Remote sensing has been a staple of mineral exploration since Landsat-1 went up in 1972, and for most of that history the workflow has been roughly the same: process an image, look for spectral signatures that correlate with alteration minerals, draw polygons around the anomalies, and hand them to the field crew. The methods are textbook, the limitations are well understood, and the technique earned its place in the standard exploration toolkit decades ago.
What's changed in the last decade is the underlying data and the analytical methods on top of it. Sensors have gotten substantially better — WorldView-3 delivers 1.24-meter multispectral and 30-centimeter panchromatic, while hyperspectral platforms like PRISMA, EnMAP, and EMIT cover hundreds of spectral bands at moderate spatial resolution. The processing methods have shifted from spectral angle mapping and basic ratio compositing toward ML-based mineral mapping that handles mixed pixels, atmospheric noise, and shadow effects more robustly. The combination has materially raised the value of remote sensing for mineral targeting, particularly in arid and exposed terrains where surface expression of alteration is interpretable.
What the Sensors Are Good At
Different sensors solve different problems, and matching the sensor to the exploration question is the difference between a useful study and an expensive one. ASTER, on the Terra satellite since 1999, has fourteen bands across visible, short-wave infrared (SWIR), and thermal infrared (TIR), and remains the workhorse for regional alteration mapping. The SWIR bands map clay, sericite, alunite, and carbonate minerals; the TIR bands map quartz, silica, and aluminosilicates. ASTER coverage is global, the data is freely available, and the methodology for processing it is mature. For a regional alteration study, ASTER is usually the first dataset to look at.
Landsat-8 and Landsat-9 — and their predecessors back to Landsat-4 — provide medium-resolution multispectral coverage with regular revisit times. Landsat is weaker than ASTER for mineral discrimination because of fewer and broader SWIR bands, but it's better for change detection, vegetation suppression, and regional iron-oxide mapping. The combination of multiple Landsat scenes through time, processed with modern cloud-removal and atmospheric correction, produces composites that handle clouds and seasonal effects better than any single image.
WorldView-3, operated by Maxar, has eight SWIR bands at 3.7-meter resolution and eight multispectral bands at 1.24 meters. The SWIR coverage is comparable to ASTER's but at roughly ten times the spatial resolution, which transforms what was a regional reconnaissance tool into a detailed mapping tool. For prospect-scale work where individual outcrops matter, WorldView-3 is the highest-quality publicly available data, with the caveat that it's commercial and pricing reflects that. Tasking new acquisitions runs into the tens of thousands of dollars for moderate-size areas.
Hyperspectral sensors — PRISMA (Italian Space Agency), EnMAP (German), EMIT (NASA, mounted on the ISS), and several upcoming commercial platforms — deliver hundreds of contiguous narrow bands across visible through SWIR. This is the data type that the spectroscopy textbooks were written for: any mineral with a diagnostic absorption feature in this range can in principle be uniquely identified. In practice, atmospheric correction, mixed pixels, and signal-to-noise ratio impose limits, but hyperspectral remote sensing is the closest available remote analog to laboratory infrared spectroscopy.
Where ML Methods Replace Older Workflows
Classical mineral mapping methods — spectral angle mapping (SAM), matched filtering, mixture-tuned matched filtering (MTMF) — remain useful but have known weaknesses. SAM is sensitive to illumination but insensitive to spectral magnitude. Matched filtering assumes a known background that's rarely available. Each method works in some scenarios and breaks in others, and choosing between them historically required spectroscopy expertise.
Modern ML approaches handle the spectral mixing problem more robustly. Unsupervised clustering on hyperspectral pixel spectra, after appropriate dimensionality reduction, identifies natural endmember populations in the data without requiring a pre-specified library. Supervised classifiers — random forests, support vector machines, shallow neural networks — trained on field-validated alteration mineralogy produce per-pixel probability maps for each mineral class. These methods don't replace classical spectroscopy interpretation; they augment it by handling the mixed-pixel and noise problems more gracefully.
Convolutional neural networks specifically address the spatial-context problem. A single pixel's spectrum can be ambiguous; the spectrum of a 3x3 neighborhood — averaged, ratioed, or processed by a CNN that learns spatial patterns — is much less ambiguous. For high-resolution imagery like WorldView-3, CNN-based classification typically outperforms pixel-by-pixel methods by a meaningful margin.
The most useful ML contribution in 2026 is in atmospheric correction and noise suppression. The published research on physics-aware neural networks for radiative transfer correction is starting to flow into operational hyperspectral processing pipelines, replacing or supplementing standard atmospheric correction algorithms like FLAASH and ATCOR. The improvements are modest but compound: better atmospheric correction means cleaner endmember spectra, which means more reliable downstream mineral identification.
The Vegetation Problem
The honest limitation of optical remote sensing for mineral exploration is vegetation. In any terrain with meaningful vegetation cover — tropical, temperate forest, even moderately vegetated semi-arid — direct mineral mapping from optical sensors is severely limited because the dominant signal is from the canopy, not the underlying lithology. Subtle geobotanical signals — vegetation stress, species composition shifts, anomalous canopy reflectance — can sometimes indicate mineralization through indirect mechanisms, but the methods are difficult and the false-positive rate is high.
