XuanJian-DINOSAR: A SAR Foundation Model Based on Global and Local Semantic Consistency

arXiv:2611.06105 · cs.CV · Submitted 2026-11-04 · 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: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "XuanJian-DINOSAR: A SAR Foundation Model Based on Global and Local Semantic Consistency".

Jane: Here's the gist of the paper. Self-supervised pre-training on synthetic aperture radar (SAR) imagery provides transferable representations for classification and detection, yet representation stability remains constrained by the imaging-domain gap between optical and SAR data and by speckle-induced perturbations The resulting framework combines DINO global semantic self-distillation, iBOT local masked modeling, and KoLeo feature-distribution regularization to learn general-purpose representations with both global semantic discrimination and local structural perception Pruning redundant reasoning tokens XuanJianDINOSAR is a SAR foundation model based on global and local semantic consistency It starts from an optically pre-trained DINOv2 model, self-supervised domain adaptation on multisource SAR imagery transfers its global semantic discrimination and local structural perception to the SAR domain The framework retains DINO global semantic selfdistillation, iBOT local masked modeling, and KoLeo featuredistribution regularization

9: to learn general-purpose representations with global semantic discrimination and local structural perception How the training signal is built The framework involves a conventional multi-crop self-distillation pathway and a shared-geometry weak–strong consistency pathway The latter reuses the global crop and changes the image-intensity statistics under an identical geometric layout, explicitly constraining teacher and student patch features at corresponding spatial locations The local-consistency loss is averaged over all corresponding patches in the shared-geometry weak–strong views The final training objective is the weighted combination of four losses: L DINO DINO iBOT iBOT KoLeo KoLeo LC LC = λ L λ L λ L λ L Where the gains come from The local-consistency branch establishes an explicit position-wise feature constraint between weak and strong views sharing the same geometry This objective complements the cross-region and cross-scale semantic consistency learned by DINOv2 through multi-crop training and jointly exploits cross-view semantic relations and local spatial correspondence The local consistency learning (LC) branch significantly enhances the model's robustness to variations in imaging conditions, such as pitch angle, azimuth angle, and polarization state Downstream performance validation Extensive experiments on multiple downstream classification and detection datasets show that XuanJianDINOSAR achieves state-of-the-art performance on most evaluation metrics For instance, it achieves 96.70% accuracy on the SOC-50 classification task and an mAP of 39.90 on RSAR detection The results show that combining multi-crop cross-view learning with same-location local consistency improves the adaptation of pre-trained representations to SAR imaging characteristics and downstream task variation Ablation studies on classification Table XI provides a detailed comparison of the classification performance of three sets of pre-trained weights on the SOC50 task, all of which adopt the same classification head and downstream fine-tuning strategy The complete model, further incorporating the local consistency learning (LC) branch, achieves an accuracy of 96.70% This result demonstrates the independent contribution of the LC branch in improving representation quality and validates its effectiveness in enforcing local feature consistency under speckle and intensity perturbations Ablation studies on detection On the XuanJian-RSAR detection benchmark, the complete model achieves an mAP of 28.76, significantly surpassing the ResNet-50 baseline of 25.35 This demonstrates that the proposed pre-training scheme also possesses superior representation transfer capability in the rotated object detection scenario The iBOT local masked prediction mechanism enhances the model's perception of local scattering structures, while the LC branch enforces feature directional consistency at the same spatial position under intensity perturbations Qualitative results and representation analysis Figure 3 shows the t-SNE feature distributions of SOC-50 class targets during classification inference It can be observed that, after LoRA fine-tuning, the backbone features already exhibit strong classification capability Figure 4 shows the input images containing targets with background and their corresponding backbone network attention heatmaps It can be seen that, for input images containing complex backgrounds, the backbone accurately focuses its attention on the target regions, effectively suppressing interference from the surrounding background Figure 5 presents qualitative comparisons between the proposed method and mainstream baselines under typical SAR imaging conditions It can be clearly observed that the proposed method exhibits significant advantages in distinguishing densely arranged targets, detecting small-scale objects, and suppressing background interference The results show that the proposed model outperforms the corresponding RoI Transformer baseline on both oriented-object detection benchmarks The detection results confirm the superiority of the universal encoder based on the ViT-B/14 backbone in extracting spatial features of rotated objects, particularly through the synergistic effect of the iBOT local masked prediction mechanism and the local consistency learning (LC) branch The LC branch effectively mitigates the interference of SAR-specific speckle noise on object localization by enforcing feature directional consistency at the same spatial position under intensity perturbations XuanJian-DINOSAR provides one implementation of this principle and offers a reference for future SAR foundation models and generic representation learning from multi-source remote-sensing data Transferable SAR representations must establish stable highlevel semantic relations across regions and scales, preserve sensitivity to local structural information, and reduce the influence of SAR-specific radiometric perturbations on feature space XuanJian-DINOSAR provides one implementation of this principle and offers a reference for future SAR foundation models and generic representation learning from multi-source remote-sensing data The results suggest that combining multi-crop cross-view learning with same-location local consistency improves the adaptation of pre-trained representations to SAR imaging characteristics and downstream task variation XuanJianDINOSAR achieves 96.70% accuracy on the SOC-50 classification task and an mAP of 39.

