Scale-Aware UAV-to-Satellite Cross-View Geo-Localization: A Semantic Geometric Approach
Authors: Yibin Ye, Shuo Chen, Kun Wang, Xiaokai Song, Jisheng Dang, Qifeng Yu, Xichao Teng, Zhang Li
Organizations: College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China · Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation, Changsha 410073, China · School of Information Science and Engineering, Lanzhou University, Lanzhou 730000, China
Cross-View Geo-Localization (CVGL) between UAV imagery and satellite images plays a crucial role in target localization and UAV self-positioning. However, most existing methods rely on the idealized assumption of scale consistency between UAV queries and satellite galleries, overlooking the severe scale ambiguity commonly encountered in real-world scenarios. This discrepancy leads to field-of-view misalignment and feature mismatch, significantly degrading CVGL robustness. To address this issue, we propose a geometric framework that recovers the absolute metric scale from monocular UAV images using semantic anchors. Specifically, small vehicles (SVs), characterized by relatively stable prior size distributions and high detectability, are exploited as metric references. A Decoupled Stereoscopic Projection Model is introduced to estimate the absolute image scale from these semantic targets. By decomposing vehicle dimensions into radial and tangential components, the model compensates for perspective distortions in 2D detections of 3D vehicles, enabling more accurate scale estimation. To further reduce intra-class size variation and detection noise, a dual-dimension fusion strategy with Interquartile Range (IQR)-based robust aggregation is employed. The estimated global scale is then used as a physical constraint for scale-adaptive satellite image cropping, improving UAV-to-satellite feature alignment. Experiments on augmented DenseUAV and UAV-VisLoc datasets demonstrate that the proposed method significantly improves CVGL robustness under unknown UAV image scales. Additionally, the framework shows strong potential for downstream applications such as passive UAV altitude estimation and 3D model scale recovery.
Figures & tables
Fig. 1 : Scale comparison in UAV-to-Satellite CVGL: (a) For existing datasets (DenseUAV, GTA-UAV, SUES-200, University-1652), UAV image scales (marked by arrows) are consistent with satellite images (purple arrow, normalized to 1.0), with maximum difference not exceeding a factor of 2; (b) Under unknown scale, imprecise satellite cropping causes huge scale/FOV discrepancies between UAV images and satellite crops.
Fig. 2 : Overview of the proposed Scale-aware CVGL framework. The method first utilizes oriented bounding boxes of small vehicles as semantic anchors. Notably, off-center targets (orange frame) exhibit significant stereoscopic effects, necessitating the incorporation of vehicle height into the modeling. To address this, we employ a decoupled stereoscopic projection model combined with statistical dimension distributions to estimate the single-instance absolute scale. Finally, robust global scale estimation is conducted to guide the scale-adaptive cropping of satellite imagery, ensuring robust CVGL.
Fig. 3 : DOTA dataset object category analysis. (a) Normalized length (blue) and width (orange) distributions. (b) Average instances per image frequency.
Dataset
Relative Altitude
Viewpoint
Scenes
Platform
UAV Images
f
Main Limitations
DenseUAV [ 7 ]
80m/90m/100m
Nadir
Urban (Campus)
Multi-rotor
9,099
≈ 1000 px
Fixed satellite crops
UAV-VisLoc [ 16 ]
300-600m
Near-Nadir
Urban/Town/Wild
Fixed Wing/Multi-rotor
6,742
≈ 4000 px
Absolute altitude only; Noisy poses;
TABLE I : Comparison of dataset characteristics. f denotes the equivalent focal length in pixels.
Fig. 4 : Overview of the refined datasets. (a) DenseUAV+: expanded satellite maps enabling scale-adaptive cropping at different altitudes. The blue dots denote the UAV trajectories. (b) UAV-VisLoc+: SfM-based refinement producing ortho-map/DSM and refined poses for accurate relative altitude.
Fig. 5 : Sensitivity to scale mismatch. Success Rate (SR, %) under different relative altitude ratios δ in H~=H(1+δ) (equivalently, relative scale mismatch). Each subfigure contains two plots: left reports results on queries where our scale estimator is applicable (with sufficient semantic anchors, e.g., τconf≥0.5 and N≥5 ), while right reports results on all queries (including those without valid scale estimates). The green vertical line indicates the ground-truth scale ( δ=0 ). The orange band denotes the distribution of our estimated scale errors, visualized as μ±σ of δ (computed on the estimable subset), showing that most estimates fall into the stable regime where CVGL performance is less sensitive to scale.
