acceptodds
Under review as a conference paper at ICLR 2027

Discover Changes Gradually: LLM-Guided Progressive Network for Remote Sensing Change Detection

Abstract

Existing remote sensing change detection (RSCD) method usually adopts one-shot prediction strategy, which is vulnerable to complex background and illumination variation, resulting in missed or false detection especially around ambiguous change boundary. Human visual perception follows a coarse-to-fine analysis process, progressively focusing from salient region to fine-grained detail. Inspired by this process, this paper formulates RSCD as a gradual reasoning process that progressively discovers change region. First, task-oriented question prompt is designed to guide large language model (LLM) to generate high-level semantic prior from bi-temporal image, including change region, boundary characteristic, and potential missed or false detection, etc. Additionally, shared dictionary–coefficient representation learning is utilized to project bi-temporal image into unified representation space, where coefficient-level visual difference information is extracted. Furthermore, text–image consistency is explored to transform global semantic prior into spatial-aware response, thereby adaptively enhancing visual difference feature.Subsequently, Segment Anything Model (SAM) is employed to achieve progressive change discovery strategy. It jointly considers change probability, prediction reliability, and remaining search region to generate multi-stage high-quality prompt, guiding SAM to progressively focus on unresolved change region. Moreover, the soft-union fusion is adopted to obtain final change result. Experiments on multiple datasets demonstrate that the proposed model achieves superior performance and exhibits stronger effectiveness in complex change identification.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote Sensing Change Detection, Progressive Network, LLM-Guided, high-quality prompt

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_gswgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LGPcd---a remote sensing change detection framework that combines LLM-derived semantic priors (text), shared dictionary representations, and progressive SAM-based refinement to recover accurate results. Evaluations on LEVIR-CD, SYSU-CD, and SECOND show superior performance compared with the reported baselines.

Critical strengths

Incorporating LLM knowledge as textual semantic priors is intuitive and reasonable, and SAM-based refinement is technically well motivated. Both components yield substantial performance gains across the evaluated remote sensing change detection benchmarks (e.g., LEVIR-CD, SYSU-CD, and SECOND).

However, the current manuscript does not sufficiently establish the necessity and effectiveness of the proposed component, for the weaknesses outlined below.

Critical weaknesses

I outline the weaknesses that limit the paper’s completeness; listed in descending order of importance.

[Missing Ablation on Text-Guided Enhancement] The authors claim that the semantic information in the text descriptions can improve performance by providing a global semantic prior (L176). While this approach is reasonable, the authors do not provide a concrete ablation that isolates the contribution of text guidance (which is important given that the paper’s main contribution lies in “LLM-Guided” and “Progressive” remote sensing change detection). Similarly, I believe "Progressive" ablation should include an N=1 baseline. This comparison is necessary to distinguish the benefit of introducing progressive refinement from that of not using refinement at all.

[Missing SAM-based baselines and Efficiency Analysis] Table 1 excludes the closely related SAM-based methods discussed in Related Work. These comparisons are particularly necessary because the proposed approach also relies on SAM. Moreover, the proposed method applies SAM multiple times through progressive processing (where inference cost can be substantially higher than that of conventional change detection models). Therefore, an efficiency analysis is needed to assess the effects of the SAM model size and the number of progressive iterations on inference time, computational cost, and memory usage, ideally in comparison with other SAM-based methods.

[Insufficient Detail and Poor Reproducibility] The paper misses essential details---including the LLM version/settings, the text encoder used to encode the LLM-generated text, the SAM variant, which components are trainable or frozen, and a detailed description of the model architecture (e.g., dictionary, coefficient encoder). It also lacks a reproducibility statement; in my point of view, the information provided in the current manuscript is highly insufficient to reproduce the reported quantitative results. This constitutes a critical weakness that should not be overlooked.

