Foundation Model-Driven Semantic Change Detection in Remote Sensing Imagery

arXiv:2602.13780 · cs.CV · Submitted 2026-02-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Foundation Model-Driven Semantic Change Detection in Remote Sensing Imagery".

Jane: Here's the gist of the paper. Remote sensing change detection (SCD) is essential for interpreting surface dynamics <ref:2602.13780#pg2> Semantic change detection (SCD) further enables pixel-level understanding of multi-class transitions yet it remains highly sensitive to pseudo-changes induced by imaging conditions <ref:2602.13780#pg6> Recent RS foundation models exhibit strong representation capabilities and their ability to extract semantically consistent features across different temporal instances is particularly critical for mitigating pseudo-changes in SCD tasks <ref:2602.13780#pg2> The authors introduce PerASCD, a robust and unified SCD framework driven by the RS foundation model PerA <ref:2602.13780#pg2> Extensive experiments on the SECOND and LandsatSCD benchmarks demonstrate that PerASCD achieves new state-of-the-art SeK scores (26.11% and 65.21%), surpassing the previous best method by 0.61% and 4.95%, respectively <ref:2602.13780#pg4> Pruning redundant reasoning tokens The paper introduces PerASCD, a robust and unified SCD framework designed to effectively bridge various RS foundation models with the SCD task <ref:2602.13780#pg2> The core innovation is the Cascaded Gated Decoder (CG-Decoder) which serves as a flexible bridge between various backbones and SCD tasks <ref:2602.13780#pg2> This decoder processes multi-scale features in a coarse‑to‑fine manner while promoting effective multi-level feature interaction and adaptive change extraction <ref:2602.13780#pg5> Furthermore, the framework proposes the Soft Semantic Consistency Loss (SSCLoss) to mitigate numerical instability often encountered during the training of large-scale models <ref:2602.13780#pg6> How the training signal is built The PerA pre-training method is a self-supervised learning approach that integrates the strengths of contrastive learning and masked image modeling <ref:2602.13780#pg2> PerA replaces random cropping in the standard contrastive pretext task of contrastive learning with random masking and incorporates learnable mask tokens into the input patches <ref:2602.13780#pg2> This design enables the model to acquire global semantic understanding from RS images via contrastive objectives while simultaneously capturing fine-grained local semantic details through pixel-level reconstruction <ref:2602.13780#pg2> The training dynamics are stabilized by introducing the Soft Semantic Consistency Loss (SSCLoss) to resolve the numerical instability issues encountered when training SCD models with SCLoss <ref:2602.13780#pg6> Where the gains come from The performance improvements are attributed to three core factors: first, the robust semantic representations provided by the PerA foundation model which significantly enhances robustness against pseudo-changes <ref:2602.13780#pg5> second, the sophisticated feature arbitration and multi-scale fusion facilitated by the CGDecoder through the dynamic gating mechanism of the CAGM <ref:2602.13780#pg5> third, the enhanced optimization stability ensured by the SSCLoss which mitigates numerical instability <ref:2602.13780#pg6> Ablation studies confirm that incorporating CAGM leads to a notable improvement of 0.79% in the SeK <ref:2602.13780#pg5> Ablation Study with Different Backbone Encoders The CG-Decoder demonstrates seamless adaptation and strong generalization across a wide range of architectures including CNNs, Transformers, and State Space Models Notably, even when equipped with a classic, lightweight CNN backbone like ResNet50 the CGDecoder achieves highly competitive results High-capacity models like Swin-L exhibit robust representation capabilities further highlighting the structural flexibility of our decoder Ablation Study on Decoder Design under Varying Training Data Scales The proposed CG-Decoder consistently demonstrates robustness against severe class imbalance in SCD datasets The well pre-trained RS foundation model plays an indispensable role in preventing pseudo-changes when the PerA foundation model is removed The framework shows exceptional transferability and data efficiency when trained with only 50% of the data achieving a SeK of 24.21% Ablation Study of Numerical Instability on LandsatSCD Dataset Numerical instability issues are strongly correlated with floating-point precision and loss function when using SCLoss under mixed-precision training Replacing SCLoss with the proposed SSCLoss significantly mitigates gradient instability The temperature parameter τ should be adjusted according to the distribution of the training dataset for optimal performance In this paper, PerASCD is a robust and unified SCD framework that synergizes the representational power of RS foundation models with a streamlined, modular Cascaded Gated Decoder (CG-Decoder) The research demonstrates that the bottleneck lies in the intelligent arbitration of multiscale semantic and change information PerASCD achieves a SeK of 65.21% on LandsatSCD surpassing Mamba-FCS by 4.95% <ref:2602.13780#pg4> The framework exhibits profound architectural flexibility and enhanced interpretability through the CAGM The SSCLoss successfully stabilizes optimization, yielding a 10.44% improvement in SeK over the baseline on LandsatSCD <ref:2602.13780#pg6> The model demonstrates exceptional data efficiency when trained with only 50% of the available data The framework proves its capability to unlock the potential of any vision encoder through its unified gating mechanism The approach provides a more precise boundary delineation and suppresses pseudo-changes in complex change regions The results validate the effectiveness of leveraging RS foundation models for practical SCD applications <ref:2602.13780#pg5> Page 1 --- <ref:2602.13780#pg6> Remote sensing change detection is essential for interpreting surface dynamics.

