw lead volume. A smaller list of verified-damage addresses beats a larger list of cold ones on every cost-per-signed-job measure.

Engineering results worth quoting

• Under 2 seconds of end-to-end inference per property, including segmentation, classification and confidence scoring.
• Territory-wide scans of thousands of properties completed within hours on batched GPU inference.
• Three-stage architecture (U-Net segmentation → fine-tuned CNN classifier → ensemble confidence scoring) that materially outperformed a single end-to-end model.
• Six categories of damage detected: hail, wind, moss and algae, structural sagging, age-related wear and storm debris.
• Continuous retraining loop feeding confirmed field outcomes back into the training set, with accuracy improving steadily over the first six months in production.

Cost and timeline expectations

Teams evaluating a comparable computer vision build usually want two numbers before anything else: what it costs and how long it takes. A production-grade detection pipeline of this shape is a multi-month engagement, not a two-week prototype — labelling guidelines and inter-annotator agreement alone consumed several iterations before any model training was worth doing. For realistic budget bands, see our breakdown of enterprise AI development costs, and for how the work is sequenced, our guide to building an enterprise AI roadmap.

Technical Lessons Learned

• Image quality variance is the real challenge. Aerial imagery varies enormously in resolution, lighting, angle, and season. Our models needed to be robust to all of these variations, which required extensive data augmentation and careful training.
• False positives are worse than false negatives. Telling a roofing company a roof is damaged when it's not wastes their time and damages trust. We tuned aggressively for precision, accepting lower recall in exchange for higher-quality leads.
• Continuous learning is essential. As we processed more imagery, we fed confirmed results back into the training pipeline. The model's accuracy improved steadily over the first six months of operation.
• Domain expert involvement is non-negotiable. Having roofing inspectors validate our training data and review edge cases was the difference between a demo and a production system.

Broader Applications

The same computer vision pipeline architecture we built for NavGen applies to numerous inspection and assessment use cases: solar panel condition monitoring, property insurance assessment, agricultural crop analysis, infrastructure inspection, and more. The pattern of aerial imagery + segmentation + classification + business workflow is broadly applicable across industries.

Frequently Asked Questions

How accurate is AI roof damage detection from aerial imagery?

Accuracy depends far more on labelling discipline and imagery quality than on model choice. In this build the pipeline was deliberately tuned for precision over recall: a flagged roof should almost always have real damage, even if some damaged roofs are missed. That trade-off is the right one commercially, because a false positive sends an estimator on a wasted trip while a false negative simply leaves a lead on the table.

What damage types can be detected from overhead images?

Six categories in production: hail damage (pitting, denting, granule loss), wind damage (lifted or missing shingles), moss and algae growth, structural sagging, age-related wear patterns, and storm debris damage. Anything requiring an under-deck or attic view — decking rot, insulation saturation — still needs a physical inspection.

How long does a territory scan take?

Under two seconds of end-to-end inference per property, so a territory of several thousand addresses completes within hours on batched GPU inference. After a hail or wind event the same territory can be re-scored the same week.

What does a system like this cost to build?

It is a multi-month engagement rather than a prototype sprint, with a meaningful share of the budget going to labelling guidelines and inter-annotator agreement before model training starts. See our enterprise AI development cost breakdown for realistic bands and our AI implementation roadmap for how the phases sequence.

Can the same approach work outside roofing?

Yes — segmentation plus classification plus a business workflow transfers to solar array inspection, insurance property assessment, crop health analysis and infrastructure surveys. If you are scoping something similar, our computer vision engineering page covers how we run these builds, and our case studies show comparable production systems.