Methods¶
Every method here turns pixel values into a layer you can threshold. None of them proves a mineral is present. What they do is rank ground so you know which pixels to take to the curve check.
The whole set at a glance¶
| Method | What it measures | Needs | Values run | Fails when |
|---|---|---|---|---|
| SFF | Fit to a mineral's individual absorption features | Hyperspectral | Closer to 1 is better | Mineral missing from the library; surface coating; correction residual |
| SAM | Angle between the whole pixel spectrum and a reference | Any | Closer to 1 is better | Spectrally similar non-target material; broad-band sensors blur diagnostics |
| GSI | Band ratios diagnostic of a group or lithology | Any | Higher is more, scale is not 0 to 1 | Ratio satisfied by something else; values do not compare across providers or dates |
| MTMF | Relative abundance of a target within a mixed pixel | Hyperspectral for abundance; multispectral ranks only | Closer to 1 is more abundant | False positives at low abundance; values do not compare between areas or masks |
| AI Spectral Targeting | How far a pixel's spectrum sits from the scene norm | 2+ minerals, a mask | Smaller is a stronger anomaly | Anomalous does not mean mineralized. Man-made ground scores well |
| AI Clusters | Groups of spectrally similar pixels | Any | Categorical, no ranking | Follows surface cover, not bedrock |
| Extracted Edges | Linear features traced from hillshade | Topography | Higher gives fewer, sharper | Traces noise and man-made lines as readily as faults |
| FCC | Three bands rendered as red, green and blue | Any | Not a measurement | Color is not a result, ever |
| Magnetic transformations | Where magnetic properties change | Magnetics | Read contrast, not absolutes | RTP artifacts at low inclination or strong remanence |
Choosing between the mineral mappers¶
Where you have hyperspectral coverage, run SFF first and treat it as your primary evidence. TerraEye runs multi-range SFF, which fits every diagnostic absorption window of the mineral at once — a stricter test than matching the overall curve — and it is the most reliable match the platform produces.
Then run SAM and GSI on the same or related minerals. They are not confirmation of SFF, because all three read the same pixel of the same scene, but a disagreement between them is worth understanding before you go further.
Without hyperspectral data, SAM and GSI are what you have, and on Sentinel-2 both answer at group level. Write your result as "clay-like" rather than as kaolinite, and say why.
A note on thresholds¶
Turn on the gradient color scale before you touch a slider. It shows the value distribution, which lets you put the threshold where strong values break away from the background instead of guessing from the map. Phase 5 walks through the motion for each layer type, including the one that runs backwards.