Tom: Pruning redundant reasoning tokens Remote sensing change detection (SCD) is essential for interpreting surface dynamics <ref:2602.13780#pg2> Semantic change detection (SCD) further enables pixel-level understanding of multi-class transitions yet it remains highly sensitive to pseudo-changes induced by imaging conditions <ref:2602.13780#pg6> Recent RS foundation models exhibit strong representation capabilities and their ability to extract semantically consistent features across different temporal instances is particularly critical for mitigating pseudo-changes in SCD tasks <ref:2602.13780#pg2> Where the gains come from The performance improvements are attributed to three core factors: first, the robust semantic representations provided by the PerA foundation model which significantly enhances robustness against pseudo-changes <ref:2602.13780#pg5> second, the sophisticated feature arbitration and multi-scale fusion facilitated by the CGDecoder through the dynamic gating mechanism of the CAGM <ref:2602.13780#pg5> third, the enhanced optimization stability ensured by the SSCLoss which mitigates numerical instability <ref:2602.13780#pg6> Where the gains come from The performance improvements are attributed to three core factors: first, the robust semantic representations provided by the PerA foundation model which significantly enhances robustness against pseudo-changes <ref:2602.13780#pg5> second, the sophisticated feature arbitration and multi-scale fusion facilitated by the CGDecoder through the dynamic gating mechanism of the CAGM <ref:2602.13780#pg5> third, the enhanced optimization stability ensured by the SSCLoss which mitigates numerical instability <ref:2602.

Title and authors: Jane: So, let's talk about who wrote this and what the title actually means for us on the ground. It’s "Foundation Model-Driven Semantic Change Detection in Remote Sensing Imagery." I mean, that sounds like a lot of jargon.

Tom: It is a mouthful, but it tells you exactly what it is: they are using these large foundation models to help with semantic change detection in remote sensing imagery. The core implication is shifting the focus from just training one specific model on one specific task to building something more adaptable.

Lu: What's interesting here is that they address three main weaknesses in the current field: sensitivity to pseudo-changes, that lack of structural adaptability, and training instability caused by category imbalance in these tasks >

The paper's summary: Tom: So, when they summarize it, they really emphasize that while foundation models have this potential for extracting semantically consistent features across different temporal instances to fight those pseudo-changes, applying them in change detection has been largely unexplored.

Jane: That’s the crucial point—the authors are saying we need to look closer at how these models handle real transitions versus just noise from the imaging process. They are looking at how foundation models can extract features that are consistent even when the environment changes.

Meng: Consistency across different temporal instances is what matters for change detection, right? If a feature looks the same in two different times but isn't actually there, that’s a false positive, and we don't want to waste compute detecting it.

