C-DRONE GUIDE · 31 JULY 2026
Artificial intelligence and automatic defect detection on drone images: what it actually changes
"Our AI detects 100% of defects" — the promise shows up in almost every drone provider's sales brochure, and it deserves a closer look. Artificial intelligence has genuinely changed how a set of several thousand images of solar panels, metres of façade or kilometres of power line gets analysed: it flags candidate defects that a human eye would take hours to review, at volumes no engineering firm would process manually at the same price. But what it actually does, what it still does not do, and what it really changes on a business's invoice deserve to be clarified before picking a provider on that sales pitch alone.
Published on 31 July 2026, reviewed on 2 August 2026 — regulations in force as of August 2026.
What AI actually does on a set of drone images
The principle is the same whatever the structure being inspected: after the flight, the raw images (photos, orthophoto mosaic, thermal imagery) are screened by a model trained to recognise a specific visual pattern — a crack in concrete or render, a hot spot on a photovoltaic cell, a trace of corrosion on sheet metal. The model does not issue a diagnosis: it flags candidate zones, with geolocated coordinates and often a confidence score, leaving a qualified operator to validate each one individually. That is an essential difference from what the term "automatic detection" suggests: automatic refers to sorting a volume of images, not to the technical conclusion.
This approach has proven itself on well-documented benchmark cases: a study by Cha, Choi and Büyüköztürk published in 2017 in Computer-Aided Civil and Infrastructure Engineering showed that a convolutional neural network trained on around 40,000 images of cracked concrete reached close to 98% accuracy at spotting cracks in new images (see the study on Google Scholar). That work, one of the first to demonstrate the approach's viability on a construction material, remains the reference cited by nearly every commercial tool that has since drawn on it — on concrete, but also, with retrained models, on the masonry, render or steel of the structures we inspect: bridges, façades before renovation or power lines.
Where the gain is real today, where it stays limited
The real benefit shows up on repetitive, high-volume assets: a solar farm with several thousand panels, several kilometres of power line, a logistics warehouse façade of several thousand square metres. Reviewing images one by one at that scale takes a human eye days, and attention inevitably drops off toward the end of the batch; a model processes the same volume in a few hours, without fatigue and with a constant detection threshold from one image to the next. On photovoltaics specifically, a study by Pierdicca, Malinverni, Piccinini, Paolanti, Felicetti and Zingaretti published in 2018 in the ISPRS archives validated a convolutional neural network trained to automatically spot damaged cells on aerial infrared images of solar plants (see the study on Google Scholar) — exactly the kind of mass sorting we now apply in thermal imaging during a solar farm campaign.
The limits come down to three points. First, a model trained on a given material or panel brand transfers poorly, without retraining, to a very different structure: cut stone, metal cladding or a different panel manufacturer produce more false positives and false negatives than the sales sheet promises. Second, no current model delivers an engineer's verdict: it flags a visual anomaly, not its structural severity or the urgency of intervention, which remain a qualified professional's call. Finally, blindly trusting the algorithm's output without a human second read risks missing a real defect wrongly classed as irrelevant — a risk worth knowing before signing, alongside the question of the provider's liability insurance.
What it changes in the service and on the invoice
In practice, AI slots in between the shoot and the report: alongside the usual images or orthophoto, the provider delivers a geolocated map of candidate defects and an annotated report where each detected anomaly is confirmed or dismissed by a trained operator — senior pilot, certified thermographer or structural engineer depending on the asset. The main gain for a professional client is turnaround time: analysing a large site drops from several weeks to a few days, which matters for a bridge's statutory ten-year inspection or a solar farm's yearly maintenance campaign. The flight itself — preparation, authorisations, piloting — is still billed on the usual basis detailed in our how much a drone service costs guide; AI-assisted analysis is generally added as an extra package tied to the volume of images.
Ranges observed in 2026 (excl. VAT), on top of a standard capture mission:
- AI analysis of a set of solar panels (up to 5,000 panels, geolocated defective-cell report): €300 to €900.
- Automatic crack detection on a façade or bridge deck (annotated report, human validation included): €400 to €1,200 depending on area.
- Automatic anomaly classification on a power line (per 10 km stretch): €200 to €600.
- Large industrial site or solar farm, custom processing: quoted.
The right questions to ask before signing
Four questions are enough to tell a marketing pitch apart from a real capability. What dataset was the model trained on, and does it match the material and structure type concerned? What false-positive and false-negative rate does the provider disclose, rather than a flattering accuracy percentage pulled out of context? Who validates each detection before the report is issued, and does that person hold the qualification required for the structure being inspected? Finally, in what format is the deliverable handed over — a usable GIS map, a PDF report, an export into an asset-management tool — and does it fit the tools already used in-house?
These questions echo the ones worth asking about any surveying and photogrammetry or façade inspection service: a serious provider presents AI as a sorting tool that speeds up and standardises analysis, not as a substitute for the human expertise still committed on the final report. Request a quote stating the structure type, the estimated volume of images and whether AI-assisted analysis interests you: a detailed answer on these four points usually says more than any sales brochure.