SCC-Loc: A Unified Semantic Cascade Consensus Framework for UAV Thermal Geo-Localization

Xiaoran Zhang; Yu Liu; Jinyu Liang; Kangqiushi Li; Zhiwei Huang; Huaxin Xiao

cs.CVApr 3, 2026

SCC-Loc: A Unified Semantic Cascade Consensus Framework for UAV Thermal Geo-Localization

Authors: Xiaoran Zhang, Yu Liu, Jinyu Liang, Kangqiushi Li, Zhiwei Huang, Huaxin Xiao

Organizations: College of Systems Engineering, National University of Defense Technology, Changsha 410073, China

Abstract

Cross-modal Thermal Geo-localization (TG) provides a robust, all-weather solution for Unmanned Aerial Vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environments. However, profound thermal-visible modality gaps introduce severe feature ambiguity, systematically corrupting conventional coarse-to-fine registration. To dismantle this bottleneck, we propose SCC-Loc, a unified Semantic-Cascade-Consensus localization framework. By sharing a single DINOv2 backbone across global retrieval and MINIMA $_{\text{RoMa}}$ matching, it minimizes memory footprint and achieves zero-shot, highly accurate absolute position estimation. Specifically, we tackle modality ambiguity by introducing three cohesive components. First, we design the Semantic-Guided Viewport Alignment (SGVA) module to adaptively optimize satellite crop regions, effectively correcting initial spatial deviations. Second, we develop the Cascaded Spatial-Adaptive Texture-Structure Filtering (C-SATSF) mechanism to explicitly enforce geometric consistency, thereby eradicating dense cross-modal outliers. Finally, we propose the Consensus-Driven Reliability-Aware Position Selection (CD-RAPS) strategy to derive the optimal solution through a synergy of physically constrained pose optimization. To address data scarcity, we construct Thermal-UAV, a comprehensive dataset providing 11,890 diverse thermal queries referenced against a large-scale satellite ortho-photo and corresponding spatially aligned Digital Surface Model (DSM). Extensive experiments demonstrate that SCC-Loc establishes a new state-of-the-art, suppressing the mean localization error to 9.37 m and providing a 7.6-fold accuracy improvement within a strict 5-m threshold over the strongest baseline. Code and dataset are available at https://github.com/FloralHercules/SCC-Loc.

Figures & tables

Fig. 1: Conceptual comparison of the existing approach and our SCC-Loc framework. To overcome severe cross-modal bottlenecks in thermal geo-localization, SCC-Loc systematically resolves: (a) spatial quantization bias via semantic alignment (SGVA); (b) dense structural outliers via cascaded filtering (C-SATSF); and (c) deceptive visual decoys via consensus selection (CD-RAPS).

Fig. 2: Overview of the proposed SCC-Loc framework: (1) shared DINOv2 feature extraction for UAV thermal and satellite images; (2) SGVA-based viewport alignment; (3) C-SATSF-based correspondence purification after MINIMA ${}_{\text{RoMa}}$ matching; and (4) CD-RAPS-based position selection via pose optimization, reliability evaluation, and geographic consensus.

Fig. 3: Visual samples from the constructed Thermal-UAV dataset. The dataset systematically captures profound modality discrepancies and diurnal thermal variations across diverse spatial topologies: (a) Urban and (b) Rural Scenes showcase thermal UAV queries of distinct semantic categories (e.g., buildings, roads, fields) during daytime and nighttime; (c) Reference Map presents the global database at varying search scales, comprising visible-light satellite ortho-photos and their spatially aligned Digital Surface Models (DSM) to supply crucial 3D elevation priors.

Fig. 4: Qualitative visualization of the proposed SCC-Loc pipeline in (a) Urban and (b) Rural scenarios. The process illustrates the adaptive correction of spatial quantization bias via the SGVA module ( Re-Cropping ) and the progressive suppression of structural outliers using the C-SATSF mechanism ( Matching ). Finally, based on the purified correspondences, the CD-RAPS strategy refines candidate poses via physically constrained optimization, and computes the total reliability ( Final score ) by fusing the multi-dimensional evaluation ( Base score ) with geographic consensus to determine the robust optimal hypothesis for precise Localization .