Lalam: The paper points out that most current methods are limited because they rely on lightweight encoders pre-trained on natural images, and these struggle to distinguish land cover changes from things like seasonal variations or illumination shifts >

The paper's improvements: Tom: Now let's talk about what they actually *did* to fix those issues. They introduced a framework called PerASCD, which is driven by the foundation model PerA. The core innovation there is the Cascaded Gated Decoder, or CG-Decoder >

Lu: This CG-Decoder acts as this flexible bridge between various backbones and different SCD tasks. It processes multi-scale features in a coarse to fine way, which lets it promote effective multi level feature interaction and adaptive change extraction >

Jane: And they also introduced something called the Soft Semantic Consistency Loss, or SSCLoss. They say this helps mitigate the numerical instability that often pops up when you try to train these large models, which is a big deal for training anything at this scale >

Tom: The gains they claim come from three main things: first, the robust semantic representations from the PerA foundation model which makes it much more resistant to those pseudo-changes we talked about before. Second, the sophisticated feature arbitration and multi-scale fusion through that dynamic gating mechanism of the CAGM in the CGDecoder >

Lu: And third, they have this optimization stability from the SSCLoss which keeps the training trajectory much more stable when dealing with these complex systems. They even showed that incorporating CAGM leads to a measurable improvement of zero point seven nine percent in their SeK score >

Meng: I'm interested in the architectural flexibility part, because if it works well with CNNs, Transformers, and even State Space Models, that means we don't have to get locked into just one type of backbone anymore. That opens up more options for our deployment pipeline >

Jane: And they also showed some really strong data efficiency. Even when trained with only fifty percent of the available data on the SECOND dataset, they still got a SeK score of twenty-four point two one percent, which is pretty good for a change detection task >

Tom: So to wrap up on the improvements, we have this robust semantic representation from PerA fighting pseudo-changes, the CG-Decoder with its dynamic gating for multi-scale fusion, and SSCLoss for stability. It sounds like they’re tackling the whole training pipeline issue >

Conclusion: Tom: So to wrap up on this paper, the big picture is that PerASCD provides a unified framework that successfully marries the representational power of foundation models with a streamlined decoder. They achieved a SeK score of sixty-five point two one percent on LandsatSCD, surpassing Mamba-FCS by four point nine five percent >

Jane: What this means for us is that if you want to build something robust that can handle the diversity of modern remote sensing data and fight those confusing pseudo-changes, leveraging these foundation models isn't just an idea anymore; it’s showing concrete performance gains >

Lu: For me, the architectural flexibility shown here with the CG-Decoder across different backbone types is what really excites me. It proves that we can move away from rigid designs and toward systems that adapt to whatever the input representation provides >

Meng: From a production angle, if you can get high performance even with only half the data, that makes the whole system much more viable for real-world scenarios where data labeling is expensive and scarce >

Lalam: I'm pretty sure this paper proves that intelligent arbitration of multiscale semantic and change information is the key here. PerASCD successfully stabilizes optimization, yielding a ten point four four percent improvement in SeK over the baseline on LandsatSCD >

Wuhan University

cs.CV

Submitted: 2026-02-14

Updated: 2026-11-05

Comments: Accepted for publication in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) on October 30, 2026

Code: https://github.com/SathShen/PerASCD

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: Remote sensing change detection (SCD) is essential for interpreting surface dynamics <ref:2602.13780#pg2> Semantic change detection (SCD) further enables pixel-level understanding of multi-class…

Key concepts

PerASCD
A robust framework that unifies remote sensing change detection by integrating RS foundation models with a modular Cascaded Gated Decoder. It is designed to effectively bridge different model backbones and extract consistent, high-quality semantic features across temporal instances.
Cascaded Gated Decoder (CG-Decoder)
A flexible component that acts as a bridge between various backbone architectures and the SCD task. It processes multi-scale features in a coarse-to-fine manner, using dynamic gating to promote effective interaction between different levels of features for adaptive change extraction.
Soft Semantic Consistency Loss (SSCLoss)
A loss function introduced to stabilize training, particularly when using SCLoss. This loss mitigates numerical instability that often occurs during the training of large-scale models, ensuring a more stable optimization trajectory and better performance.
PerA Foundation Model
The pre-training method for the foundation model that uses random masking instead of cropping. This allows the model to learn global semantic understanding from remote sensing images while simultaneously capturing fine-grained local details through pixel-level reconstruction, enhancing robustness against pseudo-changes.