[Missing Analysis on Error Accumulation] Recent benchmarks show that even frontier Multimodal LLMs (MLLMs) can incorrectly describe changes in remote sensing image pairs [1]. However, the paper assumes the LLM’s outputs are always reliable and does not examine whether errors propagate to the final predictions or are corrected through progressive refinement. To ensure the paper's completeness, analysis of the erroneous LLM outputs is essential.

[Formatting and Typo] Current citations in the Introduction appear without the parentheses required by the ICLR citation guidelines; authors should use \citep{} instead of \cite{}. Figure 1(a) also contains “Trantional,” which should be corrected to “Traditional.”

[1] C3-Bench: A Context-Aware Change Captioning Benchmark

Questions for authors

Please refer to the Weakness section.

I currently lean toward a Clear Rejection in its current form. Yet, if the authors can satisfactorily clarify and address the weaknesses outlined above, I would be happy to reconsider the assessment during the discussion period.

Reviewer_h4ggOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LGPcd, an LLM-guided Progressive Change Discovery Network. The authors employ an LLM to generate semantic descriptions of change regions, boundaries, and potential detection errors, thereby supplementing visual detection with high-level semantic information.

Critical strengths
  1. The model automatically generates prompts for each stage based on remaining unexplored regions, directing attention to areas where changes have not yet been detected.

  2. The proposed method achieves higher F1 scores and mIoU values than all the compared methods listed in the tables across three datasets, with a particularly notable improvement on SYSU-CD.

Critical weaknesses
  1. During a single inference pass with a fixed decision threshold, the cumulative prediction can only increase, while high-confidence regions are further excluded from subsequent searches. Consequently, later stages cannot correct false positives introduced at earlier stages. The paper does not adequately justify why such a design is adopted for a task involving both missed detections and false positives. The capabilities and limitations of the proposed strategy in recovering missed detections versus correcting erroneous predictions need to be clearly distinguished. The evidence supporting the progressive strategy is also limited. Table 3 only reports results for N = 2, 3, and 4, without including N = 0 or N = 1, although the main text claims that performance progressively improves as the number of stages increases from 1 to 3. Among the reported results, the F1 score increases only marginally from 81.09 to 81.15 when N increases from 2 to 3, and no variance across multiple runs is provided.

  2. Insufficient ablation studies prevent a clear assessment of the independent contributions of the core modules. The existing experiments mainly investigate loss weights and the number of progressive stages, while essential comparisons, including variants without text guidance, without the shared dictionary, and with a single-stage prediction scheme using the same backbone, are missing. As a result, it remains unclear how much improvement can be attributed to the language information and the progressive mechanism individually, or whether these components are indeed necessary.

Questions for authors
  1. Why adopt a residual completion strategy that cannot revise earlier predictions? While this design helps avoid redundant detections, it may preserve false positives introduced at early stages. Please explain the rationale and trade-offs behind this choice, and report the stage-wise changes in false positives and false negatives within the same inference pass.

2.Does Equation (5) intentionally retain only the entropy contribution of the change class? What is the rationale behind this formulation? How would the results be affected by using the complete binary entropy or simply removing the uncertainty term from Equation (8)?

  1. Please specify the configurations of the LLM, text encoder, and SAM, including their freezing or fine-tuning strategies, as well as the dimensions of Tv and FΔ and how they are aligned. What are the values of K and τj at each stage, and what rules are used to generate box prompts from candidate points?

  2. How much does each of the three components—language guidance, the shared dictionary, and progressive refinement—contribute to the overall performance? Please provide controlled ablation studies without text guidance, with mismatched text descriptions, without the shared dictionary, and with single-stage SAM using the same backbone to isolate the contribution of each component. In particular, it is essential to demonstrate that the performance gains arise from image-specific semantic content.

  3. Are the relatively small performance gains statistically stable? The F1 improvements on LEVIR-CD and SECOND are only 0.17 and 0.30 percentage points, respectively. Please report the mean and standard deviation across multiple independent runs to determine whether these improvements exceed normal training variability.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.