Open-vocabulary semantic segmentation (OVS) of remote sensing imagery is a challenging pixel-level task requiring strong generalization and adaptation to the spatial characteristics of remote sensing data. Although existing vision-language foundation models perform well in general domains, their image-level classification design weakens the spatial priors needed for high-resolution remote sensing segmentation: structural spatial relations are degraded during deep feature transformation, and fine-grained spatial details are lost during downsampling. To address these complementary deficiencies, we propose SatOV, a training-free framework for open-vocabulary remote sensing segmentation that restores spatial priors at two stages of the representation pipeline. Specifically, Residual QQ Attention (ResQQ) extracts Query-Key self-attention from an intermediate CLIP layer and fuses it with final-layer Query-Query attention via a residual combination, restoring structural spatial priors suppressed by the final-layer representation. Spatially Modulated Upsampling (SatUp) uses the original high-resolution RGB image as spatial guidance, combining spatial feature modulation with guided cross-attention to reconstruct pixel-level textures and boundaries. Extensive experiments on DOTA, UDD, LoveDA, and Vaihingen show that SatOV consistently improves training-free OVS and achieves competitive quantitative and qualitative results against state-of-the-art methods. These results validate the effectiveness of restoring spatial priors at both the representation and spatial-resolution stages for remote sensing open-vocabulary segmentation.
Figures & tables
Figure 1: Experimental analysis of spatial covariance across Transformer depth, showing the progressive loss of spatial detail with increasing semantic abstraction.
Figure 2: Left panel illustrates the restoration of structural priors within the CLIP ViT: ResQQ extracts the Query‑Key attention from the penultimate layer as the overwritten spatial prior, and fuses it with the final‑layer Query‑Query self‑attention via residual connections, recovering the spatial covariance of the low‑resolution features. Right panel demonstrates the recovery of pixel‑level detail priors: SatUp uses the original high‑resolution RGB image as a spatial index, and applies Spatial Feature Modulation (SFT) together with cross‑attention to upsample the low‑resolution semantic features back to pixel‑level space.
Figure 3: SatUp takes low-resolution CLIP features and the original high-resolution RGB image as inputs. It encodes the image into spatial queries, applies rotary positional encoding, and performs cross-attention with SFT-modulated keys and semantic values, producing pixel-aligned high-resolution features.
Dataset
Imaging Platform
Primary Type
UDD
UAV
Urban
DOTA
Satellite
Dense objects
LoveDA
Satellite
Farmland and Urban
Vaihingen
Aerial (no blue band)
Building and House
Table 1: Overview of evaluation datasets.
Model
UDD
DOTA
LoveDA
Vaihingen
Avg.
Training-Free Methods
MaskCLIP ECCV baseline
32.4
25.1
27.8
23.4
27.2
SCLIP ECCV
38.7
29.3
30.4
29.1
31.9
ClearCLIP ECCV
41.2
31.0
31.6
35.5
34.8
LPOSS CVPR
41.8
33.9
32.4
38.0
36.5
SegEarth-OV CVPR
45.3
40.3
36.9
42.3
41.2
Table 2: Performance comparison of different methods on four remote sensing datasets. The primary comparison is conducted among training-free methods under the same zero-shot evaluation setting.
Figure 4: For the segmentation mask comparison on the DOTA dataset with densely packed objects, it can be observed that MaskCLIP predictions are largely semantically correct but completely lack spatial structure. ClearCLIP shows some improvement yet still suffers from internal holes. In contrast, SatOV accurately delineates the boundaries of densely arranged buildings and narrow roads, demonstrating the effectiveness of ResQQ and SatUp in jointly recovering spatial priors.
Figure 5: These are binary classification attention maps. The second row shows ClearCLIP outputs while the third row shows our SatOV. Significant differences arise from the resolution bottleneck, and SatOV achieves true pixel-level fine-grained prediction.
Figure 6: Visualization of feature reconstruction across different upsampling methods. It can be observed that SatUp achieves the best fine-grained reconstruction.
