Slope-Weighted Focal Tversky Loss for Fault Scarp Segmentation from Satellite Imagery
Abstract
Fault scarps are geomorphological markers of past surface ruptures caused by earthquakes and provide valuable information for seismic hazard assessment. Mapping these features over large areas is therefore essential but remains a time-consuming task requiring expert interpretation. Recent advances in deep learning and remote sensing offer promising opportunities for automated fault scarp detection. In this study, we used the TerraMind foundation model using Sentinel-2 RGB and infrared imagery together with a TanDEM-X Digital Elevation Model (DEM) to map fault scarps in southern Peru. However, fault scarps are thin and barely visible linear structures, and they occupy only a small fraction of an image. Therefore, they are often confused with similar landforms such as cliffs or riverbanks. This results in severe class imbalance and numerous hard negative samples that the standard Dice loss does not handle efficiently. To address these challenges, we propose a new loss function specifically designed for fault scarp segmentation: the Slope-Weighted Focal Tversky Loss (SWFTL). This combines a Focal Tversky loss (helping the model focus on false negatives) with a slope-weighted cross-entropy term (using DEM slope as a proxy for hard negatives, thereby reducing false-positive detections). Our results show that SWFTL outperforms the standard Dice loss by approximately 44% in terms of relaxed F1-score, demonstrating its effectiveness for the specific challenges of fault scarp segmentation.
Related
- Spatio-Temporal Analysis of Urban Expansion in Awka and its Environs, Nigeria Using Sentinel-2 Satellite Imagery and Google Earth Engine (2021 to 2026)
- Workflow for estimating above ground biomass from satellite imagery with non-geolocated ground validation data
- ConLoRA-Sat: Parameter-Efficient Synthetic Satellite Imagery Generation for Data-Scarce Humanitarian Contexts
- Spatial context importance in deep learning for global horizontal irradiance estimation from satellite imagery
- A Deep‐Learning Model Cascade for Automating Large‐Scale Shoreline Mapping From Satellite Imagery
BibTeX
@misc{faure2026slopeweighted,
title = {{Slope-Weighted Focal Tversky Loss for Fault Scarp Segmentation from Satellite Imagery}},
author = {François Faure and Swann Zerathe and Laurence Audin and Lorena Rosell-Guevara and James Hollingsworth and Sophie Giffard-Roisin},
howpublished = {HAL},
year = {2026},
month = sep,
url = {https://hal.science/hal-05679240v2},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Slope-Weighted Focal Tversky Loss for Fault Scarp Segmentation from Satellite Imagery
AU - François Faure
AU - Swann Zerathe
AU - Laurence Audin
AU - Lorena Rosell-Guevara
AU - James Hollingsworth
AU - Sophie Giffard-Roisin
PY - 2026
DA - 2026/09/05
PB - HAL
UR - https://hal.science/hal-05679240v2
AB - Fault scarps are geomorphological markers of past surface ruptures caused by earthquakes and provide valuable information for seismic hazard assessment. Mapping these features over large areas is therefore essential but remains a time-consuming task requiring expert interpretation. Recent advances in deep learning and remote sensing offer promising opportunities for automated fault scarp detection. In this study, we used the TerraMind foundation model using Sentinel-2 RGB and infrared imagery together with a TanDEM-X Digital Elevation Model (DEM) to map fault scarps in southern Peru. However, fault scarps are thin and barely visible linear structures, and they occupy only a small fraction of an image. Therefore, they are often confused with similar landforms such as cliffs or riverbanks. This results in severe class imbalance and numerous hard negative samples that the standard Dice loss does not handle efficiently. To address these challenges, we propose a new loss function specifically designed for fault scarp segmentation: the Slope-Weighted Focal Tversky Loss (SWFTL). This combines a Focal Tversky loss (helping the model focus on false negatives) with a slope-weighted cross-entropy term (using DEM slope as a proxy for hard negatives, thereby reducing false-positive detections). Our results show that SWFTL outperforms the standard Dice loss by approximately 44% in terms of relaxed F1-score, demonstrating its effectiveness for the specific challenges of fault scarp segmentation.
ER -