Overview
The Computer Vision Maps Module, part of the Data Augmentation suite, generates spatial context from raster data by quantifying the local characteristics of the data around each pixel. It uses a pre-trained deep learning model, called ResNet-50, to produce high-dimensional numerical features, or embeddings, that encode characteristics like the shape, textures, contrast, and variation within each patch of input data.
ResNet-50 was originally designed for image recognition, but in the application of Data Augmentation, it is used to generate spatially aware features from existing data layers. Rather than classifying what an image contains, it summarizes the local patterns in the data numerically, as embeddings. These are then used as inputs in DORA’s predictive modelling workflows to improve performance and spatial sensitivity.
The output helps DORA recognize geological patterns that are defined by surrounding textures and gradients, not just isolated pixel values. In other words, it allows the algorithm to think more like a geologist by examining the setting of any given location, not just the location itself.
Topic | Summary |
Module Name | Computer Vision Maps |
Purpose | Extracts high-dimensional spatial features from raster data using a deep learning model |
Input Format | Raster |
Recommended Data | Geophysical Data; also relevant to other data types with uniform coverage and relevance to exploration, like DEM or remote sensing. |
Output Format | Raster |
Key Parameters | Grid resolution, AOI, number of filters |
Processing Summary | Applies RESNET-50 feature extraction to input grids |
Typical Use Cases | Enhancing data for ML models, encoding subtle spatial patterns |
Validation or QC | Not applicable (unsupervised feature engineering) |
Common Pairings | VRIFY Prospectivity Score (VPS), Multivariate Anomaly Detection, Distance Maps |
Notable Output Notes |
|
How It’s Used in Exploration
Computer Vision Maps are used within DORA to provide spatial context around each raster pixel. Instead of analyzing individual cell values in isolation, the Module summarizes the surrounding texture and structural features of the data. This allows predictive models to account for patterns, boundaries, and gradients that often carry geological meaning.
The process works by dividing the input raster into overlapping patches. Each patch is passed through the ResNet model, which returns a numerical embedding. This embedding is a set of features that captures how the data varies in that local area. These features help DORA recognize subtle structures, contrasts, and patterns that might indicate changes in lithology, alteration zones, or structural overprints.
The outputs are not meant to be visualized on their own, although they sometimes reflect observable geological features, but are intended to work behind the scenes. Their effectiveness is seen in how well they support other Modules and enhance overall model performance.
Value and Benefits
Computer Vision Maps improve exploration targeting by allowing models to evaluate the geological context surrounding each location. By encoding patterns in shape, contrast, texture, and structure, the Module gives machine learning workflows a more complete understanding of the data.
This reduces the risk of overfitting to isolated anomalies and encourages the model to recognize meaningful spatial relationships, similar to the way a geologist interprets data. The result is predictions that are more geologically reasonable, not just statistically accurate.
Because the Module applies the same process across the entire Area of Interest, this Module offers a consistent and objective way to include texture and structure in modelling. The result is a more informed model that better reflects geological context and supports confident, spatially aware targeting.
Learn more
Still have questions?
Reach out to your dedicated DORA contact or email support@VRIFY.com for more information.