Case study on the optimal scaling factor for semantic segmentation of remote sensing image

Yuanzhi Cai*, Yuan Fang, Fang Huang, Lei Fan

*Corresponding author for this work

Research output: Chapter in Book or Report/Conference proceedingConference Proceedingpeer-review

Abstract

Random scaling is a widely employed training technique to mitigate overfitting in semantic segmentation models, enabling models to handle images with varying scaling factors. However, existing deep learning frameworks for remote sensing images adopt a default scaling factor of 1.0 for single-scale testing. This raises the question of whether this scaling factor is optimal for single-scale testing. To address this question, this study investigates the optimal scaling factor across six models for three remote sensing datasets. The results show that the optimal scaling factor for the UAVid, LoveDA, and Potsdam was 0.75, 1.0, and 1.25, respectively.

Original languageEnglish
Title of host publication2024 International Conference on Machine Intelligence for GeoAnalytics and Remote Sensing, MIGARS 2024
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9798350389678
DOIs
Publication statusPublished - 2024
Externally publishedYes
Event2024 International Conference on Machine Intelligence for GeoAnalytics and Remote Sensing, MIGARS 2024 - Wellington, New Zealand
Duration: 8 Apr 202410 Apr 2024

Publication series

Name2024 International Conference on Machine Intelligence for GeoAnalytics and Remote Sensing, MIGARS 2024

Conference

Conference2024 International Conference on Machine Intelligence for GeoAnalytics and Remote Sensing, MIGARS 2024
Country/TerritoryNew Zealand
CityWellington
Period8/04/2410/04/24

Keywords

  • CNN
  • Data augmentation
  • Scaling factor
  • Semantic segmentation
  • Single-scale testing
  • Transformer

Fingerprint

Dive into the research topics of 'Case study on the optimal scaling factor for semantic segmentation of remote sensing image'. Together they form a unique fingerprint.

Cite this