A Study on the Performance Comparison of Sediment Movement Detectors Using Unet-Based Models and the Sliding Partitioning Method

12 pages•Published: August 28, 2025

Xiaosong Liu, Makoto Oda, Kei Kawamura and Tsuyoshi Wakatsuki

Abstract

Japan is covered by mountainous and hilly terrain and thus is prone to disasters such as landslides caused by earthquakes and heavy rainfall. Currently, the identification of sediment disaster areas is conducted through visual interpretation of aerial photographs taken after these disasters. To improve the efficiency of this process, research has been advancing in the automatic detection of sediment movement areas using deep learning techniques applied to aerial photographs. In this study, we compare the performance of sediment movement detectors using Unet and Unet++, in order to investigate the impact of changes in the loss function parameters during training on detection performance. Additionally, we evaluate the performance improvement when the sliding partitioning method is used for input images to the detectors. Results show that using sliding partitioning and adjusting the value of the loss function coefficient β can improve detection performance.

Keyphrases: debris flow, deep learning, image segmentation, unet, unet++

In: Jack Cheng and Yu Yantao (editors). Proceedings of The Sixth International Conference on Civil and Building Engineering Informatics, vol 22, pages 247-258.

Links:	https://easychair.org/publications/paper/C1Nz
	https://doi.org/10.29007/3kz8

BibTeX entry

@inproceedings{ICCBEI2025:Study_Performance_Comparison_Sediment,
  author    = {Xiaosong Liu and Makoto Oda and Kei Kawamura and Tsuyoshi Wakatsuki},
  title     = {A Study on the Performance Comparison of Sediment Movement Detectors Using Unet-Based Models and the Sliding Partitioning Method},
  booktitle = {Proceedings of The Sixth International Conference on Civil and Building Engineering Informatics},
  editor    = {Jack Cheng and Yu Yantao},
  series    = {Kalpa Publications in Computing},
  volume    = {22},
  publisher = {EasyChair},
  bibsource = {EasyChair, https://easychair.org},
  issn      = {2515-1762},
  url       = {/publications/paper/C1Nz},
  doi       = {10.29007/3kz8},
  pages     = {247-258},
  year      = {2025}}

Download PDF Open PDF in browser