Model
Pre-training Data
UDD
+DINO CVPR
ImageNet
41.8
+FeatUp ICLR
ImageNet
45.3
+AnyUp ICLR
ImageNet
44.2
+SatUp Ours
Million-AID
45.9
Table 3: Performance comparison of different upsamplers on the UDD dataset. Taking DINO as the baseline, all methods yield performance improvements, yet SatUp achieves the best results for remote sensing tasks.
Configuration
DOTA
∇
Full SatOV
40.4
—
− SatUp
34.8
∇5.6
− ResQQ
32.8
∇2.0
− CLS token debiasing
30.9
∇1.9
− FFN and Residual Removing
30.0
∇0.9
− Query–Query attention
29.0
∇1.0
Table 4: Ablation study on the proposed components. Starting from the full SatOV model, we progressively remove each component to quantify its individual contribution. All models are evaluated under the same training-free, zero-shot setting on the DOTA dataset.
Figure 7: Visualization results for different ablation configurations. It can be seen that ResQQ can recover CLIP features to a certain extent, while SatUp leverages guidance from the original image to achieve full pixel-level granularity.
Remote sensing semantic segmentation is hindered by costly pixel-level annotations, motivating training-free open-vocabulary methods. Recently, the recent release of DINOv3 brings DINO.txt, which equips the standalone DINO backbone with image-text contrastive learning and thus opens up the possibility of open-vocabulary segmentation. We propose DinoSplat-OV, a training-free framework that adapts DINOv3 to remote sensing without fine-tuning or additional pretraining. Targeting the dense distribution, multi-scale nature, and large size of remote sensing imagery, we design two core modules. Its Text-aware Laplacian Propagation module de-noises patch-level predictions by combining textual semantic affinities with local visual similarity, improving regional consistency while preserving boundaries. Its Gaussian Splatting Upsampling module reconstructs pixel-level features through RGB-guided anisotropic aggregation and test-time optimization. A global-anchor sliding-window strategy further supports large-scale imagery. Experiments on UDD5, DOTA, and LoveDA demonstrate competitive or superior performance over existing training-free methods, effectively filling the gap of DINO-series models in training-free open-vocabulary segmentation and providing a viable new path for further advances in this direction.
Changhao Zhao, Haoxiang Li, Yuke Li +2
College of Resources and Environment, Huazhong Agricultural University, Wuhan 430070, China
Training-free open-vocabulary remote sensing segmentation (OVRSS), empowered by vision-language models, has emerged as a promising paradigm for achieving category-agnostic semantic understanding in remote sensing imagery. Existing approaches mainly focus on enhancing feature representations or mitigating modality discrepancies to improve patch-level prediction accuracy. However, such independent prediction schemes are fundamentally misaligned with the intrinsic characteristics of remote sensing data. In real-world applications, remote sensing scenes are typically large-scale and exhibit strong spatial as well as semantic correlations, making isolated patch-wise predictions insufficient for accurate segmentation. To address this limitation, we propose ConInfer, a context-aware inference framework for OVRSS that performs joint prediction across multiple spatial units while explicitly modeling their inter-unit semantic dependencies. By incorporating global contextual cues, our method significantly enhances segmentation consistency, robustness, and generalization in complex remote sensing environments. Extensive experiments on multiple benchmark datasets demonstrate that our approach consistently surpasses state-of-the-art per-pixel VLM-based baselines such as SegEarth-OV, achieving average improvements of 2.80% and 6.13% on open-vocabulary semantic segmentation and object extraction tasks, respectively. The implementation code is available at: https://github.com/Dog-Yang/ConInfer
Wenyang Chen, Zhanxuan Hu, Yaping Zhang +2
Yunnan Normal University · Xi’an University of Posts & Telecommunications
In this paper, we propose ReSeg-CLIP, a new training-free Open-Vocabulary Semantic Segmentation method for remote sensing data. To compensate for the problems of vision language models, such as CLIP in semantic segmentation caused by inappropriate interactions within the self-attention layers, we introduce a hierarchical scheme utilizing masks generated by SAM to constrain the interactions at multiple scales. We also present a model composition approach that averages the parameters of multiple RS-specific CLIP variants, taking advantage of a new weighting scheme that evaluates representational quality using varying text prompts. Our method achieves state-of-the-art results across three RS benchmarks without additional training.