MTA-Dataset: Multiple-Tilt-Angle Dataset for UAV–Satellite Image Matching

Qifei Liu, Liang Jiang, Guoqiang Wu, Kun Huang, Haohui Sun, Gengchen LiuApplied Sciences 2026 · DOI

TL;DR

This paper introduces the MTA-Dataset, a new benchmark for UAV-satellite image matching that specifically addresses the challenges posed by oblique imagery (high tilt angles). They also propose a novel spatial-resolution-based cropping strategy that significantly improves matching accuracy, especially in challenging, high-tilt scenarios.

What Problem Does This Paper Address?

Accurately matching images from Unmanned Aerial Vehicles (UAVs) with satellite imagery is crucial for applications like autonomous navigation and environmental monitoring. However, real-world UAV operations often involve taking pictures at significant angles (oblique photography) due to operational constraints or weather. This oblique perspective introduces severe distortions (perspective and radiometric), creating a "domain gap" between UAV and satellite images. Existing datasets lack sufficient data with extreme tilt angles and detailed ground-truth annotations, making it difficult to evaluate and improve image matching algorithms in these challenging conditions. This paper aims to fill this gap by providing a new dataset and a method to improve matching performance.

Methods

The authors introduce two main contributions:

  1. MTA-Dataset Construction: They created a dataset of 1892 UAV images with tilt angles ranging from 0° to nearly 90°, captured at altitudes up to 300m. Crucially, each UAV image is annotated with the precise geographic coordinates of five keypoints, moving beyond single-point annotations and allowing for a more granular evaluation of matching accuracy across different image regions.
  2. Spatial-Resolution-Based Cropping Strategy: To combat the distortions and resolution inconsistencies between UAV and satellite images, especially at high tilt angles, they propose a preprocessing step. This strategy analyzes the spatial resolution of the UAV image's ground footprint and crops regions where the resolution falls below a certain threshold. This effectively removes redundant or overly distorted areas, creating more consistent image pairs for matching algorithms. They then use this strategy to evaluate various matching algorithms.

Datasets

  • MTA-Dataset (New): This is the primary dataset introduced in the paper. It contains 1892 UAV images with tilt angles from 0° to 89°, paired with a high-resolution (0.3m) satellite map. Each UAV image has five keypoint annotations with precise geographic coordinates.
  • Existing Benchmarks (for comparison): The paper references and compares its dataset and methods against existing datasets like University-1652, VPAIR, SUES-200, DenseUAV, UAV-VisLoc, GTA-UAV, and AnyVisLoc. The MTA-Dataset is highlighted for its extended tilt-angle range and multi-point ground truth.

Key Insights and Ideas

  • Extreme Tilt Angles are a Major Hurdle: The paper strongly emphasizes that as UAV tilt angles increase (especially beyond 50°), matching performance degrades drastically for most algorithms. This is due to severe geometric distortions and the introduction of noisy, low-resolution regions.
  • Five-Point Annotation is Crucial: Moving beyond single-point (center) annotations is vital for accurately assessing matching performance across the entire image, especially in oblique views where distortions vary significantly.
  • Resolution Mismatch is a Key Problem: The difference in spatial resolution between UAV and satellite imagery, exacerbated by perspective distortions in oblique views, is a significant challenge for feature matching.
  • Strategic Cropping Can Be Highly Effective: The proposed spatial-resolution-based cropping strategy demonstrates that intelligently removing problematic regions can lead to substantial improvements, even for state-of-the-art matching algorithms.

Key Results

  • Dataset Size: The MTA-Dataset contains 1892 UAV images.
  • Performance Degradation: As tilt angles increase, localization errors generally increase, and matching precision drops sharply. Performance deteriorates most drastically in the [50°, 90°) range.
  • Spatial-Resolution-Based Cropping Impact: For the SuperPoint + SuperGlue baseline, this strategy reduced average matching localization errors in the [50°, 90°) range by:
  • [50°, 60°): 33.49 m
    • [60°, 70°): 37.86 m
    • [70°, 80°): 98.3 m
    • [80°, 90°): 109.95 m
  • Algorithm Performance: Deep learning-based methods like ASpanFormer, SP+SuperGlue, GlueStick, and EfficientLoFTR generally performed better than traditional methods (SIFT, ORB) and some other deep learning approaches in multi-tilt-angle scenarios.

Limitations and Future Work

  • Dataset Limitations: The current MTA-Dataset was collected from a single geographic region and season, limiting its radiometric diversity.
  • Cropping Strategy Limitations: The spatial resolution thresholds were determined via grid search, not a dynamic theoretical model, and their stability across different terrains needs further investigation.
  • Future Work:
    • Enhance dataset radiometric diversity and distribution across tilt angles.
  • Develop adaptive and real-time cropping methods.
  • Investigate the theoretical benefits of larger FOVs in UAV-satellite matching, considering the trade-offs with distortions.

Summary updated

Abstract

Accurate target localization via matching real-time UAV images with reference satellite imagery is essential for autonomous environmental perception. Nonetheless, operational constraints and weather conditions often necessitate oblique photography. This large-tilt mode causes significant perspective and radiometric distortions, resulting in a substantial domain gap between UAV and vertical satellite imagery. The scarcity of datasets featuring extreme viewpoint shifts and fine-grained ground-truth labels hinders the validation of image matching algorithms in multi-tilt-angle environments. To address this issue, we introduce the multiple-tilt-angle dataset (MTA-Dataset), containing 1892 UAV images with tilt angles spanning 0°,90° and flight altitudes up to 300 m, supported by high-precision five-point manual annotations. Based on this benchmark, we evaluate state-of-the-art matching algorithms and propose a spatial-resolution-based cropping strategy. Experimental results demonstrate that, as the UAV tilt angle increases within the range of 0°,90°, although the expanding field of view provides richer contextual information, the localization errors of all methods increase significantly and matching precision drops sharply due to severe geometric distortions in far-field regions and interference from redundant background information, with performance deteriorating most drastically in the 50°,90° range. With the integration of our strategy, the average matching localization errors of SuperPoint + SuperGlue baseline for UAV images within the tilt-angle ranges of 50°,60°, 60°,70°, 70°,80°, and 80°,90° are reduced by 33.49 m, 37.86 m, 98.3 m, and 109.95 m, respectively. Our study provides a more comprehensive evaluation framework for robust UAV–satellite image matching algorithms in multi-tilt-angle scenarios.
BibTeX
@inproceedings{liu2026mta,
  author = {Qifei Liu and Liang Jiang and Guoqiang Wu and Kun Huang and Haohui Sun and Gengchen Liu},
  title = {{MTA-Dataset: Multiple-Tilt-Angle Dataset for UAV–Satellite Image Matching}},
  booktitle = {Applied Sciences},
  year = {2026},
  doi = {10.3390/app16052488},
}
Figure