Title and authors: Tom: So, the paper calls this "XuanJian-DINOSAR: A SAR Foundation Model Based on Global and Local Semantic Consistency." It’s basically saying that we need a model that understands the big picture of an image while also being very sensitive to local details, even when the image quality is poor due to speckle.

Jane: That sounds like a lot of moving parts. Can you break down what they mean by "global and local semantic consistency" in plain English?

Lu: They’re saying that the model needs to learn consistent patterns across different regions and scales, which is what DINOv2 achieves through multi-crop training, but they are adding something new to handle those perturbations.

Meng: So it’s like teaching the model how things relate globally, but also making sure that if you look at two patches right next to each other, their features stay consistent even if one patch has a lot of speckle noise. That sounds complicated to train at scale.

Tom: It is complex because they are combining several mechanisms. They start with DINOv2's global distillation and add this local consistency branch, which they call the shared-geometry weak–strong local consistency branch, to explicitly enforce that feature constraints between corresponding patches under intensity variation and speckle perturbations.

Jane: So the core idea is to use DINOv2 for the broad semantic learning, but then fine-tune it with a specific mechanism designed to handle SAR's noise issues locally. That makes sense for a domain adaptation task.

Lalam: I think what’s powerful is that they aren't just tacking on a fix; they are training this new local consistency branch jointly with the original cross-region and cross-scale objectives from DINOv2, which should help it integrate smoothly.

The paper's summary: Tom: Okay, so the summary says they developed XuanJian-DINOSAR as a self-supervised domain adaptation framework specifically for SAR imagery. They use the image-level and patch-level learning mechanisms from DINOv2 on unlabeled SAR data to create this reusable backbone that has both global semantic discrimination and local structural perception.

Jane: That’s the high level. So they are taking a model that knows how to see things generally, like DINOv2 does, but adapting its knowledge specifically for SAR imagery by using multi-source unlabeled data.

Lu: They introduce the shared-geometry weak–strong local consistency branch as their second major contribution. This branch explicitly promotes local semantic consistency under intensity variation and multiplicative speckle perturbations while being optimized alongside the original cross-region and cross-scale multicrop objectives of DINOv2.

Meng: So they are using this new constraint to force feature stability right at corresponding spatial locations, which is key because speckle noise is localized. It’s like giving the model a strict rule about what features should look like next to each other spatially.

Tom: And then they validate this by running extensive evaluations on several downstream tasks, including XuanJian-RSAR and their oriented fine-grained target detection dataset, where they claim state-of-the-art results on most reported metrics.

Jane: So the summary boils down to: use DINOv2 for global learning, add a local consistency branch to handle SAR noise specifically, and then show it performs better on real tasks than other models.

Lalam: It sounds like they are successfully addressing the paper's initial problem—the instability caused by speckle—by making the model learn both globally and locally robustly.

The paper's improvements: Tom: The improvements section highlights a few key things. First, they detail how the training signal is built: they use a conventional multi-crop self-distillation pathway alongside this shared-geometry weak–strong consistency pathway. That latter pathway reuses the global crop but changes the image intensity statistics under an identical geometric layout, explicitly constraining teacher and student patch features at corresponding spatial locations.

Jane: That sounds like a very clever way to generate training signals without needing extra labels for every single patch constraint. They are using geometric layout to create controlled variations in intensity for consistency testing.

Lu: The local-consistency loss is averaged over all corresponding patches in the shared-geometry weak–strong views, and the final training objective is a weighted combination of four losses: L DINO, iBOT, KoLeo, and KoLeo LC. That’s a lot of terms being balanced together.

Meng: So they are weighting DINO for global stuff, iBOT for local masked prediction, KoLeo for feature distribution regularization to keep features in a good space, and then the LC loss specifically to enforce that spatial consistency under intensity variation and speckle perturbations.

Tom: The gains come from this local-consistency branch which establishes an explicit position-wise feature constraint between weak and strong views sharing the same geometry. This objective complements the cross-region and cross-scale semantic consistency learned by DINOv2 through multi-crop training, jointly exploiting both cross-view semantic relations and local spatial correspondence.

Jane: So, it’s not just one thing working; they are using the local branch to enhance robustness against variations like pitch angle, azimuth angle, and polarization state in the SAR imagery. That shows they are targeting real imaging challenges.

