Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Non-Registration Change Detection: A Novel Change Detection Task and Benchmark Dataset

Zhe Shan, Lei Zhou, Liu Mao, Shaofan Chen, Chuanqiu Ren, Xia Xie

TL;DR

NRCD addresses detecting semantic changes when multi-temporal remote-sensing images are not perfectly registered. The authors define $D_{train} = \\{(X_i^{T1}, X_i^{T2}, Y_i)\\}_{i=1}^{N_{train}}$ and preserve $Y_i \\in \\{0,1,...,C\\}$ while applying transformations to $X_i^{T2}$; the reference image is $T_1$ and the post-event image is $T_2$. They propose eight transformations — RR, RP, CJ, OD, SR, GB, CD, GS — implemented with torchvision/Albumentations to synthesize NRCD datasets, and show these can catastrophically degrade SOTA detectors on datasets like LEVIR-CD. They release code and datasets to enable benchmarking.

Abstract

In this study, we propose a novel remote sensing change detection task, non-registration change detection, to address the increasing number of emergencies such as natural disasters, anthropogenic accidents, and military strikes. First, in light of the limited discourse on the issue of non-registration change detection, we systematically propose eight scenarios that could arise in the real world and potentially contribute to the occurrence of non-registration problems. Second, we develop distinct image transformation schemes tailored to various scenarios to convert the available registration change detection dataset into a non-registration version. Finally, we demonstrate that non-registration change detection can cause catastrophic damage to the state-of-the-art methods. Our code and dataset are available at https://github.com/ShanZard/NRCD.

Non-Registration Change Detection: A Novel Change Detection Task and Benchmark Dataset

TL;DR

NRCD addresses detecting semantic changes when multi-temporal remote-sensing images are not perfectly registered. The authors define and preserve while applying transformations to ; the reference image is and the post-event image is . They propose eight transformations — RR, RP, CJ, OD, SR, GB, CD, GS — implemented with torchvision/Albumentations to synthesize NRCD datasets, and show these can catastrophically degrade SOTA detectors on datasets like LEVIR-CD. They release code and datasets to enable benchmarking.

Abstract

In this study, we propose a novel remote sensing change detection task, non-registration change detection, to address the increasing number of emergencies such as natural disasters, anthropogenic accidents, and military strikes. First, in light of the limited discourse on the issue of non-registration change detection, we systematically propose eight scenarios that could arise in the real world and potentially contribute to the occurrence of non-registration problems. Second, we develop distinct image transformation schemes tailored to various scenarios to convert the available registration change detection dataset into a non-registration version. Finally, we demonstrate that non-registration change detection can cause catastrophic damage to the state-of-the-art methods. Our code and dataset are available at https://github.com/ShanZard/NRCD.
Paper Structure (18 sections, 2 figures, 1 table)

This paper contains 18 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Preliminary
  4. Non-Registration Change Detection
  5. Random Rotation
  6. Random Perspective
  7. Color Jitter
  8. Optical Distortion
  9. Solarization
  10. Gaussian Blur
  11. Coarse Dropout
  12. Grayscale
  13. Experiment and Discussion
  14. Evaluation Metrics
  15. Experimental Settings
  16. ...and 3 more sections

Figures (2)

  • Figure 1: Here are examples of several types of non-registration change detection images: (a) The image captured at time $T_1$, representing the state before changes occur. (b) The image captured at time $T_2$, showing the state after changes have occurred. (c) The change area mask. (d) The flight paths at times $T_1$ and $T_2$ do not align. (e) Inconsistencies in image modality between $T_1$ and $T_2$. (f) The image at $T_2$ is affected by electromagnetic interference, and some areas are distorted.
  • Figure 2: The visualization results of the proposed eight non-registration scenarios are shown in the LEVIR-CD, GVLM, and SYSU-CD datasets (from left to right). The abbreviations RR, RP, CJ, OD, SR, GB, CD, and GS stand for random rotation, random perspective, color jitter, optical distortion, solarization, Gaussian blur, coarse dropout, and grayscaling, respectively.