Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RemoteVAR: Autoregressive Visual Modeling for Remote Sensing Change Detection

Yilmaz Korkmaz, Vishal M. Patel

TL;DR

Change detection in remote sensing is challenged by cross-time nuisances like illumination and misalignment. RemoteVAR addresses this by a coarse-to-fine, scale-wise autoregressive framework that tokenizes bi-temporal data with a frozen multi-scale VQ-VAE and conditions token generation via cross-attention on fused continuous features, followed by decoder refinement. The approach delivers state-of-the-art or competitive performance on WHU-CD and LEVIR-CD, with ablations confirming the importance of cross-attention, RGB-based mask tokens, token randomness to mitigate exposure bias, and decoder refinement. This work demonstrates that autoregressive models can be a practical and effective alternative to diffusion and transformer baselines for dense change detection in remote sensing, with implications for efficiency and controllability in practical deployments.

Abstract

Remote sensing change detection aims to localize and characterize scene changes between two time points and is central to applications such as environmental monitoring and disaster assessment. Meanwhile, visual autoregressive models (VARs) have recently shown impressive image generation capability, but their adoption for pixel-level discriminative tasks remains limited due to weak controllability, suboptimal dense prediction performance and exposure bias. We introduce RemoteVAR, a new VAR-based change detection framework that addresses these limitations by conditioning autoregressive prediction on multi-resolution fused bi-temporal features via cross-attention, and by employing an autoregressive training strategy designed specifically for change map prediction. Extensive experiments on standard change detection benchmarks show that RemoteVAR delivers consistent and significant improvements over strong diffusion-based and transformer-based baselines, establishing a competitive autoregressive alternative for remote sensing change detection. Code will be available \href{https://github.com/yilmazkorkmaz1/RemoteVAR}{\underline{here}}.

RemoteVAR: Autoregressive Visual Modeling for Remote Sensing Change Detection

TL;DR

Change detection in remote sensing is challenged by cross-time nuisances like illumination and misalignment. RemoteVAR addresses this by a coarse-to-fine, scale-wise autoregressive framework that tokenizes bi-temporal data with a frozen multi-scale VQ-VAE and conditions token generation via cross-attention on fused continuous features, followed by decoder refinement. The approach delivers state-of-the-art or competitive performance on WHU-CD and LEVIR-CD, with ablations confirming the importance of cross-attention, RGB-based mask tokens, token randomness to mitigate exposure bias, and decoder refinement. This work demonstrates that autoregressive models can be a practical and effective alternative to diffusion and transformer baselines for dense change detection in remote sensing, with implications for efficiency and controllability in practical deployments.

Abstract

Remote sensing change detection aims to localize and characterize scene changes between two time points and is central to applications such as environmental monitoring and disaster assessment. Meanwhile, visual autoregressive models (VARs) have recently shown impressive image generation capability, but their adoption for pixel-level discriminative tasks remains limited due to weak controllability, suboptimal dense prediction performance and exposure bias. We introduce RemoteVAR, a new VAR-based change detection framework that addresses these limitations by conditioning autoregressive prediction on multi-resolution fused bi-temporal features via cross-attention, and by employing an autoregressive training strategy designed specifically for change map prediction. Extensive experiments on standard change detection benchmarks show that RemoteVAR delivers consistent and significant improvements over strong diffusion-based and transformer-based baselines, establishing a competitive autoregressive alternative for remote sensing change detection. Code will be available \href{https://github.com/yilmazkorkmaz1/RemoteVAR}{\underline{here}}.
Paper Structure (10 sections, 4 figures, 2 tables)

This paper contains 10 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Methodology
  4. Experiments and Results
  5. Implementation Details
  6. Experimental setup.
  7. Baselines
  8. Results
  9. Ablation Studies
  10. Conclusion

Figures (4)

  • Figure 1: Overview of the RemoteVAR architecture and training pipeline. For clarity, we visualize only the first three token scales with grid sizes $1{\times}1$, $2{\times}2$, and $3{\times}3$. Pre-image, post-image, and fused feature streams are color-coded in red, blue, and green, respectively. Trainable modules are marked with a fire icon, while frozen components are marked with an ice icon.
  • Figure 2: Scale-wise autoregressive mask generation shown across progressively finer token resolutions from $\mathbf{1\times1}$ to $\mathbf{16\times16}$.
  • Figure 3: Decoder refinement procedure is illustrated.
  • Figure 4: Qualitative prediction comparisons are shown for the WHU-CD whu (top row) and LEVIR-CD levir (bottom row) datasets. True positives are colored white, false positives green, and false negatives red.