Gains:
- Ability to explain how lithology, fault and alteration mapping is done with AI in satellite and aerial images.
- Ability to design the workflow that extracts mineral and structure markers from multispectral/hyperspectral data
- Ability to test remote sensing outputs with ground truth and geological context
A geological map is the projection of the underground story of a region on the surface: which rock unit is where, where do the faults pass, in which direction are the layers inclined. Classically, this map is a product that the geologist produces by traveling from outcrop to outcrop, with months of fieldwork. Today, remote sensing and artificial intelligence accelerate most of this process. In this unit, we will see how lithology, fault and alteration (chemical change of rock with hydrothermal waters; a critical indicator in mineral exploration) mapping is done with artificial intelligence from satellite / aerial images, and most importantly, how these outputs are tested with the reality of the land.
A warning from the beginning: the remote sensing image is a measurement, not an interpretation; But every map derived from that measurement is an interpretation and there is error in the interpretation. An "alteration zone" marked by artificial intelligence could be an actual mineral indicator, a dried field, a shadow, or an atmospheric disturbance. The map is a hypothesis until verified in the field.
Basic Physics of Remote Sensing
Each mineral and rock reflects sunlight at different wavelengths at different rates; this is called the spectral signature. The human eye only sees visible light (red-green-blue), but satellite sensors measure a much wider range, including infrared. Two basic concepts:
Multispectral imaging measures in several broad wavelength bands (typically 4-12 bands); for example, Landsat and Sentinel satellites. It distinguishes rough rocks. Hyperspectral imaging, on the other hand, measures in hundreds of narrow bands and can recognize minerals almost from their "spectral fingerprint"; It is used to distinguish alteration minerals (such as kaolinite, alunite, illite).
AI does two jobs here: first, it classifies this high-dimensional spectral data (which pixel is which lithology/mineral). Second, it extracts fault and structural features from the geometric patterns of the image (lineaments, texture, drainage network). When the Digital Elevation Model (DEM for short; three-dimensional elevation map of the land) is added, slope, aspect and topographic lineaments are also included in the analysis.
Step by step: an alteration map workflow
- Prepare the image. Atmospheric correction (removing the distorting effect of the atmosphere), cloud masking, vegetation mask. If this step is skipped the entire result will be contaminated.
- Narrow your area of interest. Limit the study area to the geological context (known fault, formation, outcrop); Give context to the model.
- Calculate spectral indices. Band ratios that highlight specific mineral groups (e.g. for iron oxides, clay minerals).
- Classify. Artificial intelligence classifies pixels/regions and marks possible alteration zones.
- Schedule terrain verification. Select sample points from the marked zones and check them in the field. Without this step the map will not be signed.
The Role and Limits of Artificial Intelligence in Image Analysis
In particular, AI reduces two types of errors: it catches subtle anomalies that are overlooked, and it standardizes inconsistent interpretations. But it produces three systematic errors. The first is a false positive: vegetation, shade, human structure, or soil moisture may appear to be alteration. The second is training bias: if the model is trained on data from another region, it will be wrong on the different lithology of your site. Third is the illusion of scale: one pixel covers tens of meters on the ground; Fine structures are lost, mixed pixels (multiple materials in one pixel) are misleading.
So the golden rule: ground truth always wins. The map resulting from the image cannot be considered definitive without testing it with samples and observations at selected points in the field. Verification points are not chosen randomly, but strategically, to cover both "certain" and "suspicious" zones.
Three Mini Cases: By the Numbers
Case 1 — Narrowing the search area. An exploration company was evaluating a license area of 1,200 km². Artificial intelligence-assisted alteration classification from hyperspectral data reduced the area to 7 target zones of 38 km². Instead of visiting the entire area, the field team prioritized these 7 zones; initial field campaign time was reduced by an estimated 70%. 5 of the 7 zones showed real alteration in the field, 2 were false positives (clayed soil) — without verification, these 2 zones would have been wasted drilling.
Case 2 — Fault mapping preliminary sketch. DEM and image-extracted lineament analysis suggested two minor lineaments alongside the known main fault. One turned out to be an actual fault in the field; the other turned out to be the topographic trace of a pipeline route. The model captured the geometric pattern correctly but could not distinguish its origin; He distinguished between geological plausibility and field control.
Case 3 — Cloud shadow fallacy. In one area the model marked a strong "iron oxide alteration" on the northern slope. Atmospheric correction was incomplete and the signature was actually a spectral distortion of a thin cloud cover. When the image was reprocessed and corrected, the anomaly disappeared. Lesson: preprocessing error fools even the most advanced model.
