UASTHN introduces CropTTA, a simple crop-based test-time augmentation that provides model-agnostic data uncertainty estimation for deep homography estimation in satellite-thermal geo-localization, achieving 7m geo-localization error with a 97% success rate.
Geo-localization is an essential component of Unmanned Aerial Vehicle (UAV) navigation systems to ensure precise absolute self-localization in outdoor environments. To address the challenges of GPS signal interruptions or low illumination, Thermal Geo-localization (TG) employs aerial thermal imagery to align with reference satellite maps to accurately determine the UAV's location. However, existing TG methods lack uncertainty measurement in their outputs, compromising system robustness in the presence of textureless or corrupted thermal images, self-similar or outdated satellite maps, geometric noises, or thermal images exceeding satellite maps. To overcome these limitations, this paper presents UASTHN, a novel approach for Uncertainty Estimation (UE) in Deep Homography Estimation (DHE) tasks for TG applications. Specifically, we introduce a novel Crop-based Test-Time Augmentation (CropTTA) strategy, which leverages the homography consensus of cropped image views to measure data uncertainty effectively. This approach complements Deep Ensembles (DE), offering comparable performance with improved efficiency and seamless integration with any DHE model. Extensive experiments across multiple DHE models demonstrate the effectiveness and efficiency of CropTTA in TG applications. Analysis of detected failure cases underscores the improved reliability of CropTTA under challenging conditions. Finally, we demonstrate the capability of combining CropTTA and DE for a comprehensive assessment of both data and model uncertainty.
UASTHN captures six categories of high data-uncertainty samples leading to geo-localization failure, where predicted displacements significantly deviate from the ground truth. Thermal images are overlaid on predicted displacements on the satellite imagery for visualization.
UASTHN combines CropTTA for data uncertainty with Deep Ensembles for model uncertainty. CropTTA augments thermal images by cropping with specific offsets. The homography network FH with an uncertainty estimation module calculates aggregated displacements and data uncertainty. High-uncertainty samples are rejected. Optionally, Deep Ensembles estimate model uncertainty, which can be combined with CropTTA for comprehensive assessment.
Visualization of crop offsets and sampling methods. Colored boxes represent cropping regions with different offsets.
Comparison of uncertainty estimation methods across DHE baselines at WS = 1536. All baselines trained with real and synthesized thermal data.
| DHE Method | UE Method | Uncertainty | DC = 128m | DC = 256m | DC = 512m | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| MACE | CE | SR | MACE | CE | SR | MACE | CE | SR | |||
| DHN | |||||||||||
| — | — | 73.60 | 73.58 | 100% | 171.93 | 171.02 | 100% | 342.41 | 341.11 | 100% | |
| DE | model | 61.91 | 61.90 | 97.3% | 162.77 | 162.75 | 96.4% | 346.26 | 346.21 | 93.6% | |
| DM | data | 66.77 | 66.76 | 99.3% | 164.19 | 164.17 | 99.7% | 335.81 | 335.77 | 91.7% | |
| CropTTA | data | 64.18 | 62.78 | 98.6% | 162.04 | 161.82 | 95.1% | 337.90 | 337.84 | 91.5% | |
| CropTTA+DE | data+model | 64.09 | 62.70 | 96.6% | 161.84 | 161.62 | 96.1% | 336.41 | 336.47 | 93.6% | |
| IHN | |||||||||||
| — | — | 7.27 | 7.24 | 100% | 16.78 | 16.42 | 100% | 16.42 | 15.90 | 100% | |
| DE | model | 6.07 | 6.06 | 97.9% | 12.31 | 12.13 | 97.6% | 13.73 | 13.37 | 94.5% | |
| DM | data | 7.02 | 6.99 | 100% | 11.81 | 11.40 | 100% | 11.48 | 11.14 | 94.1% | |
| CropTTA | data | 7.46 | 7.47 | 97.4% | 11.91 | 10.80 | 97.5% | 9.27 | 8.06 | 95.0% | |
| CropTTA+DE | data+model | 7.25 | 7.26 | 95.7% | 11.57 | 10.44 | 97.1% | 10.67 | 9.41 | 93.8% | |
| STHN (two-stage) | |||||||||||
| — | — | 7.51 | 6.66 | 100% | 14.99 | 14.34 | 100% | 12.70 | 12.12 | 100% | |
| DE | model | 9.45 | 8.72 | 98.1% | 9.98 | 9.09 | 95.3% | 8.29 | 7.58 | 97.3% | |
| DM | data | 9.75 | 8.68 | 100% | 13.64 | 12.91 | 100% | 11.35 | 10.64 | 100% | |
| CropTTA | data | 8.26 | 7.75 | 98.5% | 7.85 | 7.31 | 95.8% | 7.93 | 7.25 | 97.5% | |
| CropTTA+DE | data+model | 8.16 | 7.65 | 98.1% | 7.50 | 6.97 | 94.5% | 7.83 | 7.15 | 97.0% | |
Blue bold = best result. Underlined = second best.
| Method | Early Stop | w/o UE | CropTTA | DE | CropTTA+DE |
|---|---|---|---|---|---|
| IHN | ✗ | 35.2 | 64.6 | 114.6 | 164.2 |
| IHN | ✓ | — | 54.6 | 63.1 | 92.1 |
| STHN | ✗ | 63.9 | 87.0 | 130.2 | 186.0 |
| STHN | ✓ | — | 78.2 | 81.9 | 118.6 |
Inference time (ms) with 5 samples on NVIDIA RTX 2080Ti. Early stopping reduces overhead while maintaining accuracy.
Thermal images overlap with satellite images, showing ground truth and predicted displacements. Thermal images are overlaid on predicted displacements on the satellite imagery for visualization.
@INPROCEEDINGS{11128423,
author={Xiao, Jiuhong and Loianno, Giuseppe},
booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)},
title={UASTHN: Uncertainty-Aware Deep Homography Estimation for UAV Satellite-Thermal Geo-Localization},
year={2025},
volume={},
number={},
pages={14066-14072},
keywords={Location awareness;Uncertainty;Satellites;Measurement uncertainty;Estimation;Autonomous aerial vehicles;Thermal noise;Robustness;Noise measurement;Robotics and automation},
doi={10.1109/ICRA55743.2025.11128423}}