EchoChange: A Diffusion Language Model with Dual Pass Remasking for Factual Remote Sensing Disaster Change Captioning

2026-08-03

AI Summary

This work addresses the issue of cascading factual errors in autoregressive generation for bitemporal remote sensing disaster change captioning, which often stems from early misjudgments. To mitigate this, the authors propose a multimodal discrete diffusion language model framework that formulates the task as an iterative masked denoising process conditioned on image pairs. The approach enables non-autoregressive sentence-level rewriting and factual correction through draft-aware two-stage training, a progressive masking curriculum, and a confidence-guided remasking strategy. Evaluated on the RSCC benchmark, the method significantly outperforms both general-purpose and remote sensing–specific baselines across lexical and semantic metrics, achieving state-of-the-art performance.

Abstract

Bi-temporal remote-sensing disaster change captioning often needs to identify sparse and spatially localized changes across large pre- and post-event scenes and then translate them into coherent, factual descriptions. However, existing change captioning methods always follow an autoregressive decoding paradigm to generate the change description and thus an early misinterpretation of the changed object, event, or spatial relation becomes an irreversible premise for subsequent text, amplifying visual ambiguity into cascading factual errors. To address this limitation, we propose EchoChange, a multimodal discrete diffusion language model that formulates change captioning as iterative masked-token denoising rather than left-to-right generation. By repeatedly revising the entire caption while conditioning on the image pair, EchoChange can reconsider uncertain content and correct imperfect intermediate predictions. We further introduce draft-aware dual-pass training, a progressive masking curriculum, and confidence-guided remasking to align training with iterative inference. Extensive experiments on the RSCC benchmark show that EchoChange substantially outperforms both general-purpose and remote-sensing-specific baselines across lexical and semantic metrics. The EchoChange Project is at https://github.com/sundongwei/EchoChange_Project

Problem

• remote sensing
• change captioning
• factual error
• autoregressive generation
• disaster monitoring

Innovation

• diffusion language model
• dual-pass remasking
• factual captioning
• remote sensing change detection
• iterative denoising