Coarse Retrieval Phase
Hyperparameter Description	Symbol	Value
Geographic search area	$A_{\text{search}}$	$600\times 600\text{ m}^{2}$
Sliding window overlap	$\rho_{\text{overlap}}$	$60\%$
GSD scaling factor	$s_{\text{gsd}}$	$1.5$
Minimum error tolerance	$\epsilon_{\text{min}}$	$10^{-6}$
Pooling exponent	$\psi$	$4$

TABLE I: Detailed hyperparameter configurations for the proposed SCC-Loc framework.

Method

Top-3 Candidates

Top-5 Candidates

Top-10 Candidates

Recall@3

Acc@5/10/20

\pm

Time

Recall@5

Acc@5/10/20

\pm

Time

Recall@10

Acc@5/10/20

\pm

Time

CAMP + XoFTR

75.15

1.53 / 5.96 / 16.51

280.52\pm 374.93

1.03

85.70

1.66 / 6.94 / 20.04

193.75\pm 306.82

1.24

96.13

1.66 / 8.13 / 23.79

121.00\pm 216.86

1.79

CAMP + RoMa

75.15

3.06 / 12.04 / 31.49

91.89\pm 139.31

4.01

85.70

4.26 / 14.00 / 36.98

69.00\pm 105.93

6.30

96.13

4.38 / 15.83 / 39.70

49.17\pm 82.95

11.97

CAMP + MINIMA

{}_{\text{RoMa}}

75.15

4.72 / 15.57 / 40.43

84.26\pm 131.58

4.22

85.70

4.77 / 17.62 / 46.17

61.79\pm 111.69

6.60

96.13

6.77 / 21.11 / 49.83

40.92\pm 78.34

12.53

DINOv2 + XoFTR

91.23

1.66 / 7.57 / 21.45

185.70\pm 307.10

0.92

96.77

2.17 / 8.64 / 24.51

141.61\pm 254.98

1.30

99.45

2.30 / 9.57 / 25.87

100.86\pm 188.14

1.96

DINOv2 + RoMa

91.23

3.57 / 14.98 / 39.28

55.21\pm 81.07

4.28

96.77

4.09 / 16.81 / 43.45

48.04\pm 70.72

6.36

99.45

4.89 / 16.85 / 42.21

40.50\pm 56.01

12.02

TABLE II: Quantitative comparison of the proposed SCC-Loc framework against various two-stage baselines, which comprise different combinations of retrieval and matching models. The best results are highlighted in bold .

Method	Search Area: $600\times 600$			Search Area: $800\times 800$			Search Area: $1200\times 1200$			Memory
Method	Acc@5/10/20	ME $\pm$ SD	Time	Acc@5/10/20	ME $\pm$ SD	Time	Acc@5/10/20	ME $\pm$ SD	Time	Memory
NIVnet	0.10 / 0.80 / 2.60	$91.64\pm 35.06$	0.02	0.10 / 0.60 / 2.10	$92.37\pm 35.63$	0.03	0.20 / 0.70 / 1.90	$99.21\pm 47.49$	0.04	15
STHN	0.50 / 3.10 / 11.50	$40.70\pm 15.49$	0.03	0.70 / 3.10 / 12.10	$38.76\pm 16.57$	0.03	1.00 / 3.40 / 12.00	$49.60\pm 44.40$	0.04	5.5
DINOv2 + MINIMA ${}_{\text{RoMa}}$	6.85 / 22.85 / 54.09	$31.32\pm 51.18$	12.28	6.26 / 20.89 / 52.51	$36.92\pm 73.59$	12.50	5.11 / 19.06 / 48.13	$69.05\pm 148.08$	13.62	2335.5
SCC-Loc (Ours)	52.09 / 86.38 / 93.53	$\mathbf{9.37\pm 25.26}$	14.21	51.40 / 81.70 / 89.87	$\mathbf{16.84\pm 55.63}$	14.30	46.30 / 76.00 / 83.45	$\mathbf{46.62\pm 127.04}$	16.12	1174.7

TABLE III: Comparison with domain-specific baselines across varying search areas on our Thermal-UAV dataset under the Top-10 retrieval setting. Best results are in bold .

