There is a political argument for indigenous defence technology, and it is a reasonable one. But it is not the argument we lead with, because for marine perception specifically there is a technical argument that is simply stronger — and it survives contact with a sceptical procurement officer far better.
Models inherit the sea they were trained on
A detection model's performance is bounded by the distribution of its training data. This is not a controversial claim; it is the central fact of supervised learning. What it means in practice for marine vision is more specific and more consequential than most buyers realise.
Consider what a model trained predominantly on North Sea, Baltic and Mediterranean footage has actually learned. It has seen a particular catalogue of hull forms — steel-hulled commercial traffic, leisure craft, a familiar set of fishing vessel types. It has seen a particular sea-state distribution and a particular set of atmospheric conditions. Crucially, it has learned in an environment where the overwhelming majority of vessels above a small size transmit AIS, so its optical detections have been trained and validated against a rich cooperative-target baseline.
Now deploy that model off the Konkan coast in August. Nearly every one of those learned priors is wrong at once. The hull forms are different — wooden country craft, dhows, small gillnetters with low freeboard and no radar signature worth the name. The atmosphere is different: south-west monsoon rain, and after the rains a haze that behaves nothing like North Atlantic fog. The traffic density is different, and the proportion of it that is electronically invisible is dramatically higher.
The imported system is not badly engineered. It is well engineered for a different sea. That is a harder problem to fix than bad engineering, because nothing in it looks broken.
The traffic that matters is the traffic that is not reported
Automatic Identification System carriage requirements bind cargo ships above three hundred gross tonnes on international voyages and all passenger ships. Below that threshold there is no general obligation. In Indian coastal waters that exempted category is not a rounding error — it is the majority of the vessels you are likely to encounter close inshore, and it contains almost all of the ones a collision-avoidance system needs to see.
Radar does not close the gap either. A low-freeboard wooden craft in a moderate sea returns very little, and what it does return competes with sea clutter. This is the well-documented reason cameras have moved from a nice-to-have to a complementary primary sensor in maritime situational awareness: they resolve targets inside radar's minimum range and they see the non-metallic hulls radar loses.
So the highest-value detection problem in Indian waters is precisely the one an imported model has the least training signal for. That is not a policy observation. That is a distribution-shift problem with a straightforward engineering answer: collect and annotate here.
What retraining actually requires
The obvious objection is that any vendor can fine-tune on local data. True in principle. In practice, three things get in the way.
- Data access. Retraining needs annotated footage from the target environment across the seasonal cycle. For naval and coastal-security applications that footage is frequently not exportable, which makes an offshore training pipeline a non-starter before it begins.
- Commercial priority. A global vendor allocates retraining effort by market size. A customer buying tens of units does not move that queue, whatever the contract says.
- Iteration latency. The gap between observing a failure mode at sea and having a corrected model in service is where the real cost sits. When that loop runs through a foreign organisation, a different time zone and an export review, it stretches from weeks into quarters.
The supply argument, stated plainly
Alongside the technical case there is a supply case, and it is worth stating without inflation. India's defence indigenisation lists progressively close import routes for named sub-systems. Spares from abroad arrive on someone else's schedule and someone else's political weather. A model you cannot retrain is a model you cannot fix. None of these are hypothetical risks; all of them are ordinary programme risks that a procurement officer already knows how to weigh.
What we would resist is the version of this argument that treats indigenous as a synonym for adequate. It is not. An Indian system that underperforms an imported one is not a win for anybody, and the indigenisation programmes are not served by companies that treat a badge as a substitute for a specification.
Where we are honest about the limits
Our own bill of materials is not fully indigenous and will not be for some time. Cooled mid-wave detectors are imported. Standard CMOS sensors are imported. AI inference silicon is imported. We publish this because the alternative — implying a fully domestic supply chain that does not exist — devalues the programme for everyone operating inside it.
What is genuinely in-country is everything above the detector: optical integration and calibration, the training data, the models, the fusion layer, the decision logic, the compute integration, the test regime and the support. That is where the performance difference in Indian waters actually comes from, and it is also the layer that can be iterated in weeks rather than quarters.
It is a narrower claim than the marketing version. It is also one we can defend at a design review.