EchoChange: A Diffusion Language Model with Dual Pass Remasking for Factual Remote Sensing Disaster Change Captioning
Dongwei Sun, Bowen Yao, Yujie Zhang, Pei Liu, Jing Yao, Xiangyong Cao
cs.AI
2026-08-03
EchoChange treats disaster change captions as discrete diffusion with dual-pass remasking, reaching 26.18 ROUGE-L and 65.34% single-error Strict Hit on RSCC.
Post-disaster assessment needs more than a change map. Analysts have to name what changed, where it happened, and at what scale. Bi-temporal remote-sensing change captioning turns a pre-event and post-event image into that statement. Real change is sparse and local. Illumination, season, sensor shift, and misregistration can look like change. If the caption gets the object, event, or spatial relation wrong early, the rest of the sentence inherits the error.
Most existing methods decode left to right. An early mistake becomes the prefix that conditions every later token, and there is no edit step. Teacher forcing makes the mismatch worse: training sees gold prefixes, inference sees the model's own errors. Prior RS diffusion work does not close this gap. Diffusion-RSCC denoises continuous word embeddings. Mask Approximation Net diffuses a visual change mask and still uses an autoregressive caption decoder. The caption hypothesis itself is not an editable discrete state.
EchoChange, from Xi'an Jiaotong University and the Aerospace Information Research Institute of the Chinese Academy of Sciences, starts from Qwen3.5-9B. The caption is an editable answer region. The model denoises discrete masked tokens while the ordered pre/post images and the instruction stay visible. Only answer tokens are corrupted. At inference, CLIPLen estimates answer length from the image pair with no learned length head and no leak from the reference.
The training fix is dual-pass remasking, aimed at the clean-context gap: training usually sees gold visible tokens, inference sees the model's previous draft.
The loss is masked-token cross-entropy at weight 1.0, visible-token cross-entropy at 0.2, and confidence calibration at 0.1. Calibration punishes confident errors and timid correct predictions, because inference uses confidence to keep or rewrite a token. Curriculum timestep sampling stretches the mask ratio from local infilling to near-full masks, and with probability 0.1 uses the fully masked state that inference starts from.
Inference runs 16 confidence-guided denoising steps from a fully masked answer, then 4 mask-free polish steps. Polish is not a separately trained module. It reuses the draft-repair behavior learned in Pass 2.
Evaluation uses RSCC: 62,351 bi-temporal disaster pairs, mean reference length 72 tokens, 3,119 official test samples, compared against 11 general-purpose and RS-specific models.
| Method | ROUGE-L | METEOR | ST5-SCS |
| RSCCM 7B (strongest RS baseline) | 14.99 | 16.05 | 58.52 |
| Qwen3.5-9B (base) | 12.89 | 19.04 | 64.09 |
| InternVL3-8B | 12.76 | 15.77 | 51.84 |
| EchoChange 9B | 26.18 | 30.86 | 77.40 |
Gains over the strongest baseline on each metric are +11.19, +11.82, and +13.31. Against the untuned base, ROUGE-L moves from 12.89 to 26.18.
Diagnostic sets split revision from generation. Single-error Strict Hit is 65.34% versus 27.50% for InternVL3 and 18.68% for the base. Multi-error Strict Hit is 41.23% versus 7.23% for InternVL3. Relation errors are repaired at 95.59%, events at 81.21%, quantities at 69.64%. Masked-token accuracy on fragment recovery is 72.32% versus 37.11% for RSCCM. Clean Keep is 79.26%, below Qwen2-VL's 88.00%, but Qwen2-VL's single-error Strict Hit is 5.11%. It almost never edits.
Denoising and polishing are not interchangeable. Twenty polish steps alone reach 22.75 ROUGE-L. Sixteen denoising steps reach 25.20. Sixteen plus four polish steps reach 26.18. Mean latency falls from 5.30 s for the base Qwen model to 1.42 s, a 3.72x speed-up.
For RS interpretation and disaster briefing, a caption no longer has to be right on the first token. If an autoregressive decoder calls farmland "urban", it will keep writing around that error. EchoChange can remask the uncertain span after the rest of the draft exists. The paper also measures three behaviors separately: fill missing facts, fix wrong visible facts, keep correct ones. That is closer to how an analyst revises a draft than a single caption score.
The cost is real. Training runs two forwards on Qwen3.5-9B, on six 40 GB A100s. For short fluent captions this is heavy. For long descriptions that must keep entities, counts, and spatial relations consistent, revisable decoding is the more direct fix than scaling the same autoregressive stack. The lift comes from changing how the base model writes, not from a stronger vision encoder.
There is no standalone limitations section. The numbers already show the trade-offs. Non-target preservation (NTP) is 23.26% versus 67.53% for Qwen2-VL. Single-error CCSR is 16.15%, only slightly above the base at 14.01%. The model edits, but not locally enough, and can rewrite context that was already correct. Change-type Strict Hit is 21.87%, below InternVL3's 34.40%. Direction is 49.06%, below the base at 62.26%. Near-synonym change verbs and orientation remain weak.
Correction and recovery sets inject errors or masks. They are not drafts from human analysts. CLIPLen has no ablation; a wrong length estimate hard-caps the answer slots. All numbers are on RSCC only. Whether the 72-token setting transfers to short change captions is untested. Qualitative examples still miss fine counts: a tornado case predicts 60 buildings where the reference has 64.