Terminology

Summary

Remote sensing change detection (SCD) is essential for interpreting surface dynamics <ref:2602. p. 013780#pg2> Semantic change detection (SCD) further enables pixel-level understanding of multi-class transitions yet it remains highly sensitive to pseudo-changes induced by imaging conditions <ref:2602. p. 0 p. 013780#pg6> Recent RS foundation models exhibit strong representation capabilities and their ability to extract semantically consistent features across different temporal instances is particularly critical for mitigating pseudo-changes in SCD tasks <ref:2602. p. 0 p. 013780#pg2> The authors introduce PerASCD, a robust and unified SCD framework driven by the RS foundation model PerA <ref:2602. p. 013780#pg2> Extensive experiments on the SECOND and LandsatSCD benchmarks demonstrate that PerASCD achieves new state-of-the-art SeK scores (26. p. 011% and 65.21%), surpassing the previous best method by 0. p. 061% and 4.95%, respectively <ref:2602. p. 013780#pg4>

Pruning redundant reasoning tokens

The paper introduces PerASCD, a robust and unified SCD framework designed to effectively bridge various RS foundation models with the SCD task <ref:2602.13780#pg2> The core innovation is the Cascaded Gated Decoder (CG-Decoder) which serves as a flexible bridge between various backbones and SCD tasks <ref:2602. p. 013780#pg2> This decoder processes multi-scale features in a coarse‑to‑fine manner while promoting effective multi-level feature interaction and adaptive change extraction <ref:2602. p. 013780#pg5> Furthermore, the framework proposes the Soft Semantic Consistency Loss (SSCLoss) to mitigate numerical instability often encountered during the training of large-scale models <ref:2602. p. 013780#pg6>

How the training signal is built

The PerA pre-training method is a self-supervised learning approach that integrates the strengths of contrastive learning and masked image modeling <ref:2602.13780#pg2> PerA replaces random cropping in the standard contrastive pretext task of contrastive learning with random masking and incorporates learnable mask tokens into the input patches <ref:2602. p. 313780#pg2> This design enables the model to acquire global semantic understanding from RS images via contrastive objectives while simultaneously capturing fine-grained local semantic details through pixel-level reconstruction <ref:2602. p. 313780#pg2> The training dynamics are stabilized by introducing the Soft Semantic Consistency Loss (SSCLoss) to resolve the numerical instability issues encountered when training SCD models with SCLoss <ref:2602. p. 113780#pg6>

Where the gains come from

The performance improvements are attributed to three core factors: first, the robust semantic representations provided by the PerA foundation model which significantly enhances robustness against pseudo-changes <ref:2602.13780#pg5> second, the sophisticated feature arbitration and multi-scale fusion facilitated by the CGDecoder through the dynamic gating mechanism of the CAGM <ref:2602. p. 8 p. 8 p. 813780#pg5> third, the enhanced optimization stability ensured by the SSCLoss which mitigates numerical instability <ref:2602. p. 8 p. 8 p. 813780#pg6> Ablation studies confirm that incorporating CAGM leads to a notable improvement of 0.79% in the SeK <ref:2602. p. 013780#pg5>

Ablation Study with Different Backbone Encoders

The CG-Decoder demonstrates seamless adaptation and strong generalization across a wide range of architectures including CNNs, Transformers, and State Space Models Notably, even when equipped with a classic, lightweight CNN backbone like ResNet50 the CGDecoder achieves highly competitive results High-capacity models like Swin-L exhibit robust representation capabilities further highlighting the structural flexibility of our decoder

Ablation Study on Decoder Design under Varying Training Data Scales

The proposed CG-Decoder consistently demonstrates robustness against severe class imbalance in SCD datasets The well pre-trained RS foundation model plays an indispensable role in preventing pseudo-changes when the PerA foundation model is removed The framework shows exceptional transferability and data efficiency when trained with only 50% of the data achieving a SeK of 24.21%