Fig. 6 : Visualization of scale mismatch in scale-adaptive CVGL We use one UAV-VisLoc+ query and vary the relative altitude ratio δ in H~=H(1+δ) . Top: local GT view, showing the UAV query (left) and the GT satellite crops under different δ . Middle: global search view, showing the satellite search region (left) and the corresponding similarity heatmaps over the full region. Bottom: retrieved satellite patches and final localization outcomes (green/red: success/failure), with detected vehicles used as semantic anchors shown on the right. When δ=0 (scale matched), the similarity surface exhibits a clear single dominant peak near the GT location. Under scale mismatch (e.g., δ=±0.25,±0.50 ), the similarity surface becomes multi-modal and localization easily drifts to spurious peaks. Our estimated scale (Est, corresponding to a small relative error ratio, e.g., δ≈+0.05 ) recovers reliable localization.
Dataset
Region
True Alt. Range (m)
Total Images
Used
Used Ratio
Avg. Cars
Mean Err.
UAV-VisLoc+
R01
382–452
690
532
77.1%
106.4
4.3%
R02
350–404
818
221
27.0%
34.2
5.5%
R03
454–462
768
515
67.1%
70.4
3.7%
R04
535–551
738
446
60.4%
41.1
3.7%
R05
377–416
473
112
23.7%
24.0
5.2%
R06
325–364
344
72
20.9%
14.1
4.8%
TABLE II : UAV flight altitude estimation performance. True Alt. represents the ground truth relative altitude range; Used denotes the number of images with sufficient semantic anchors ( τconf≥0.5,N≥5 ); Mean Err. indicates the Mean Absolute Percentage Error (MAPE) between estimated and true altitude.
Dataset
Region
True Alt. Range (m)
Used Ratio
Mean Err. (MAPE)
SR (%) Est / GT
UAV-VisLoc+
R01
382–452
77.1%
4.3%
38.0 / 38.2
R02
350–404
27.0%
5.5%
53.8 / 56.6
R03
454–462
67.1%
3.7%
50.4 / 52.9
R04
535–551
60.4%
3.7%
63.2 / 63.9
R05
377–416
23.7%
5.2%
71.4 / 73.2
R06
325–364
20.9%
4.8%
65.3 / 65.3
TABLE III : Scale-adaptive CVGL performance (SR, %) using the estimated altitude (Est) versus ground-truth altitude (GT), together with altitude statistics and estimation accuracy. MAPE denotes the Mean Absolute Percentage Error of the estimated altitude.
Fig. 7 : Qualitative comparison with monocular depth estimation. Rows (a)–(b): UAV-VisLoc Dataset; rows (c)–(d): DenseUAV Dataset. Each row shows (left to right) the UAV image, vehicle detections, Depth Anything V3 relative depth/3D visualization, and the estimated ground distance. GT: ground truth; MDE: Depth Anything v3; Ours: our method.
Method
DenseUAV+ MAPE (%)
UAV-VisLoc+ (Avg) MAPE (%)
Full model (Default)
2.9
4.4
Naive (box edges as vehicle)
8.1 ( ↑ 5.2)
5.9 ( ↑ 1.5)
TABLE IV : Comparison between our full model and a naive baseline. MAPE denotes the Mean Absolute Percentage Error of the estimated altitude. The default setting is indicated in gray.
Method/Settings
Parameters
DenseUAV+
UAV-VisLoc+ (Avg)
τconf
Nmin
Ratio (%)
MAPE (%)
Ratio (%)
MAPE (%)
Different τconf
RTMDet
0.2
5
49.0
15.7
78.5
14.6
RTMDet
0.3
5
39.4
5.8
61.2
7.7
RTMDet (Default)
0.5
5
33.7
2.9
50.4
4.4
RTMDet
0.7
5
29.9
2.3
38.6
3.6
TABLE V : Ablation studies on hyperparameters and detection models. Ratio indicates the usage ratio (availability), and MAPE denotes the Mean Absolute Percentage Error of the estimated altitude. The default setting is indicated in gray.
Fig. 8 : Metric-Aware Generative Urban Planning. We simulate a site planning task on an unscaled orthophoto crop. (a) The input crop from Region 01 representing a vacant lot. (b) Ours (Metric-Aware Generation): Using the GSD ( 10.2 cm/px) estimated by our method, standard sports facilities were strictly scaled to pixel dimensions, manually laid out, and then rendered using Nano Banana Pro. The result exhibits correct physical proportions consistent with the environment. (c) Direct Generation: The result rendered directly by Nano Banana Pro without metric constraints. The generated facilities show significant scale deviation (e.g., the stadium is unrealistically large).
Region
Cars Used
GT GSD
Est. GSD
Abs. Err.
Abs. Err.
