SAMF Teaches Drones to Locate Themselves on a Satellite Map Using Thermal Imaging, No GNSS Required

A Chinese research group has posted a preprint on arXiv describing SAMF (Structure-Aware Matching Framework). The method helps a drone determine its position by matching a live thermal camera frame against a satellite map. The task matters wherever GNSS is unavailable or unreliable and a visible-light camera goes blind: at night, in smoke or fog, and under jamming and spoofing. The authors claim the best results on two datasets, Boson-nighttime and the new Thermal-UAV-Paired. The abstract gives no specific numbers.

What makes the matching problem hard

A satellite image is a color picture taken in daylight from above. A thermal frame from a drone shows the radiation of surfaces. A roof that looks light on the map may be colder than the asphalt at night. Add differences in viewing angle and brightness, and classical appearance-based methods fail to find reliable correspondences.

According to the authors, existing approaches mostly fight the global shift between modalities and align feature distributions as a whole. In doing so, they miss what stays the same in both views: scene geometry. The outlines of roads and the boundaries of buildings and fields do not depend on whether the camera captures light or heat.

How SAMF works

The framework is built around two modules that run in sequence.

• DSEM (Directional Structure Enhancement Module) strengthens stable structural representations. It extracts directional and low-frequency spatial patterns, meaning large lines and shapes rather than texture.
• MARM (Modality-Aware Affine Remapping Module) applies an affine correction to the already enhanced features. This pulls the features of the thermal and satellite branches closer together.

The logic is to first isolate what is stable, then carefully correct the residual difference between modalities. According to the authors, ablations confirm that the modules work better together than separately. Without the full text of the paper, details are unclear: the backbone is unknown, as are the accuracy and the inference cost.

Data and setup

The authors' other contribution is the Thermal-UAV-Paired benchmark. They built it from the existing Thermal-UAV dataset by assembling "drone thermal frame – satellite image" pairs. Paired data is scarce in this niche, so an open dataset may be useful to other groups as well.

• Task: UAV geolocalization from thermal images against georeferenced satellite images
• Modules: DSEM and MARM
• Evaluation datasets: Boson-nighttime, Thermal-UAV-Paired (new)
• Claimed result: state-of-the-art on both datasets
• Numerical metrics in the abstract: not provided
• Code and data: github.com/YuxinWang879/SAMF

What to check before practical use

This is a v1 preprint and has not yet been peer reviewed. The state-of-the-art claim should be verified against the tables: which baselines were compared, on which metrics, and at what flight altitude.

Engineers building GNSS-free navigation may want to pull the repository and run SAMF on their own thermal recordings. The main questions are how well the method handles a seasonal gap between the satellite image and the flight, whether it holds up when altitude changes, and whether it fits within the compute budget of an onboard computer. Evaluating these requires inference latency and robustness to outdated maps, and the abstract has neither.

Researchers will find the benchmark itself useful. A shared set of pairs makes it easier to compare methods, and the idea of relying on structure rather than appearance carries over to other modality pairs, such as radar and optical imagery.