Fig. 5: Sensitivity analysis of the roll and pitch penalty weights in physically constrained non-linear pose optimization under the Top-5 retrieval setting. (a) Acc@5, (b) Acc@10, and (c) mean error (ME) under different combinations of $\lambda_{\mathrm{roll}}$ and $\lambda_{\mathrm{pitch}}$ .

Components			Performance Metrics
SGVA	C-SATSF	CD-RAPS	Recall@5	Acc@5/10/20	ME $\pm$ SD	Time
✓			96.98	5.91 / 23.49 / 56.38	33.74 $\pm$ 61.53	6.81
	✓		96.72	7.15 / 21.87 / 53.40	37.40 $\pm$ 68.48	6.82
		✓	96.72	17.40 / 46.47 / 74.72	26.08 $\pm$ 55.59	8.60
✓	✓		96.98	8.00 / 24.47 / 57.70	32.73 $\pm$ 60.38	7.00
✓		✓	96.98	20.89 / 51.91 / 78.00	21.71 $\pm$ 46.61	8.58

TABLE IV: Ablation study of the proposed framework under the Top-5 retrieval setting. The best results are highlighted in bold .

Stage	Description	Time (s)
Retrieval	Top-10 candidate search	0.769
SGVA	Semantic viewport alignment	0.001
Dense Matching	MINIMA ${}_{\text{RoMa}}$ matching	11.917
C-SATSF	Cascaded correspondence filtering	0.087
CD-RAPS	Pose optimization and selection	1.355
Other	Data conversion and I/O overhead	0.081

TABLE V: Runtime breakdown of SCC-Loc under the Top-10 retrieval setting.

SGVA Parameters		Performance Metrics
$\alpha$	$\beta$	Recall@5	Acc@5/10/20	ME $\pm$ SD
0	1.0	96.89	50.47 / 79.32 / 87.62	16.99 $\pm$ 48.01
	1.2	96.94	49.66 / 77.79 / 87.83	17.80 $\pm$ 50.78
	1.5	96.98	51.49 / 79.23 / 87.36	16.54 $\pm$ 46.29
0.2	1.0	96.98	51.66 / 81.32 / 89.87	15.30 $\pm$ 44.54
	1.2	97.11	52.00 / 82.60 / 89.91	14.75 $\pm$ 43.66

TABLE VI: Sensitivity analysis of SGVA parameters under the Top-5 retrieval setting. The best results are highlighted in bold .

$(w_{1},w_{2},w_{3},w_{4})$	Acc@5	Acc@10	Acc@20	ME $\pm$ SD
(0.25, 0.25, 0.25, 0.25)	50.17	80.60	89.23	15.27 $\pm$ 42.93
(0.1, 0.3, 0.3, 0.3)	50.68	81.40	89.74	14.90 $\pm$ 43.07
(0.1, 0.2, 0.35, 0.35)	51.66	81.32	89.87	15.30 $\pm$ 44.54
(0.2, 0.2, 0.3, 0.3)	51.15	81.62	89.45	14.97 $\pm$ 42.52
(0.1, 0.1, 0.4, 0.4)	53.40	82.81	90.00	15.12 $\pm$ 44.55

TABLE VII: Sensitivity analysis of CD-RAPS reliability weights under the Top-5 setting. The best results are highlighted in bold .