Weak Prompt / Strong Prompt
Weak prompt:
Show places where there may be mines in this satellite image.
Powerful prompt:
Your role: Remote sensing geologist. Interpret the following spectral index and band ratio values (image Sentinel-2, arid zone, sparse vegetation).- For each possible alteration zone: write which index/ratio triggered it.- Suggest AT LEAST one false positive source for each zone (vegetation, shadow, soil, man-made, atmospheric).- Sort the zones by "terrain verification priority" (high/medium/low).- Do not present any zone as "mined"; "alteration token must be verified" data: [index/rate table + context]
The strong prompt forces the model to list false positives itself and avoid claiming accuracy.
Four Copiable Templates
1) Pre-processing checklist:
Check for preprocessing steps that may have been omitted in the following remote sensing workflow: atmospheric correction, cloud/shadow mask, vegetation mask, topographic correction, sensor calibration. Explain in one sentence for each how the missing steps could bias your conclusion.Workflow: [definition]
2) Alteration zone interpretation:
Extract possible alteration mineral groups from the following spectral index values (clay, iron oxide, carbonate, silica). For each group: trigger index,confidence level, alternative explanation (false positive), proposed field check.Claiming certainty. Data: [indexes]
3) Lineament/fault separation:
Consider the list of linearities below. List the possible origins for each: fault, plate boundary, dyke, drainage, human structure (road/line/field), vegetation boundary. State the most likely origin with geological context, but leave out the only possibility. List: [linearities + context]
4) Verification point plan:
Propose a field verification plan for the following alteration/lithology map:- How many sample points for each class, why.- Balanced sampling of "definite" and "questionable" zones.- Access and security note. Map summary: [classes and areas]
Image Type and Area of Use
Data type
What measures
His strength is
limit
Multispectral (Sentinel/Landsat)
broadbands
Rough lithology, regional scanning
Mineral separation is limited
hyperspectral
Hundreds of narrow bands
Mineral/alteration distinction
Cost, processing overhead
DEM (height)
topography
Fault/lineament, drainage
Lithology does not give
thermal infrared
surface temperature
Some silicate/carbonate
sensitive to atmosphere
Radar (SAR)
Surface roughness/deformation
Under cloud, subsidence/landslide
Interpretation is complicated
Tip: Use remote sensing output as a “site prioritizing scan,” not a “drilling map.” Its greatest value is that it tells you where to go first; Field evidence and the engineer decide where to drill.
Caution: Even the most advanced classification produces systematic errors if pre-processing (atmospheric and topographic correction, masking) is omitted. In case of "output is weird" query the preprocessing first, not the model. The garbage in, garbage out principle is ruthless here.
Common mistakes
- Skipping preprocessing. Mistaking anomalies from uncorrected images for real geology.
- Forgetting false positives. Mistaking plants, shade, soil moisture and human structures as alterations/faults.
- Signing without ground truth. Basing the drilling decision on the image map without verifying it in the field.
- Scale illusion. Forgetting the pixel size and ignoring the fact that fine structures are lost or mixed pixels are misleading.
- Ignoring educational bias. Blindly applying a model trained in another region to a field with a different lithology.
In summary
- Remote sensing measures the spectral signature of minerals at different wavelengths; It distinguishes multispectral coarse and hyperspectral fine.
- Artificial intelligence classifies this high-dimensional data and extracts fault/linearity from geometric patterns; but it produces false positives, training bias, and scale bias.
- Preprocessing (atmospheric/topographic correction, masking) determines the quality of the result; If it is omitted, a systematic error will occur.
- Remote sensing is a screening/prioritization tool; The map is not considered accurate without being tested with ground truth.
- Ground truth always wins; Verification points should cover both certain and suspicious zones in a balanced manner.
Application task
Anonymize the spectral index/band ratio values or a DEM lineament list you have for a worksite (or an open data region). Interpret possible alteration zones or faults with the powerful prompt and have the model suggest at least one false positive source for each zone. Then draw up a site control plan with the “verification point plan” template; Note in one sentence how many points are divided into "suspicious" zones and why you are performing balanced sampling.
checklist
- [ ] I can distinguish what multispectral and hyperspectral data measure and what they are used for.
- [ ] I know how preprocessing steps (trimming, masking) affect the result.
- [ ] I question sources of false positives (vegetation, shadow, man-made) in alteration/fault outputs.
- [ ] I view the remote sensing map as a scanning tool and test it with terrain verification.
- [ ] I can plan the verification points to balance certain and suspicious zones.