Artificial Intelligence and Data Science
Online ISSN : 2435-9262
Comparative study of multiple evaluation metrics for deep-learning-based extraction results of uphill-facing scarps
Toshu FUJINOKI, Hiroshi OKAWA, Kazuki KIKUCHI, Heitaro KANEDA, Kazuo KASHIYAMA
Author information
JOURNAL OPEN ACCESS

2026 Volume 7 Issue 3 Pages 844-854

Details
Abstract

This paper compares the characteristics of multiple evaluation metrics for deep-learning-based extraction results of uphill-facing scarps. Since uphill-facing scarps appear as narrow and elongated linear features, the Global Dice coefficient based on pixel-level overlap alone may not sufficiently evaluate line continuity, component-level positional validity, or spatial distribution. In this study, extraction results obtained under different input data conditions were evaluated using the Global Dice coefficient, clDice, Centroid-matching F1, and Grid-based F1 from the viewpoints of pixel-level overlap, linear structure, component-level detection, and spatial distribution. The results showed that MPI+DEM(4ch), in which DEM information was added as a fourth channel to the MPI red relief image map, achieved the highest performance in most metrics covering pixel-wise overlap, component-level proximity, and spatial distribution, supported by multiple evaluation viewpoints. Meanwhile, clDice for linear structure evaluation showed that MPI+DEM(3ch) achieved comparable performance to MPI+DEM(4ch), confirming that different metrics capture distinct extraction characteristics. These findings demonstrate that, when evaluating fine linear landforms, combined interpretation of multiple metrics according to the intended use is effective rather than relying on a single metric.

Content from these authors
© 2026 Japan Society of Civil Engineers
Previous article Next article
feedback
Top