Custom AI Models for Drone Inspection: From Site Data to a Deployable Detector

When to use a custom model instead of a preloaded detector

A preloaded model is the fastest route when its classes and operating conditions match the problem. Custom training becomes useful for proprietary components, unusual safety objects, site-specific vegetation, specialized debris or viewpoints poorly represented in general datasets.

Do not train a model to compensate for an unobservable target. If the camera cannot resolve a small crack at the required stand-off distance, more labels will not restore missing detail. First confirm sensor, lens, route and lighting feasibility with sample capture.

• The target class is not available in an existing model.
• Generic results fail because the aerial viewpoint or thermal appearance is unusual.
• False alerts come from repeatable site-specific backgrounds.
• A proprietary object or condition creates competitive or operational value.
• The organization can provide representative data and expert labels.

Define the class around an observable visual rule

Asset experts and model developers should write the annotation guide together. Include positive examples, exclusions, minimum visible portion, occlusion rules and what to do with uncertain cases. Separate classes only when the distinction is visible and operationally useful.

Tie every class to a response. If two labels produce the same action, one class may be enough. If a condition requires measurement or context outside the image, the model should flag a review region rather than claim a diagnosis.

Build a dataset that represents missions, not isolated pictures

Drone imagery varies with altitude, angle, focal length, motion, compression, season and background. Thermal appearance changes with weather and time. Sample across these dimensions and include difficult negatives. Split data by site or flight to prevent nearly identical video frames from leaking into the test set.

Track consent, ownership and retention for training data. Label quality should be audited with inter-annotator review and class-frequency reports. Preserve a frozen, unseen test set for final acceptance and another small field set for the target hardware.

1. Capture representative positive and negative missions.
2. Create and train annotators on the class guide.
3. Audit labels, duplicates and class balance.
4. Split by mission, site or time.
5. Version the dataset and protect the frozen test set.

Choose on-board or Brain-Box deployment early

The target hardware shapes model architecture, input size and optimization. DJI’s Matrice 4 on-board pathway uses supported models and quantization before device distribution. It suits compatible, low-latency aircraft-side detection. The converted artifact must be revalidated.

Brain-Box supports on-premises models across drone, dock, video and robot workflows. It may be preferable for several streams, larger storage, selected bring-your-own-model formats or integration. Train toward the production constraint instead of compressing an oversized model at the end.

Monitor the model after deployment

Field data changes. New paint, vegetation, weather, construction phases, sensor settings or route height can reduce performance. Record reviewer outcomes and sample normal footage as well as alerts; otherwise the team sees false positives but not missed targets.

Use a controlled retraining cycle. Curate new examples, review labels, compare the candidate with the approved model on the same frozen test sets and re-run hardware acceptance. Spectro AI offers on-premises algorithms, DJI on-board algorithms and custom training for defined use cases.

• Model, dataset and threshold version attached to each release.
• Precision and recall by class, site and target size.
• Review of false alerts and sampled missed cases.
• Change approval and rollback for operational models.
• Revalidation after camera, firmware, route or environment changes.

Frequently asked questions