Monitoring Lithium Fields with Geo-Embeddings
The rapid adoption of electric vehicles and the massive energy infrastructure expansion required to support AI data centers have driven lithium demand sharply higher. As lithium suppliers race to keep up with global demand, significant resources have been poured into scaling up the sought-after metal’s two main sources: hard rock mines and salt brine evaporation fields. The latter involves pumping salty groundwater into shallow ponds and letting the sun concentrate the lithium, a 12–24 month process accompanied by stunning visual changes that make the fields a perfect candidate for AI analysis and geo-embeddings.
The Problem
Over half of the world’s identified lithium resources per USGS lie beneath the salt flats of the so-called “Lithium Triangle” – a ~150,000 square-mile region spanning Chile, Argentina, and Bolivia. To understand the world’s shifting lithium market, it’s crucial to monitor the Lithium Triangle. This is no trivial task; while lithium fields are easily distinguishable in satellite imagery, manually observing and measuring change over a region as large as the Lithium Triangle can be costly and time intensive.
To answer questions such as “what new fields were constructed between 2020 and 2024?” or “how much did a certain field expand year-over-year?”, embeddings can deliver far more efficient results than traditional image-based analysis. Using those insights, lithium suppliers or commodity researchers can back into far more accurate yield forecasts, reliably locate every field in the region, observe recent changes, and more. The catch is that embeddings are spread across many foundation models, resolutions, and imagery types, so just assembling the right ones for a region this size can be a project in itself.
The Solution
During my summer fellowship at LGND, I built a lithium field detector application to solve this problem.
L:ithum Field Detector built using LGND Platform
The LGND platform handles the heavy lifting: processing collections, scanning large regions, and learning what a new field looks like. That let me focus on the analyst workflow on top of it: the loop of labeling, training, and reviewing that turns a probe into a map of change. Here’s the summary workflow, step by step:
- Label positives and negatives for ponds (e.g., known evaporation ponds vs. surrounding salt flat, roads, and buildings).
- Click train.
- Tune the score threshold and the change gate to control how strict “new” is.
- Flip through the flagged clusters, mark false positives, and repeat steps 1–3.
- Once satisfied, export the flagged areas as GeoJSON polygons; the tool traces the outlines for you.
The real power comes from the LGND’s classifier, which lets the user train a simple linear probe: label a few examples of positives and negatives for the lithium fields, train, refine, and repeat. Every pixel got a pond score for 2024 and for 2020. A pixel is flagged as a new field only when it scores as a pond in 2024, scores as not-a-pond in 2020, and its embedding actually changed between the two years. That last check, a “change gate,” screens out pixels whose score flipped while the ground stayed the same. By toggling through these clusters, users can quickly flag false positives and confirm true new fields.
The Results
To see how the workflow holds up on real ground, I ran it end to end over the Salar de Atacama in northern Chile: a test area of about 27 × 29 km (roughly 780 km²) that takes in SQM’s two pond complexes and Albemarle’s. The embeddings were Google’s AlphaEarth Foundations annual embeddings (64 dimensions, sampled at 20 m) for 2020 and 2024. The value comes from what LGND’s change-analysis tool does with them: labeling, a trained probe, a change check across years, and export.
LGND’s lithium change analyzer on the Salar de Atacama test area, with 25 pond labels (green) and 25 not-pond labels (red) placed on 2024 imagery.
I placed 25 pond and 25 not-pond labels on 2024 imagery. The pond labels spanned every color a brine pond goes through as it concentrates: blue, cyan, green, olive, orange-brown, near-black, and white or pink salt crust, plus a few small, new-looking cells. The negatives skipped easy open desert and went after look-alikes: plant buildings and worker camps, tanks, well pads, salt stockpiles, roads, brine canals, an empty bermed pond bed, and four kinds of bare salar (reddish, dark and damp, light crust, and a grey alluvial fan).
First pass: a fast fit, and too many flags
Training took 15 milliseconds, and the probe got 46 of its 50 labels right (92%). The four misses were telling. Two were ponds that scored low: a green cell in SQM’s north complex and a large dark-green pond at Albemarle. The other two were salt stockpiles that scored as ponds. Piled salt shares the bright, crystalline look of a salt-crusted pond, which makes it exactly the kind of negative the probe needs more of.
Across the whole test area, the first pass over-reached. In 2024 the probe scored 46% of the area as pond, far more than the pond complexes actually cover, including a long diagonal band of bare salar (Figure 2). The new-field view inherited the problem: at the default settings (score threshold 0.50, change gate 0.10) it flagged about 2,660 hectares, which exported as 1,707 polygons.
First-pass 2024 pond scores over the Salar de Atacama test area (red = high, blue = low). The probe lights up the pond complexes, but also a broad band of bare salar.
The largest cluster, about 270 hectares, sits on open salar just east of Albemarle’s ponds (Figure 3). Its shapes are irregular streaks rather than the rectangular cells of an engineered pond, which points to natural change in the salt-flat surface rather than construction. Two large ponds at the edge of the complex are flagged too; whether they are new, refilled, or a scoring artifact needs a check against 2020 imagery.
The largest “new field” cluster (amber) at the default settings, on open salar east of Albemarle’s ponds. Streaky, irregular shapes like these mark natural surface change: the false positive the review step exists to catch.
Why it over-flagged
The salt flat changes on its own. Moisture and crust shift from year to year, and those shifts move a pixel’s embedding much like building a pond does. At our labeled bare-salar sites, the change between 2020 and 2024 reached 0.33, more than three times the 0.10 gate. The gate let that natural change straight through.
Scores drift between years. The probe was trained on 2024 embeddings. The same 25 labeled ponds that averaged a score of 0.73 in 2024 averaged only 0.52 in 2020. When the score shifts between years on its own, “pond in 2024, not in 2020” can fire on ponds that were there all along. The change gate is the guard against that, which is why its setting matters so much.
The threshold isn’t the fix. Sweeping the score threshold barely moved the total: about 2,840 hectares flagged at 0.40, 2,660 at 0.50, and 2,390 at 0.60. The false positives don’t come from a borderline cutoff. They come from the labels and the change gate.
New-field flags (amber) at score thresholds of 0.40, 0.50, and 0.60, change gate 0.10. The scattered flags on open salar barely change, which is the point: the threshold isn’t where the false positives come from.
The fix is the loop the tool is built for: add damp-salar and salt-pile negatives right where the false clusters sit, raise the change gate toward the roughly 0.33 the tool’s own diagnostics suggest, retrain, and review what’s left.
Takeaways
“Find every new lithium field in the Lithium Triangle” seemed like a tall ask. Embeddings get you a first answer fast; the Atacama test shows that a first answer isn’t a final one. Creating my field detector hinged on whether I could easily centralize these collections of embeddings, and LGND removed that blocker entirely. I’ve seen firsthand that embeddings are remarkably good at answering geospatial questions, and LGND’s strength is removing the friction in getting to those answers.