Noise Config		Performance Metrics
$\delta_{\text{pitch}}$	$\delta_{\text{yaw}}$	Recall@5	Acc@5/10/20	ME $\pm$ SD
0	0	96.98	51.66 / 81.32 / 89.87	15.30 $\pm$ 44.54
	10	97.36	51.02 / 82.00 / 89.70	15.42 $\pm$ 44.24
	20	97.36	52.34 / 81.91 / 89.06	15.48 $\pm$ 43.87
10	0	96.72	28.98 / 51.74 / 71.45	21.83 $\pm$ 44.71
	10	97.19	28.17 / 51.66 / 70.51	22.31 $\pm$ 44.87

TABLE VIII: Robustness evaluation of the proposed framework under varying pitch and yaw angles in the Top-5 retrieval setting.

Cross-View Geo-Localization

Cross-view geo-localization matches ground, drone, or aerial images to maps or satellite views to estimate where they were captured.

See all 94 papers

ScalLoc: Continuous Oblique Thermal UAV Geo-Localization with a 2D Satellite Map
Oct 10, 2026Xiaoran Zhang, Yu Liu, Jinyu Liang +1Cross-View Geo-Localization GNSS-Denied Localization
College of Systems Engineering, National University of Defense Technology, Changsha 410073, China
Structure-Aware Matching for Robust UAV Thermal-to-Satellite Geo-Localization
Oct 9, 2026Chun Liu, Yuxin Wang, Zhenxiang Huo +3Cross-View Geo-Localization Cross-Modal Feature Matching
State Key Laboratory of Spatial Datum, Henan University, Zhengzhou 450046, China · College of Remote Sensing and Geoinformatics Engineering, Henan University, Zhengzhou 450046, China · School of Computer and Information Engineering, Henan University, Zhengzhou 450046, China +1
Lights Out: A Nighttime UAV Localization Framework Using Thermal Imagery and Semantic 3D Maps
Apr 29, 2026Ryan Allen, Melissa GreeffUAV Navigation GNSS-Denied Localization
Robora Lab (www.roboralab.com), Queen’s University · affiliated with Ingenuity Labs Research Institute

Oct 10, 2026J/K move · Enter open · S save

ScalLoc: Continuous Oblique Thermal UAV Geo-Localization with a 2D Satellite Map

Xiaoran Zhang, Yu Liu, Jinyu Liang, Huaxin Xiao

College of Systems Engineering, National University of Defense Technology, Changsha 410073, China

Cross-View Geo-Localization GNSS-Denied Localization UAV Navigation Map-Based Localization Thermal Imaging

Open paper AskPDF arXiv

Cross-modal Thermal Geo-localization (TG) provides an illumination-robust positioning solution for Unmanned Aerial Vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environments. However, existing thermal geo-localization methods mainly focus on near-nadir observations or depend on map-side elevation priors, limiting their applicability to continuous oblique thermal localization with only geo-referenced 2D satellite maps. To address this problem, we propose ScalLoc, a scale-aligned local-to-map localization framework for continuous oblique thermal UAV geo-localization. ScalLoc follows a reconstruction--refinement--anchoring--propagation design: it reconstructs local window geometry from thermal sequences, refines camera poses using track-based pose optimization, adaptively anchors each window to the satellite map, and propagates validated states to recover frame-level global trajectories. To support systematic evaluation, we construct Thermal-Oblique, a real thermal UAV sequence dataset containing 216 urban and rural day--night trajectories and 28,112 thermal frames with frame-level annotations. Extensive experiments demonstrate that ScalLoc reduces the mean localization error from 112.69 m to 26.97 m compared with the VGGT-only configuration, corresponding to a 76.07% reduction, and achieves the lowest frame-weighted mean position error among the compared methods. These results demonstrate the feasibility of continuous oblique thermal geo-localization using only lightweight 2D satellite maps, without requiring digital surface models or preconstructed 3D maps. Code and dataset are available at https://github.com/FloralHercules/ScalLoc.

Abstract

Cross-View Geo-Localization

Explore similar work

ScalLoc: Continuous Oblique Thermal UAV Geo-Localization with a 2D Satellite Map

Structure-Aware Matching for Robust UAV Thermal-to-Satellite Geo-Localization

Lights Out: A Nighttime UAV Localization Framework Using Thermal Imagery and Semantic 3D Maps