Lalam: The ablation studies confirm this is important because when you look at the classification performance on SOC50, that complete model with all those losses achieves ninety-six point seven zero percent accuracy, showing the local consistency learning branch has a measurable contribution to representation quality and its effectiveness under speckle and intensity perturbations.

Conclusion: Tom: To wrap up, XuanJian-DINOSAR provides one implementation of this principle for SAR foundation models and generic representation learning from multi-source remote sensing data. The results suggest that combining multi-crop cross-view learning with same-location local consistency improves the adaptation of pre-trained representations to SAR imaging characteristics and downstream task variation.

Jane: So, the main implication is that you can create these general representations for SAR by focusing on both global relations across regions and ensuring those local features are stable even when there are noise or intensity changes in the data.

Lu: I think it sets a reference point for future work because it shows how to successfully combine the image-level and patch-level learning of DINOv2 with a specific, geometrically constrained branch to solve domain adaptation challenges effectively.

Meng: For us at startup, what this means is that if we want to deploy a model reliably in SAR sensing, we can build it on these foundation models instead of training everything from scratch for every single new task. It lowers the barrier for creating robust detection systems.

Lalam: I’m confident in what they did because the results are strong: ninety-six point seven zero percent accuracy on the SOC50 classification task and an mAP of thirty-nine point nine zero on RSAR detection, which significantly surpasses baselines like ResNet-fifty for detection.

Tom: That's a solid summary of what XuanJian-DINOSAR achieved. It really shows that by being careful about both global structure and local perturbation stability, you get much better results when moving from optical to SAR imagery.

Chunxiao Wu, Huixin Huang, Jialei Pan, Qiang Liang, Tao Wu, Cheng Wu, Long Huang, Yuanji Li, Qiang Cheng, Aifang Liu,  Linghao Xia

Nanjing Institute of Electronic Technology and National Key Laboratory of Radar Detection and Sensing

cs.CV

Submitted: 2026-11-04

Updated: 2026-11-04

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

Importance score: 82/100

The gist: Self-supervised pre-training on synthetic aperture radar (SAR) imagery provides transferable representations for classification and detection…

Key concepts

DINO Global Semantic Self-Distillation
This technique uses a DINOv2 model trained on optical images to transfer its global understanding of objects to SAR data. It ensures that the learned representations maintain high-level semantic relationships across different parts of the image, allowing the model to recognize objects based on their overall meaning rather than just local pixel patterns.
iBOT Local Masked Modeling
iBOT is a mechanism used for local masked modeling that helps the model perceive fine details and scattering structures within SAR images. It allows the network to predict missing parts of a patch by leveraging the context of surrounding pixels, which is essential for accurately detecting small-scale objects and understanding local scattering patterns.
Local Consistency (LC) Branch
This branch introduces an explicit constraint that forces feature tokens at the exact same spatial location to be consistent between different views or conditions. It is trained to remain stable even when intensity statistics change due to SAR-specific noise like speckle, thereby enhancing the model's robustness against imaging perturbations.
Multi-crop Self-Distillation Pathway
This is a conventional training method that uses multiple crops of the same image for self-distillation. It helps establish cross-region and cross-scale semantic consistency by comparing features from different sections of the same image, ensuring the model learns coherent representations across various spatial scales.

Terminology

Summary

Self-supervised pre-training on synthetic aperture radar (SAR) imagery provides transferable representations for classification and detection, yet representation stability remains constrained by the imaging-domain gap between optical and SAR data and by speckle-induced perturbations The resulting framework combines DINO global semantic self-distillation, iBOT local masked modeling, and KoLeo feature-distribution regularization to learn general-purpose representations with both global semantic discrimination and local structural perception

Pruning redundant reasoning tokens

XuanJianDINOSAR is a SAR foundation model based on global and local semantic consistency It starts from an optically pre-trained DINOv2 model, self-supervised domain adaptation on multisource SAR imagery transfers its global semantic discrimination and local structural perception to the SAR domain The framework retains DINO global semantic selfdistillation, iBOT local masked modeling, and KoLeo featuredistribution regularization [9] to learn general-purpose representations with global semantic discrimination and local structural perception

How the training signal is built

The framework involves a conventional multi-crop self-distillation pathway and a shared-geometry weak–strong consistency pathway The latter reuses the global crop and changes the image-intensity statistics under an identical geometric layout, explicitly constraining teacher and student patch features at corresponding spatial locations The local-consistency loss is averaged over all corresponding patches in the shared-geometry weak–strong views The final training objective is the weighted combination of four losses: L DINO DINO iBOT iBOT KoLeo KoLeo LC LC = λ L λ L λ L λ L

