Problem description
For this challenge, your task is to develop models and algorithms that provide accurate information about the presence of structures that are indicative of geothermal resources—namely, geological faults.
A unique feature of this challenge is that ground truth data for currently known faults is publicly available! While detailed, it is known that this set of faults is not complete and may even contain some inaccurate data.
For this challenge, we have consulted with fault experts who have manually identified faults that are not contained within the current public USGS database. These new faults will comprise the test dataset for the initial prize round of the challenge. After the initial prize round, an expert panel will use submitted predictions to update fault labels for the full region. Submissions will then be rescored against the entire updated label set for the final round awards.
Looking for a starting point? This reference solution implements a simple approach.
Competition structure
This competition has a unique prize structure to address the challenge of incomplete and potentially inaccurate ground truth data, and to encourage participants to submit predictions for what they truly believe are faults in the region as opposed to simply optimizing for the existing labels.
For this competition, the sponsors have compiled a set of newly identified faults in the GeoDAWN region that are not included in the existing USGS fault database. These new faults have been manually labeled by fault experts and comprise the test dataset for the competition. The GeoDAWN region is chunked and split into a public test set and a private test set. Competitors' performance against the public test set is shown on the public leaderboard during the competition.
In the Initial Prize Round, submissions are evaluated by their performance on the private test set at the close of the competition. An expert panel then uses submitted predictions to update fault labels for the full region. Submissions are rescored against the entire updated label set for the Final Prize Round awards. Competitors must choose a single submission for scoring across both rounds before the deadline, without knowing their private test set performance.
The diagram below illustrates how the same set of submissions flows through both rounds:
| Round | Ground truth used for scoring | How your submission contributes |
|---|---|---|
| Initial Prize Round — $50,000 | A fixed private set of new faults that the sponsor's experts labeled before the competition started | Each submission is scored independently. Top 5 each win $10,000. |
| Final Prize Round — $250,000 | An expanded label set — Initial Prize Round labels plus previously-unknown faults that experts verify after reviewing every team's submission | Your submission is re-scored against the expanded set. Predictions that helped experts identify previously-unmapped faults can score higher here than in the Initial Prize Round. |
The Final Prize Round is where participants directly contribute new geological knowledge: faults that you flagged, that experts confirm, become part of the map of the region.
Datasets
The challenge focuses on the region covered by the GeoDAWN study. GeoDAWN (Geoscience Data Acquisition for Western Nevada) consists of a set of high-resolution magnetic and radiometric surveys conducted to support geologic and geophysical mapping and modeling. A map of the areas covered by GeoDAWN is shown below:
The datasets provided by this challenge come from the United States Geological Survey and from the Great Basin Center for Geothermal Energy’s INGENIOUS project. We will provide some of the data for this challenge directly via the data download page, including data from the GeoDAWN study and the ground truth fault vector and raster data from the USGS. We will also provide a list of URLs where the high-resolution Digital Elevation Model (DEM) data can be downloaded.
Provided features
On the data download page, you will find a GeoTIFF file called training_features.tif. This GeoTIFF is in a projected coordinate system for UTM zone 11N (EPSG 32611) at 100m resolution. The GeoTIFF has many layers:
- Surface conductivity and depth to conductive base surface
- Detrended elevation and the slope of detrended elevation
- Dilatation rate, shear strain rate, and the second invariant of the strain rate tensor
- Isostatic gravity anomaly and the slope of the isostatic gravity anomaly
- Magnetics including reduced-to-pole magnetic anomaly, total magnetic intensity, the vertical and horizontal slope of total magnetic intensity, and the top-of-crustal magnetic source depth estimate
- Density of earthquakes
In addition, you will find a CSV file called 1m_DEM_links.csv that contains links where DEM data at 1m resolution can be downloaded.
Labels
The labels for this challenge come from the USGS quaternary fault maps and from INGENIOUS. Labels are provided in both vector and raster formats on the data download page.
External datasets
Participants are allowed to use any additional data sources, provided that the participants possess a license that permits the data to be used in this challenge and shared with the sponsor for evaluation purposes. For more information, see the complete rules document.
Performance metric
The performance metric for this competition is a distance-weighted Tversky index. The Tversky index is a similarity metric parameterized by coefficients alpha and beta representing the penalty term applied, respectively, to false positives and false negatives.
