Or an image detection model. Fraction of the compute and can run even on edge embedded. And easy to train with your own data

Clouds and nighttime are a barrier to visual detection. Even with good effectiveness the conditions needed for that would mean that you have far less than 50% uptime, and your downtime is predictable to your adversary.

A cheap radar takes an order of magnitude less power to run on hardware that is cheaper than an LLM and can see way farther than a camera.