What is ground truthing, and why does a custom classifier still need it?

Ground truthing is the step where you take whatever a classifier called a feature, whether that's a stand of invasive phragmites, a culvert, or a stockpile, and confirm against something more reliable than the imagery that it actually is what the model says it is. Usually that means a field visit with a GPS unit and a tablet. Sometimes it means cross-checking against a higher-resolution image, a drone pass, or an as-built drawing you already have on file. Either way, it's the difference between a layer you can hand to a regulator and a layer you're still guessing about.
Off-the-shelf classifiers need this check too, but a custom one needs it more, because the training set is yours and nobody else has run it through a hundred projects first. Skipping it under deadline pressure is how a layer with silent false positives ends up in a client deliverable.
Why a model trained on a handful of examples still drifts
A custom feature class, by definition, has no published accuracy figures behind it. Nobody has run a benchmark on "abandoned well pads in this specific basin" or "retention ponds with algal bloom at this time of year." You built the training set yourself, probably from ten or twenty labelled instances, and the model generalized from there across your whole project area.
That generalization is exactly where things go sideways. A shadow pattern that looked like your target in the training tiles might get picked up a hundred times somewhere else in the imagery. A real instance that's partially occluded by tree canopy might get missed entirely. None of this shows up by eyeballing the output layer in QGIS. It shows up when someone drives to a point the model flagged and finds nothing there, or stands next to a feature the model never caught.
Any model asked to learn a category from a handful of examples instead of a labeled dataset with thousands of instances behind it will drift this way, no matter which vendor built the workflow. The tradeoff that makes few-shot classification useful, standing up a feature class without a six-month labeling project, is the same tradeoff that means you can't treat the output as final until someone's checked it against reality.
What a verification pass needs to cover
You don't need a full statistical accuracy assessment with a stratified random sample and a confusion matrix, not for most project deliverables. But you do need enough of a check to know where the layer can be trusted and where it can't. A few things worth doing before the layer goes into a report or a client deliverable:
Pull a subset of the flagged polygons, weighted toward the edges of your study area and anywhere the imagery quality dropped (haze, off-nadir angle, seasonal canopy). Models tend to perform best near their training examples and get shakier further out.
Spot-check a handful of places the classifier didn't flag, particularly ones that look visually similar to your target class. A model with a false-negative problem won't announce it; you have to go looking for what it missed.
Note the context of every error you find, not just that it was wrong. A false positive triggered by bare soil reads differently than one triggered by a parking lot, and that distinction tells you whether to retrain with a few more examples or just flag a known limitation in the metadata.
Document the sample size and what you checked, even informally. If this layer ever gets questioned in a regulatory submission, "we spot-checked 15% of flagged instances against field GPS points" is a defensible sentence. "We ran it and it looked right" is not.
None of this takes a dedicated accuracy team. It takes a field day or two, or in some cases just a careful desk review against higher-res reference imagery, scoped to the size of the dataset and what the deliverable is riding on.
A workflow that lets you label a handful of examples and get a feature class back for the exact thing you need mapped, like the one Few-Shot Mapping is building early access around, still hands you a layer that needs this same look before it goes anywhere official. That's simply what "custom" means when there's no existing benchmark to lean on.
If you've got a feature class that doesn't exist in any standard classifier and a handful of examples to work from, that's the exact gap Few-Shot Mapping's early access is built for.