Natural fracture networks govern subsurface fluid flow, rock-mass stability, and strain accommodation in the brittle crust, yet their automated delineation from outcrop imagery remains challenging due to multi-scale size variability, low contrast between fracture boundaries and host-rock texture, and scene clutter from vegetation, shadows, and blast artifacts. Standard encoder-decoder networks apply fixed-size receptive fields that inadequately span the fractal-like scale range of natural discontinuity networks, while gradient-based classical detectors lack the semantic context required to suppress non-structural scene edges. We introduce GeoFractNet: a dilated U-Net coupling ConvNeXt encoder blocks with multi-scale dilated convolutions (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(d \in \{1,\!2,\!4,\!8\}\)</EquationSource></InlineEquation>), gated edge-aware skip connections fusing Scharr and Gabor filter responses, and a Sobel Edge-Aware (SEA) Loss that directly penalizes gradient misalignment at fracture boundaries. The model output is a binary semantic edge map (per-pixel classification of fracture-boundary pixels versus background) rather than a vectorized fracture trace network; conversion to discrete trace objects requires subsequent post-processing. Trained on GeoCrack, an open-access fracture edge dataset of 12,158 annotated patches from 49 structurally and lithologically diverse outcrops across five countries, GeoFractNet achieves mIoU <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(= 0.91\)</EquationSource></InlineEquation>, Dice <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(= 0.92\)</EquationSource></InlineEquation>, and Boundary F<InlineEquation ID="IEq4"><EquationSource Format="TEX">\(_1 = 0.90\)</EquationSource></InlineEquation>, outperforming all evaluated classical and deep-learning baselines. Compared to the best performing network in literature, YOLACT++, GeoFractNet reduces false positives by 38% and recovers 94% of long-range fracture edges missed by classical (Canny) edge detection on a curated hard-negative evaluation subset (420 patches containing vegetation, blast holes, and strong illumination gradients). Ablation studies confirm that the gated edge-aware skip connections and SEA Loss provide the largest incremental gains in boundary-sensitive metrics, with no architectural component degrading performance at any ablation step. Code, model weights, and the GeoCrack dataset are openly released, providing a reproducible benchmark for automated fracture characterization in structural geology, geomechanics, and reservoir analogue studies.

Semantic edge detection of fractures in geological outcrops using edge aware dilated convolutional networks

期刊论文
出版模式：

开放获取
发表日期：

2026年7月21日

查看全文

Mohammed Yaqoob,
Mohammed Ishaq,
Mohammed Yusuf Ansari,
Thomas Daniel Seers

摘要

Natural fracture networks govern subsurface fluid flow, rock-mass stability, and strain accommodation in the brittle crust, yet their automated delineation from outcrop imagery remains challenging due to multi-scale size variability, low contrast between fracture boundaries and host-rock texture, and scene clutter from vegetation, shadows, and blast artifacts. Standard encoder-decoder networks apply fixed-size receptive fields that inadequately span the fractal-like scale range of natural discontinuity networks, while gradient-based classical detectors lack the semantic context required to suppress non-structural scene edges. We introduce GeoFractNet: a dilated U-Net coupling ConvNeXt encoder blocks with multi-scale dilated convolutions (\(d \in \{1,\!2,\!4,\!8\}\)), gated edge-aware skip connections fusing Scharr and Gabor filter responses, and a Sobel Edge-Aware (SEA) Loss that directly penalizes gradient misalignment at fracture boundaries. The model output is a binary semantic edge map (per-pixel classification of fracture-boundary pixels versus background) rather than a vectorized fracture trace network; conversion to discrete trace objects requires subsequent post-processing. Trained on GeoCrack, an open-access fracture edge dataset of 12,158 annotated patches from 49 structurally and lithologically diverse outcrops across five countries, GeoFractNet achieves mIoU \(= 0.91\), Dice \(= 0.92\), and Boundary F\(_1 = 0.90\), outperforming all evaluated classical and deep-learning baselines. Compared to the best performing network in literature, YOLACT++, GeoFractNet reduces false positives by 38% and recovers 94% of long-range fracture edges missed by classical (Canny) edge detection on a curated hard-negative evaluation subset (420 patches containing vegetation, blast holes, and strong illumination gradients). Ablation studies confirm that the gated edge-aware skip connections and SEA Loss provide the largest incremental gains in boundary-sensitive metrics, with no architectural component degrading performance at any ablation step. Code, model weights, and the GeoCrack dataset are openly released, providing a reproducible benchmark for automated fracture characterization in structural geology, geomechanics, and reservoir analogue studies.