Where the gains come from

The local-consistency branch establishes an explicit position-wise feature constraint between weak and strong views sharing the same geometry This objective complements the cross-region and cross-scale semantic consistency learned by DINOv2 through multi-crop training and jointly exploits cross-view semantic relations and local spatial correspondence The local consistency learning (LC) branch significantly enhances the model's robustness to variations in imaging conditions, such as pitch angle, azimuth angle, and polarization state

Downstream performance validation

Extensive experiments on multiple downstream classification and detection datasets show that XuanJianDINOSAR achieves state-of-the-art performance on most evaluation metrics For instance, it achieves 96.70% accuracy on the SOC-50 classification task and an mAP of 39. p. 9 p. 990 on RSAR detection The results show that combining multi-crop cross-view learning with same-location local consistency improves the adaptation of pre-trained representations to SAR imaging characteristics and downstream task variation

Ablation studies on classification

Table XI provides a detailed comparison of the classification performance of three sets of pre-trained weights on the SOC50 task, all of which adopt the same classification head and downstream fine-tuning strategy The complete model, further incorporating the local consistency learning (LC) branch, achieves an accuracy of 96.70% This result demonstrates the independent contribution of the LC branch in improving representation quality and validates its effectiveness in enforcing local feature consistency under speckle and intensity perturbations

Ablation studies on detection

On the XuanJian-RSAR detection benchmark, the complete model achieves an mAP of 28.76, significantly surpassing the ResNet-50 baseline of 25.35 This demonstrates that the proposed pre-training scheme also possesses superior representation transfer capability in the rotated object detection scenario The iBOT local masked prediction mechanism enhances the model's perception of local scattering structures, while the LC branch enforces feature directional consistency at the same spatial position under intensity perturbations

Qualitative results and representation analysis

Figure 3 shows the t-SNE feature distributions of SOC-50 class targets during classification inference It can be observed that, after LoRA fine-tuning, the backbone features already exhibit strong classification capability Figure 4 shows the input images containing targets with background and their corresponding backbone network attention heatmaps It can be seen that, for input images containing complex backgrounds, the backbone accurately focuses its attention on the target regions, effectively suppressing interference from the surrounding background Figure 5 presents qualitative comparisons between the proposed method and mainstream baselines under typical SAR imaging conditions It can be clearly observed that the proposed method exhibits significant advantages in distinguishing densely arranged targets, detecting small-scale objects, and suppressing background interference The results show that the proposed model outperforms the corresponding RoI Transformer baseline on both oriented-object detection benchmarks The detection results confirm the superiority of the universal encoder based on the ViT-B/14 backbone in extracting spatial features of rotated objects, particularly through the synergistic effect of the iBOT local masked prediction mechanism and the local consistency learning (LC) branch The LC branch effectively mitigates the interference of SAR-specific speckle noise on object localization by enforcing feature directional consistency at the same spatial position under intensity perturbations

Conclusion

XuanJian-DINOSAR provides one implementation of this principle and offers a reference for future SAR foundation models and generic representation learning from multi-source remote-sensing data Transferable SAR representations must establish stable highlevel semantic relations across regions and scales, preserve sensitivity to local structural information, and reduce the influence of SAR-specific radiometric perturbations on feature space XuanJian-DINOSAR provides one implementation of this principle and offers a reference for future SAR foundation models and generic representation learning from multi-source remote-sensing data The results suggest that combining multi-crop cross-view learning with same-location local consistency improves the adaptation of pre-trained representations to SAR imaging characteristics and downstream task variation XuanJianDINOSAR achieves 96.70% accuracy on the SOC-50 classification task and an mAP of 39.

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 representation learning by jointly exploiting cross-view semantic relations and local perturbation stability through XuanJian-DINOSAR. This enables models to learn global and local semantic consistency jointly, allowing them to preserve cross-region and cross-scale semantic consistency learned by DINOv2 through multi-crop training while simultaneously ensuring local representations to remain consistent under radiometric perturbations.

  2. Enhanced robustness against SAR-specific noise by introducing a local consistency branch that enforces constraints on patch tokens at corresponding spatial locations. This branch is trained to ensure local feature consistency under the specified speckle and intensity perturbations, which mitigates the interference of SAR-specific multiplicative speckle noise on target localization.

  3. Superior fine-grained classification and object detection performance across varied imaging conditions by transferring representations from an optically pre-trained model to SAR imagery. The framework achieves highest accuracy on all five EOC tasks and outperforms baselines in both classification (96.70% accuracy) and object detection (mAP of 39.90 on RSAR).

  4. Better localization of small-scale targets by enhancing the perception of local scattering structures through the iBOT local masked prediction mechanism. This mechanism allows the model to recover local representations from visible context, which is crucial for accurately detecting small-scale objects and suppressing background interference.

Sources

Related papers