For this challenge, you will submit a GeoTIFF raster with fault predictions represented as pixel-wise probabilities or confidence scores between 0 and 1. Rasterization is lossy and can induce off‑by‑one errors near pixel boundaries. In addition, portions of the existing fault data may be misaligned from the true location of the surface fault, which is the prediction target. To mitigate these effects, we weight the contributions of true positives, false negatives, and false positives by the distance to the nearest ground truth pixel using a linear (triangular) kernel with 300 m support.
Mathematical representation
Let |$p(x) \in [0,1]$| denote the predicted probability of a fault at pixel |$x$|, and let |$g(x)$| denote the ground truth label of pixel |$x$|.
The Tversky index |$TI$| is defined as:
$$ \mathrm{TI}(\alpha,\beta) = \frac{\sum_x p(x) g(x)}{\sum_x p(x) g(x) + \alpha \sum_x p(x) (1-g(x)) + \beta \sum_x (1-p(x)) g(x)} $$
where |$\alpha$| and |$\beta$| are non-negative parameters that control the penalty applied to false positives and false negatives, respectively.
To compute a distance-weighted Tversky index, we define a triangular kernel |$k$| as $$ k(d) = \bigl(1-d/R\bigr)_+ = \max\left(1-\frac{d}{R},\,0\right). $$
where the range |$R$| is 300 meters (i.e., 3 pixels at 100m resolution). This kernel is linearly larger when distances are smaller and vice versa.
For probabilistic predictions, we compute each term in the Tversky index in a distance-weighted manner:
$$ \text{distance-weighted true positive count}\,\mathrm{TP}_w = \sum_{g \in G} \max_{x\,:\,d(x,g)\leq R} p(x)\,k\bigl(d(x,g)\bigr) $$ $$ \text{distance-weighted false positive count}\,\mathrm{FP}_w = \sum_{x\,:\,p(x)>0}p(x)\,\Bigl[1-\max_{g \in G}k\bigl(d(x,g)\bigr)\Bigr] $$ $$ \text{distance-weighted false negative count}\,\mathrm{FN}_w = \sum_{g \in G}\Bigl[1-\max_{x\,:\,d(x,g)\leq R}p(x)\,k\bigl(d(x,g)\bigr)\Bigr] $$
and the distance-weighted Tversky index is given by
$$ \mathrm{DTI}(\alpha,\beta) = \frac{\mathrm{TP}_w}{\mathrm{TP}_w+\alpha\,\mathrm{FP}_w+\beta\,\mathrm{FN}_w + \epsilon} $$
For this competition, we set |$\alpha = 0.2$| and |$\beta = 0.8$|, which reduces the penalty for false positive predictions and increases the penalty for false negative predictions.
Scoring example
Consider the following example of a ground truth raster. The ground truth is a single vertical line. We then apply a triangular kernel with |$R=3$| to determine the kernel weights.
To calculate the distance-weighted Tversky index, we compute the weighted true positive, false positive, and false negative counts according to the equations given above. The original and distance-weighted contributions from each pixel are shown below for clarity. Note that false positives that are closer to ground truth pixels have a lower penalty due to the distance weighting, and false negatives that are closer to high-confidence predictions also have a reduced penalty.
We then select the values from the weights that correspond to pixels for that label criterion and sum them, weighting by the supplied values for |$\alpha$| and |$\beta$|.
$$ \mathrm{TP}_w = 3.00, \mathrm{FP}_w = 1.89, \mathrm{FN}_w = 2.00 $$
$$ \mathrm{TI}_w (\alpha=0.2, \beta=0.8) = \frac{3.00}{3.00 + 0.2 \cdot 1.89 + 0.8 \cdot 2.00} = 0.60 $$
Submission format
For this competition, you will submit a GeoTIFF file containing your predictions for all faults in the region.
Your submitted GeoTIFF should meet the following requirements:
- Your submission is in the same projected coordinate reference system as the training data (projected coordinate system for UTM zone 11N, EPSG 32611)
- Your submission is at the same resolution as the training data (100m)
- Your submission has the same bounds as the training data, and data outside the bounds is null or nan.
- Your submission contains a single layer with datatype of 32-bit float (
float32) with values between 0 and 1 indicating the confidence or probability of fault presence, with higher values indicating higher probability.
A sample submission that predicts total fault absence is provided for your reference on the data download page. You can use this as a template to ensure that your submission is correctly formatted.
Good luck!
Good luck and enjoy this problem! If you have any questions you can always visit the competition forum!