Copper-Gold Porphyry Systems in Pakistan: How AI Spots Alteration Zones From Orbit

I was looking at a Sentinel-2 scene over the Chagai belt last March when something caught my eye. A faint yellow-brown halo, maybe 2.3 km across, sitting right next to a small intrusive body that nobody had logged. Two weeks later I had a field team on the ground. The rocks came back with visible chalcopyrite stringers.

That's the job. That's literally what we do at GeoMine AI every single day.

Porphyry copper-gold systems are the biggest prize in Pakistani mining. Reko Diq alone holds an estimated 5.9 billion tonnes of ore. And here's the thing — the Chagai arc that hosts it runs for roughly 480 km across Balochistan, and most of it has never been properly mapped at high resolution. Not even close.

What a porphyry actually looks like from space

A porphyry system isn't a single rock. It's a plumbing network. Hot fluids escape from a cooling intrusion and cook the surrounding rock into concentric zones — potassic in the center, then phyllic, then argillic, then propylitic out at the edges. Each zone has its own mineral signature. Sericite, illite, kaolinite, chlorite, epidote. And every one of those minerals absorbs light at specific wavelengths.

That's the part most people miss. You're not looking for copper from space. Copper itself doesn't show up. You're looking for the alteration halo — the cooked rock — which can be 5 to 10 times larger than the actual ore body.

ASTER's SWIR bands (especially bands 5, 6, 7, and 8) are tuned almost perfectly for this. Band ratios like (5+7)/6 light up phyllic alteration. (7+9)/8 picks up propylitic. Sentinel-2 doesn't have the same SWIR resolution but its 20-meter bands 11 and 12 still catch hydroxyl-bearing minerals well enough to be useful as a first pass.

I used to think ASTER alone was enough. I was wrong. Cloud cover, old acquisitions (the sensor's been degrading since 2008), and gaps in coverage mean you need a stack. Sentinel-2 for freshness, ASTER for mineral specificity, SAR for structure, SRTM for the topographic context. That's the stack we run.

Where the AI part actually matters

Honestly, band ratios alone have been around since the 90s. Geologists at the Geological Survey of Pakistan have used them for decades. So what's new?

The new bit is pattern recognition at scale. A trained convolutional model doesn't just look at one pixel — it looks at the spatial arrangement of pixels. Porphyry alteration has a specific geometry. Concentric. Roughly circular to elliptical. Usually 1 to 4 km across. Often sitting at the intersection of two fault systems. A human analyst can recognize this in a single scene. An AI can do it across 40,000 square kilometers in an afternoon.

We trained our model on a labeled dataset that included Reko Diq, Saindak, Dasht-e-Kain, and some validated targets from Iran's Kerman belt (geologically similar setting). The model now flags candidates with a confidence score. Anything above 0.78 we mark for field follow-up. Below that, we still log it but don't prioritize.

False positives are real. Sometimes hydrothermally altered rock from non-economic systems shows the same SWIR signature. Sometimes agricultural soil in the wrong moisture state mimics argillic alteration. We've had embarrassing moments. One target I was excited about in 2023 turned out to be a kaolinite-rich weathering profile over granodiorite — no mineralization at all. The model flagged it. I flagged it harder. We were both wrong.

So we added structural filters. If the alteration zone doesn't sit near a mapped or inferred fault, the score drops. If there's no associated topographic anomaly (porphyries often erode into circular depressions or resistant knobs depending on the stage), the score drops more. Layer the evidence. That's the only way.

What this means for Pakistan specifically

The Chagai belt is the obvious target. But it's not the only one. The Kohistan island arc up north has porphyry-style mineralization that's barely been touched. The Waziristan ophiolite margins show alteration patterns we still don't fully understand. And there are anomalies in the Bela ophiolite zone that nobody's published on.

I own 15 leases in Gilgit-Baltistan and I'll tell you straight — the satellite work changed how I think about each one. Two of my leases I almost dropped. Then the alteration mapping showed me something I'd missed walking the ground for three years. Different scale, different perspective.

Look, satellite data doesn't replace drilling. Anyone who tells you that is selling something. What it does is narrow the target. If you have a 200 sq km lease, knowing which 3 sq km to trench first can save you a year and maybe $400,000 in wasted exploration. That's the actual value proposition for copper exploration AI in Pakistan — not magic, just better triage.

The report we generate for a porphyry target usually includes: ASTER alteration maps (argillic, phyllic, propylitic separated), Sentinel-2 false color composites, SAR-derived lineament analysis, SRTM-based structural interpretation, and the AI confidence map overlaid on all of it. About 40 pages. Costs less than two days of a drilling rig.

Is every target real? No. We've had a hit rate around 31% on field-verified targets so far, which sounds low until you compare it to traditional grassroots exploration where you're lucky to get 5%. Six times better than walking blind. That's the number that matters.

If you've got ground in Chagai, Kohistan, or anywhere along the Tethyan suture and you haven't run satellite alteration mapping on it yet — what are you waiting for?