( N )
(cm/px)
(cm/px)
(cm)
(%)
01
13,066
10.0
10.2
0.2
2.4%
02
1,562
10.0
10.0
0.0
0.0%
03
7,404
12.0
11.8
0.2
1.7%
04
3,679
14.0
13.7
0.3
1.8%
05
475
13.0
12.9
0.1
1.1%
TABLE VI : Absolute scale recovery results on UAV VisLoc+ orthophotos. The Ground Sampling Distance (GSD) is reported in cm/pixel. All errors are absolute values.
Cross-view geo-localization for Unmanned Aerial Vehicles (UAVs) operating in GNSS-denied environments remains challenging due to the severe geometric discrepancy between oblique UAV imagery and orthogonal satellite maps. Most existing methods address this problem through a decoupled pipeline of place retrieval and pose estimation, implicitly treating perspective distortion as appearance noise rather than an explicit geometric transformation. In this work, we propose a geometry-aware UAV geo-localization framework that explicitly models the 3D scene geometry to unify coarse place recognition and fine-grained pose estimation within a single inference pipeline. Our approach reconstructs a local 3D scene from multi-view UAV image sequences using a Visual Geometry Grounded Transformer (VGGT), and renders a virtual Bird's-Eye View (BEV) representation that orthorectifies the UAV perspective to align with satellite imagery. This BEV serves as a geometric intermediary that enables robust cross-view retrieval and provides spatial priors for accurate 3 Degrees of Freedom (3-DoF) pose regression. To efficiently handle multiple location hypotheses, we introduce a Satellite-wise Attention Block that isolates the interaction between each satellite candidate and the reconstructed UAV scene, preventing inter-candidate interference while maintaining linear computational complexity. In addition, we release a recalibrated version of the University-1652 dataset with precise coordinate annotations and spatial overlap analysis, enabling rigorous evaluation of end-to-end localization accuracy. Extensive experiments on the refined University-1652 benchmark and SUES-200 demonstrate that our method significantly outperforms state-of-the-art baselines, achieving robust meter-level localization accuracy and improved generalization in complex urban environments.
Haoyuan Li, Wen Yang, Fang Xu +4
School of Electronic Information, Wuhan University, Wuhan 430072, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China · Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing 100094, China +1
Cross-view UAV geolocalization is fundamentally a challenging large-scale image retrieval task, aiming to determine the geographic coordinates of Unmanned Aerial Vehicle (UAV) queries by matching them against an extensive geo-tagged satellite image database. Most existing methods learn separate feature representations for each view and determine the final prediction using naive heuristics to assess feature similarity, thereby neglecting to model the crucial cross-view relationships. In this paper, we propose SkyLink, a novel plug-and-play ranking framework that pioneers joint relational modeling of inter-view relationships to enhance cross-view UAV geolocalization. SkyLink leverages a Large Vision-Language Model (LVLM) to model the intricate visual-semantic relationships between UAV and satellite views, facilitating effective cross-view matching. To further refine the learning process, we introduce a relational-aware loss. It leverages soft labels to provide a more nuanced supervision signal, mitigating the harsh penalty on near-positive pairs. This approach enhances both training stability and the model's discriminative capacity. Extensive experiments conducted across multiple base retrieval architectures and benchmark datasets demonstrate that SkyLink significantly boosts the ranking effectiveness of existing models, consistently achieving superior performance in various challenging scenarios.
Bowen Liu, Pengyue Jia, Wanyu Wang +9
Department of Data Science, City University of Hong Kong, Hong Kong · Information Systems, City University of Hong Kong, Hong Kong · College of Computer Science and Technology, Zhejiang University of Technology, Zhejiang
Cross-view geo-localization (CVGL) is fundamental for precise localization and navigation in GPS-denied environments, aiming to match ground or UAV imagery with satellite views. Existing approaches often rely on global feature alignment, but they suffer from substantial domain shifts induced by varying regional textures and weather conditions. This issue becomes even more pronounced in UAV-based scenarios, where the broader perspective inevitably introduces dense, fine-grained objects, creating significant visual clutter. To address this, we draw inspiration from Object-Centric Learning (OCL) and propose InfoGeo, an information-theoretic framework designed to enhance robustness and generalization. InfoGeo reformulates the optimization as an information bottleneck process with two core objectives: (i) maximizing view-invariant information by aligning the object-centric structural relations across views, and (ii) minimizing view-specific noisy signals through cross-view knowledge constraints. Extensive evaluations across diverse benchmarks and challenging scenarios demonstrate that InfoGeo significantly outperforms state-of-the-art methods.
Hongyang Zhang, Maonan Wang, Ziyao Wang +2
School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen), Guangdong 518172, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong SAR 999077, China · Shanghai AI Laboratory, Shanghai 200032, China