Overview
This article explains how to interpret the Depth Accuracy (R²) gauge and Real vs Predicted scatter plot on your Prediction Map. These outputs help you evaluate how well the model predicts the vertical position of mineralized targets, and what to do if results are not reliable.
Depth Accuracy Gauge
When you first open your results, you’ll see a Depth Accuracy display in the form of a circular gauge, or odometer, showing:
Accuracy Score – A percentage representing how closely predicted depths match actual depths
Label – A qualitative indicator of model performance
Labels
81-100% - High (green)
51-80% - Moderate (yellow)
0-50% - Low (red)
A higher percentage indicates better alignment between predicted and actual depths. Lower scores may suggest the model is struggling to learn vertical patterns from your data. Ideally, you want to see the High label.
To explore the results further, click the arrow in the upper right to view detailed model outputs.
MAE and Samples
MAE
MAE (Mean Absolute Error) shows the average size of the model’s errors. It tells you how far predictions are from the real values, on average.
Samples
Samples shows the number of data points where the model’s predicted depth is compared to the real depth.
What is the Real vs Predicted Scatter Plot?
Below the gauge, you’ll find the Real vs Predicted scatter plot. This graph visualizes how closely the model’s predicted depths align with actual depths across multiple test-train splits.
How the plot works:
X-axis: True depth
Y-axis: Predicted depth
Blue dots: Depth samples used for validation
Green band: Good fit region
Red zones: Underestimation or overestimation regions
Dashed line: Perfect 1:1 fit
Each dot shows one sample where both the real depth and the model’s predicted depth are known. The closer the dots fall to the 1:1 line, the better the predictions.
What the coloured bands mean:
Points in the upper red band show underestimation (model predicts too shallow).
Points in the lower red band show overestimation (model predicts too deep).
Points in the green band indicate a good fit.
This layout gives you an immediate sense of whether the model is predicting too deep, too shallow, or reliably across the depth range.
How to Interpret
Strong Performance
A High Depth Accuracy label
MAE values that are reasonable for the deposit type
A tight cluster of dots following the 1:1 line
Most dots falling within the green band
Under-Prediction (Shallower Than Reality)
Most dots fall above the 1:1 line.
The model predicts depths that are too shallow.
Over-Prediction (Deeper Than Reality)
Most dots fall below the 1:1 line.
The model predicts depths that are too deep.
Weak or Unstable Patterns
Dots scatter widely with no clear diagonal trend.
The model may be defaulting to average depths instead of learning vertical patterns.
Few dots may indicate not enough valid depth samples.
What to Do if Results Aren’t Optimal
If the Real vs Predicted scatter plot shows a weak or inconsistent pattern, your model may not be effectively predicting depth. Try the following adjustments:
Lower the classification threshold
In Step 3: Set Up Learning Data, reduce the value used to define mineralized points. A high threshold can result in too few positive samples, limiting valid training data for depth modeling.
Review your input features
In Step 2: Select Input Features, assess whether the rasters you’ve selected are suitable for predicting depth. Remove features with little vertical variation or excessive noise, and experiment with different raster combinations. Replacing underperforming features with more geologically relevant inputs can significantly improve model accuracy.
Adjust model training parameters
In Step 4: Apply Data Fusion, increase the number of epochs to allow more training cycles. In Step 5: Build Predictive Model, refine cluster settings (e.g. cluster size or minimum points) to better capture subsurface patterns.
Learn More
Interpret other DORA’s Result Graphs:
Create a DORA Prediction Map:
Still Have Questions?
Reach out to your dedicated DORA contact or email support@VRIFY.com for more information.