Ablation Study of Numerical Instability on LandsatSCD Dataset

Numerical instability issues are strongly correlated with floating-point precision and loss function when using SCLoss under mixed-precision training Replacing SCLoss with the proposed SSCLoss significantly mitigates gradient instability The temperature parameter τ should be adjusted according to the distribution of the training dataset for optimal performance

Conclusion and Discussion

In this paper, PerASCD is a robust and unified SCD framework that synergizes the representational power of RS foundation models with a streamlined, modular Cascaded Gated Decoder (CG-Decoder) The research demonstrates that the bottleneck lies in the intelligent arbitration of multiscale semantic and change information PerASCD achieves a SeK of 65.21% on LandsatSCD surpassing Mamba-FCS by 4. p. 895% <ref:2602.13780#pg4> The framework exhibits profound architectural flexibility and enhanced interpretability through the CAGM The SSCLoss successfully stabilizes optimization, yielding a 10. p. 844% improvement in SeK over the baseline on LandsatSCD <ref:2602. p. 113780#pg6> The model demonstrates exceptional data efficiency when trained with only 50% of the available data The framework proves its capability to unlock the potential of any vision encoder through its unified gating mechanism The approach provides a more precise boundary delineation and suppresses pseudo-changes in complex change regions The results validate the effectiveness of leveraging RS foundation models for practical SCD applications <ref:2602. p. 113780#pg5>

--- Page 1 ---

<ref:2602.13780#pg6> Remote sensing change detection is essential for interpreting surface dynamics. p. 0

Remote sensing change detection (SCD) is essential for interpreting surface dynamics <ref:2602.13780#pg2> Semantic change detection (SCD) further enables pixel-level understanding of multi-class transitions yet it remains highly sensitive to pseudo-changes induced by imaging conditions <ref:2602.13780#pg6> Recent RS foundation models exhibit strong representation capabilities and their ability to extract semantically consistent features across different temporal instances is particularly critical for mitigating pseudo-changes in SCD tasks <ref:2602.13780#pg2>

The performance improvements are attributed to three core factors: first, the robust semantic representations provided by the PerA foundation model which significantly enhances robustness against pseudo-changes <ref:2602.13780#pg5> second, the sophisticated feature arbitration and multi-scale fusion facilitated by the CGDecoder through the dynamic gating mechanism of the CAGM <ref:2602.13780#pg5> third, the enhanced optimization stability ensured by the SSCLoss which mitigates numerical instability <ref:2602.13780#pg6>

The performance improvements are attributed to three core factors: first, the robust semantic representations provided by the PerA foundation model which significantly enhances robustness against pseudo-changes <ref:2602.13780#pg5> second, the sophisticated feature arbitration and multi-scale fusion facilitated by the CGDecoder through the dynamic gating mechanism of the CAGM <ref:2602.13780#pg5> third, the enhanced optimization stability ensured by the SSCLoss which mitigates numerical instability <ref:2602.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved Robustness Against Pseudo-changes: The system can mitigate pseudo-changes induced by imaging conditions by leveraging foundation model representations that are temporally invariant, which is achieved through the PerA pre-training method.

  2. Enhanced Architectural Flexibility: The system gains modularity through the Cascaded Gated Decoder, allowing it to adapt to diverse and hierarchical multi-scale feature representations produced by various backbones, such as CNNs, Transformers, and State Space Models.

  3. Dynamic Feature Arbitration: The Change-Aware Gating Module can intelligently arbitrate change information by learning global and local weighted maps, which allows it to suppress confusing non-change objects susceptible to environmental variations without rigid supervision.

  4. Numerical Stability in Training: The introduction of the Soft Semantic Consistency Loss (SSCLoss) resolves numerical instability encountered during training, providing a more stable training trajectory for high-capacity models.

  5. Improved Data Efficiency: The framework demonstrates superior data efficiency by achieving state-of-the-art performance when trained with only 50% of the available data on the SECOND dataset.

Sources

Related papers