Mining · Mineral Exploration
Kakamega sits on one of East Africa's most promising gold frontiers. We mapped where to look first.
Two independent models, trained on 68 confirmed gold occurrences and 61 layers of evidence, rank 226.5 km² of ground as high priority. Both were tested on ground they had never seen.
Built on Kaistrum Minerals
Targeting work, mapped
- Gold, Kakamega County
- Your ground here
- Your ground here
- Your ground here
The ground
Gold this far underground does not announce itself.
The Liranda Corridor is a 12 km structural zone inside the Busia-Kakamega Greenstone Belt. It already hosts the Isulu and Bushiangala high grade orogenic gold deposits, in the Nyanzian volcanic sequence.
That is what makes targeting hard. The gold sits in shear zones, so finding more of it means reading structure rather than surface. Soil chemistry smears the signal over hundreds of metres, and the available fault mapping is too coarse to show the small splays that actually carry the mineralisation.
61
Data layers integrated
geochemistry, remote sensing, structure, terrain and lithology
68
Confirmed occurrences
known gold occurrences the models were trained on
0.981
Random Forest
ranking score on held-out ground
0.975
Support Vector Machine
ranking score on held-out ground
226.5km²
High priority ground
Random Forest High and Very High classes combined
0
Occurrences missed
both models recovered every known occurrence
How it works
Four streams of evidence, 61 layers in all.
Geochemistry
Regional soil geochemical data covering 24 pathfinder and major elements, plus seven element ratios. Gold pathfinders include arsenic, antimony, tungsten, bismuth, copper, lead, zinc, manganese, silver and the iron to manganese ratio. These tie directly to the sulphide assemblage confirmed at Isulu and Bushiangala: arsenopyrite, sphalerite, galena and molybdenite.
Remote sensing alteration
Alteration indices derived from satellite imagery: Clay-AlOH, ferrous mineral and iron oxide. These pick up the mineral changes that hydrothermal fluids leave behind in the rock around a deposit.
Structure
Euclidean distance to mapped fault structures. Gold here is structurally controlled, so how close a point sits to a known fault carries real predictive weight.
Terrain and lithology
DEM derivatives covering slope, plan curvature, profile curvature and general curvature, alongside 21 lithological units encoded as binary layers. Soil organic matter and pH are included as covariates.
Two models
One model tells you what it found. Two tell you whether to believe it.
A single model can latch onto a quirk of its own mathematics and present the result with total confidence. So we ran two that work in completely different ways, trained separately, and compared where they landed. They independently converge on the same high priority target zones. Where both of them point at the same ground, that agreement is worth more than either score on its own.
Map A, model comparison
Drag the handle, or focus it and use the arrow keys, to wipe between the two models.
Validation
Tested on ground it had never seen.
The study area was cut into 133 blocks on a 5 km fishnet, and whole blocks were held back from training. Training and validation data were always spatially separated. That is what stops simple proximity to a known deposit from inflating the score.
Training used 409 points in total: 68 confirmed gold occurrences and 341 background points. Class imbalance was handled with balanced class weighting. Validation ran as groupKFold cross validation using the spatial blocks as groups, plus a groupshufflesplit spatial holdout reserving 25 percent of blocks.
The ranking score below is the spatial holdout, ROC-AUC figure. It measures ordering, not correctness: it is how often the model scores a known occurrence above a random background point. A ranking score of 1.0 is perfect ordering. 0.5 is a coin toss.
Solid bars are Random Forest, lighter bars are Support Vector Machine. Both reached 100% recall on the 68 known occurrences, meaning zero omission errors.
What the data shows
Three things the models agree on.
- 01Banded iron formation and phyllite dominate as host lithology across the Very High zones.
- 02The highest confidence targets cluster in the Liranda Corridor and in Lurambi and Hamisi sub counties.
- 03Elevated arsenic, antimony and tungsten signatures run across the target zones.
Work with Kairos
Prospectivity targeting for exploration companies, ministries and regulators.
If you hold ground, license it, or regulate it, we can build the same targeting workflow over your data.
Built with
ArcGIS Pro, ArcGIS Notebooks, Python with scikit-learn, GDAL, arcpy
Prospectivity targeting within Kaistrum Minerals.
Working coordinate system: Arc 1960 / UTM Zone 36N (WKID 21096)