🎓 Lesson 6 D4

Convolutional Architectures for Core Image Segmentation

Convolutional architectures are AI models that scan rock core images like a geologist’s eye—spotting patterns in textures, colors, and boundaries to automatically label mineral zones.

🎯 Learning Objectives

  • Explain how convolutional kernels detect geological textures in core imagery
  • Design a minimal U-Net architecture for binary ore/waste segmentation on 512×512 core scans
  • Analyze model performance using IoU and boundary F1-score on segmented drill core patches
  • Apply transfer learning from pretrained ImageNet weights to fine-tune on limited field-collected core data

📖 Why This Matters

In grade control, misclassifying even 5% of pyrite-rich zones in core can shift resource estimates by millions of dollars. Manual logging is slow, subjective, and inconsistent across geologists—while CNN-based segmentation delivers repeatable, sub-millimeter boundary detection at >100 cores/hour. This lesson bridges AI theory to the drill rig: showing how convolutional models turn pixel data into actionable geological decisions.

📘 Core Principles

CNNs operate in two phases: feature extraction (encoder) and spatial reconstruction (decoder). The encoder uses stacked convolutional layers with ReLU activation and max-pooling to progressively reduce spatial resolution while increasing channel depth—capturing edges → textures → lithological motifs. The decoder (often symmetric) upsamples feature maps and fuses them with encoder skip connections to preserve high-resolution localization—critical for sharp contact delineation in core. Architectural choices (kernel size, stride, padding) directly impact sensitivity to thin veining (<2 mm) or diffuse alteration halos. Modern variants (e.g., Attention U-Net) add spatial attention gates to suppress noise in low-contrast zones like clay-rich wallrock.

📐 Segmentation Accuracy Metrics

IoU (Intersection over Union) quantifies pixel-level agreement between predicted and ground-truth masks—essential for evaluating contact fidelity in core segmentation. It penalizes both over-segmentation (false positives) and under-segmentation (false negatives), making it more geologically meaningful than pixel accuracy alone.

Intersection over Union (IoU)

IoU = |P ∩ G| / |P ∪ G|

Measures pixel-wise overlap between predicted segmentation mask (P) and ground-truth mask (G); primary metric for segmentation quality in geoscience applications.

Variables:
SymbolNameUnitDescription
P Predicted mask pixel set Binary or multi-class segmentation output from the CNN
G Ground-truth mask pixel set Expert-drawn reference segmentation used for training and validation
Typical Ranges:
Operational grade control: 0.75 – 0.92
Research-grade lithological mapping: 0.65 – 0.85

💡 Worked Example

Problem: A U-Net predicts an ore zone covering 1,842 pixels; the expert-labeled ground truth covers 1,796 pixels; their overlap (intersection) is 1,630 pixels.
1. Step 1: Compute intersection = 1630 pixels
2. Step 2: Compute union = (1842 + 1796 − 1630) = 2008 pixels
3. Step 3: Apply IoU = intersection / union = 1630 / 2008
Answer: The result is 0.812, which falls within the acceptable range of ≥0.75 for operational grade control segmentation per CIM Best Practices (2022).

🏗️ Real-World Application

At the Red Lake Gold Mine (Ontario), a custom U-Net trained on 4,200 labeled 512×512 patches from hyperspectral core scans reduced ore/waste misclassification at the quartz-carbonate–pyrite contact by 63% vs. manual logging. Model outputs were integrated into LeapFrog Geo™ via GeoJSON polygons, enabling real-time grade shell updates during diamond drilling—cutting assay turnaround lag from 14 days to <48 hours for critical development drives.

📋 Case Connection

📋 Gold Mine Real-Time Grade Control at Development Drift Face

Manual chip sampling caused 24–48 hr delay in stope boundary decisions, leading to 14% dilution

📋 Iron Ore Mine Sensor Fusion for Banded Iron Formation (BIF) Delineation

Conventional geophysics failed to resolve thin hematite bands (<2m) within jaspilite, causing 22% grade variance in ROM...

📋 Limestone Mine Digital Twin for Karst-Related Grade Uncertainty

Solution cavities caused unpredictable grade drops (CaCO₃ purity <85%) in otherwise uniform deposits, resulting in 11% p...

📚 References