For vegetated jurisdictions, the realistic role of optical remote sensing is in mapping the visible outcrops, road cuts, stream beds, and historic workings, then extrapolating geology using the structural and topographic signatures that the imagery does show clearly. Lidar — particularly bare-earth lidar that strips the vegetation canopy — has become a more useful tool than spectral imagery in heavily vegetated exploration areas, because it reveals structural and geomorphic patterns under the canopy that optical sensors cannot see.
What a Modern Workflow Looks Like
A 2026 remote-sensing-driven targeting workflow in an arid jurisdiction looks roughly like this. Pull regional ASTER coverage as a baseline, process for alteration minerals using the standard ratio combinations and band-math approaches. Identify the regional alteration footprints. For high-priority sub-areas — anomalies, known mineralization, prospective ground — task or acquire WorldView-3 or commercial hyperspectral coverage. Process the higher-resolution data with ML-based classification to produce detailed alteration maps. Integrate with regional geophysics (magnetics, radiometrics) and any available geochemistry. Output a target package combining alteration footprints, geophysical signatures, and structural interpretation.
The compute is now trivial on a modern laptop or modest cloud instance. The data acquisition is the bottleneck — high-quality commercial imagery is expensive and tasking adds weeks of latency. The cost-effective approach is to plan acquisitions carefully: use free regional data to identify where to spend money on detailed coverage, rather than acquiring commercial coverage of the whole license area speculatively.
What's Worth Knowing About Cloud Platforms
Google Earth Engine has become the de facto platform for medium-to-regional-scale remote sensing analysis. The platform hosts most of the relevant satellite archives (Landsat, Sentinel, ASTER, MODIS), runs analyses on Google's compute infrastructure, and is free for research and substantially subsidized for commercial users. For regional exploration studies, Earth Engine has shifted the cost structure from local infrastructure to cloud-hosted scripts, which has democratized regional remote-sensing work.
For higher-resolution commercial imagery — WorldView, Planet, hyperspectral platforms — cloud delivery is increasingly the norm, with platforms like Maxar's GeoHIVE, Planet's Planet Insights, and integrated analytics offerings from the major imagery providers. The processing tools embedded in these platforms have improved substantially; what used to require local ENVI seats and PhD-level processing expertise is now accessible through cloud workflows with a credit card.
The catch is that commercial cloud platforms encourage workflow lock-in. A targeting workflow built around Earth Engine, Planet Insights, or Maxar's analytics is portable in principle but practically tied to the platform's specific tooling. For one-off studies that doesn't matter. For longer-term capability building, juniors should be deliberate about which workflows they want to keep portable and which they're comfortable handing off to a vendor.
Cost and Realistic Adoption
The cheapest entry point for ML-augmented remote sensing is open-source: a Python environment with rasterio, scikit-learn, and Earth Engine access, applied to free ASTER and Landsat data. Total cost is essentially the analyst's time. A first-pass regional alteration study at this cost level produces useful but not exhaustive coverage.
A mid-tier study, contracted to a remote-sensing specialist, runs in the $15,000 to $40,000 range for a moderate-size area using a mix of free and modest commercial data. The deliverable is a target package and methodology documentation. This is the right level of spend for a junior with a focused district-scale targeting question.
High-end studies using commercial high-resolution and hyperspectral imagery run into the six figures for area-and-resolution combinations that justify it — typically prospect-scale work on confirmed targets where every percent of targeting precision matters. These are the right investments for an advanced project entering drill programs, not for early-stage reconnaissance.
What This Doesn't Replace
Remote sensing identifies surface signatures. It does not see through cover, it does not see through vegetation reliably, and it does not directly indicate subsurface continuity. The targets it generates are surface-projection targets that still require geophysics, drilling, and the rest of the exploration toolkit to convert into actual discoveries. The mistake some junior boards make is treating a remote-sensing-derived target package as if it were a drill-ready set of targets; in reality, it's an input to a longer process of integration and validation.
The technique is also limited by the fact that not every deposit type expresses on the surface. Buried mineralization without surface alteration, deposits under deep cover, deposits in heavily vegetated terrain — these are dominated by other geophysical signatures and remote sensing contributes little. For target portfolios in deep-cover jurisdictions, the right remote-sensing budget is small.
A Practical First Step
If your project covers exposed terrain and you've never used remote sensing systematically, the first useful step is a free-data baseline: an ASTER alteration study and a Landsat structural and vegetation analysis over your property package. Open-source tools, a few days of analyst time, and the result is a regional context that informs every other exploration decision.
If a baseline is already in hand and you have a focused target area, the next step is commercial high-resolution coverage on that area specifically — WorldView-3 SWIR is the standard choice — with ML-based mineral mapping on the result. This is the prospect-scale work that produces drill-ready alteration maps when the geology cooperates.
For an outside view on what remote sensing would add to your specific project, our free workflow audit covers exploration data workflows, or contact us to discuss what a pilot study would look like.