How Engineers Know a Missile Hit Its Target

Ukraine says its missiles hit the Kazan aviation plant deep inside Russia. Official statement came hours after the strike. But OSINT analysts on Telegram had georeferenced footage, acoustic sensor cross-references, and fire brigade dispatch logs corroborated within 45 minutes. No government needed.

Most engineers don't realize how much of modern conflict verification runs on the same infrastructure they use to debug microservices. Let me break down what's actually happening technically when a strike gets confirmed.

The Sensor Fusion Problem

Confirming a strike location is a data fusion problem. You've got heterogeneous, unreliable sources coming in asynchronous order, and you need a confidence score fast. Sound familiar? It's exactly like aggregating health checks from a distributed system where half the nodes are lying.

The primary inputs: Sentinel-2 and Planet Labs satellite imagery (3-5m resolution, 12-hour revisit cycles minimum), Flightradar24 and ADS-B Exchange for airspace closures that signal active operations, Telegram channels as an unstructured event stream, Windy.com atmospheric data to model smoke plume drift, and Wigle.net-style RF sensor networks that some groups run to detect radar activity.

Groups like Bellingcat and the GeoConfirmed project built custom tooling on top of Google Earth Engine, QGIS, and Overpass API to cross-reference reported coordinates against known facility footprints in OpenStreetMap. The pipeline is mostly Python glue code, Telegram bots pulling message metadata, and shared Google Sheets that would make any data engineer cry.

The Georeferencing Stack

Here's where it gets interesting from an engineering standpoint. When a shaky phone video surfaces, analysts run it through a process called georeferencing: matching architectural features, shadow angles, terrain lines, and antenna configurations against satellite basemaps.

Tools used: SunCalc.org to compute solar azimuth at a given timestamp, which lets you reverse-engineer location from shadow direction to within a few degrees. Google Earth Pro's historical imagery layer. Mapillary for street-level reference photos where they exist. And increasingly, people are running custom YOLO-based models fine-tuned on Soviet-era industrial architecture to auto-flag candidate locations from building silhouettes.

I've built alerting pipelines at Stripe that were less rigorous than what some 22-year-old running a Telegram channel has put together in their spare time. That's not a compliment to us.

The Data Integrity Problem Nobody Talks About

In 2019, working on fraud signal aggregation with a team of about 60 engineers, we had a pipeline that ingested 40M events per day and had a 3% false positive rate we thought was acceptable. It wasn't. At that volume, 3% is 1.2M bad signals daily compounding into downstream decisions.

OSINT networks have the same problem. A single viral misattributed video poisons every downstream analysis that cites it. The Kazan plant claim has multiple independent corroborations, which is exactly the right architecture: you want N independent confirmations from non-overlapping sources, not N citations of the same upstream event.

This is why the OSINT community built tools like Hunch.ly for evidence chain tracking and Hypothesis for annotation. They're solving data provenance. Same problem Databricks and dbt are solving for analytics pipelines.

What Governments Are Actually Using

NGA and equivalent agencies run automated change detection on satellite imagery using convolutional nets trained on multispectral bands, not just RGB. Thermal IR for fire signatures. SAR (synthetic aperture radar) penetrates cloud cover and works at night. Planet Labs sells API access. Maxar has a direct feed to defense customers.

The gap between what a well-funded OSINT analyst can access commercially in 2024 versus classified systems is smaller than it's ever been. That's the actual story here. The democratization of remote sensing APIs has made geospatial intelligence a software engineering problem, not a classified one.

Don't underestimate what you can build with Planet's APIs, a decent object detection model, and a Postgres database with PostGIS. I've seen this exact stack used to monitor construction permits, supply chain logistics, and now, apparently, missile strikes.