Inspect IQ

AI Drone Inspection: How Machine Learning Reads Inspection Imagery and Where Engineers Decide

By Kush Agarwal, President and CEO, YelloSKYE. Published 2026-10-29.

Executive Summary

• The Question: What does AI actually do with drone inspection imagery, and which inspection decisions can it be trusted with?
• The Answer: AI sorts large image sets: it classifies frames, boxes likely defects, outlines areas pixel by pixel and highlights what changed since the last survey. That turns thousands of images into a shortlist. Whether each flag is real, how serious it is and what to do about it remains an engineering decision.
• The YelloSKYE Angle: Inspect IQ uses AI to find and geotag candidate defects, then hands them to people for review, and records each confirmed finding with its location, severity and history in the DEX platform.

AI drone inspection is the use of machine learning to analyse the images a drone captures of an asset, so that likely defects are found, outlined, located and compared with earlier surveys automatically. It does not inspect the asset in the engineering sense. It reduces thousands of photos to a shortlist of places an engineer should look at, and keeps a consistent record of what changed.

For owners of facades, roofs, solar farms, towers, bridges and industrial plant, that is a significant change in how inspection time is spent. This guide explains the main techniques, what they need from the flight, how human review fits in, how to judge whether a tool is accurate, and where the boundary between software and engineer should sit. At YelloSKYE, this analysis runs inside Inspect IQ, the solution we use for asset inspection.

Why inspection imagery needs machine help

A drone inspection is a data problem as much as a flying problem. To see small defects, the camera has to be close and the images have to overlap, so even a modest asset produces hundreds or thousands of frames. Add thermal imagery, multiple visits a year and a portfolio of sites, and manual review becomes the bottleneck. In YelloSKYE's solar farm thermal inspection use case, a single portfolio can produce more than 350,000 images.

People reviewing that volume by eye get tired, and tired reviewers miss things and judge inconsistently. A model looks at the last image with the same attention as the first. That consistency, more than speed, is the main reason to use AI in inspection.

Four things AI does with inspection images

Most inspection software combines a few standard computer vision techniques. Knowing which one is behind a feature helps you understand its output and its failure modes.

Classification

Classification is the simplest technique and often the first filter. A model sorts images or map tiles into groups, such as "defect likely" and "clear", so reviewers start with the frames that matter. On thermal surveys, it can sort anomalies by type before anyone opens them.

Object detection

Detection is what most people picture when they hear "AI inspection": a box around each suspected defect, with a label and a confidence score. It suits defects that are discrete and countable, such as a cracked insulator, a broken glazing panel or a missing fixing.

Segmentation

Segmentation goes further and labels each pixel. That makes it possible to measure area and shape, for example how much of a coating has failed or how far spalling has spread. Segmentation is also used to separate the scene into broad classes, such as structure, vegetation, water and ground, which helps with tasks like checking vegetation encroachment along a power line.

Change detection

Change detection compares surveys of the same asset over time. It is often the most useful technique for maintenance, because the question an engineer asks is rarely "is there a crack?" but "is this crack getting worse?". It depends completely on consistent capture: same route, same angles, similar light. A docked drone flying a stored route is one way to achieve that, as our drone-in-a-box guide explains; a disciplined crewed programme is another.

What the model needs from the flight

AI cannot recover detail the camera did not capture. Most disappointing results come from capture problems, not from the model.

There is also the model's own training data. A model trained mostly on one asset type, one climate or one camera will be less reliable on something different. Ask what it was trained on, and whether it has seen assets like yours. For thermal work specifically, our guide to drone thermal inspection covers the survey conditions in detail.

From flag to decision: where human review fits

An AI flag is a question, not a finding. A sound workflow moves each one through clear steps, with a person accountable at each decision point.

1. Flag. The model marks a candidate defect, with type, location and confidence.
2. Triage. A trained reviewer confirms or rejects it, merges duplicates and adds anything the model missed.
3. Assessment. An engineer judges severity and likely cause, using the asset's history and design.
4. Record. The confirmed finding is stored with its images, location, severity and date.
5. Action. A repair, monitoring interval or further investigation is assigned to an owner.
6. Recheck. The next survey shows whether the repair held or the defect has progressed.

Setting the sensitivity

Every detection model trades missed defects against false alarms. Turn the sensitivity up and it finds more real defects but buries reviewers in false flags. Turn it down and the list is cleaner but some real defects slip through. The right setting depends on the consequence of a miss. On a facade above a public street or a component carrying structural load, accept more false alarms. On low-risk cosmetic surveys, a cleaner list may be the better trade.

Where AI helps and where engineers still decide

The bottom three rows do not change with better models. A drone image shows a surface. Loose cladding fixings, hollow render, internal corrosion and section loss still need someone to touch, tap, test or open up. The facade inspection from a desk use case on our site reaches a similar conclusion: the map decides where people go, and people still go.

How to judge whether an AI inspection tool works

Accuracy claims are easy to make and hard to compare. Before relying on a tool, ask for:

• Two numbers, not one. The share of flags that were real defects (precision) and the share of real defects that were flagged (recall), measured on assets like yours.
• A blind test on your own data. Run it on a past survey where your engineers already know the defects, and compare.
• Per-defect results. A tool can be strong on hot spots and weak on fine cracks. An average hides that.
• Review workload. How many flags per hundred images reach a reviewer, and how long does triage take?
• Training data and updates. What assets and conditions the model learned from, and how it is improved when it gets things wrong.

Our broader guide to AI in construction covers similar questions for other kinds of AI tool.

How Inspect IQ handles it

Inspect IQ captures assets with high-resolution RGB and thermal sensors, so facades, roofs, towers and industrial structures can be inspected without scaffolding, rope access or a work-at-height team. AI models trained to recognise cracks, spalling, corrosion, thermal anomalies and other surface defects geotag and categorise each finding, and the findings then go to your team for review. Thermal data is combined with RGB imagery into a single report.

What the client receives is a defect database rather than a folder of photos: each confirmed finding carries its marked-up image, its coordinates, a severity grade and a recommended fix, and the database can be exported as PDF or Excel or in BIM-linked formats. Inside DEX, maintenance teams can track open issues, severity and resolution speed across a portfolio, and slide between historical captures to show when a defect first appeared.

Two use cases show this in practice. On utility-scale solar, described in the solar farm thermal inspection use case, the software picks out hot spots, failed diodes and disconnected strings across the array and gives each one a GPS position; the owner saw a 75 per cent fall in inspection costs. Our AI-powered asset inspection use case covers facades and industrial structures, where defects are mapped on a 3D model of the asset. In both, civil and geospatial engineers on YelloSKYE's team check the automated output before it is relied on.

A practical checklist for AI drone inspection

1. Define the smallest defect that matters for each asset, and set capture resolution from it.
2. Fix routes, angles and settings so surveys can be compared over time.
3. Ask for precision and recall per defect type, measured on assets like yours.
4. Run a blind test on a past survey before using any tool on live work.
5. Set sensitivity by consequence, accepting more false alarms where a miss is dangerous.
6. Name a reviewer and an engineer for every survey, and record who confirmed each finding.
7. Keep tactile and intrusive checks in the programme for what imagery cannot show.
8. Close each finding on evidence from the next survey, not on a report that the repair was done.

Explore Inspect IQ