Hüseyin Burak Ünal from the Control of Networked Systems Group at the University of Klagenfurt developed the landing site detection system for RAMSES, the group’s experiment in the Austrian Space Forum’s AMADEE-24 Mars analog mission, as part of his master’s thesis. The work was supervised by Univ.-Prof. Dr. Stephan Weiss, Assoc. Prof. Dr. Jan Steinbrener and Dr. Thomas Jantos. The system was then tested on real aerial imagery recorded during the mission in Armenia. Here is a look into his research.

When a drone has to land on ground that nobody has mapped out, such as a rocky, Mars-like desert, one of the key challenges is deciding which patch of ground below is actually safe to land on.

Loose gravel that looks flat from above can actually be dangerous to land on, and even a field of boulders can contain open, flat patches wide enough to land on. For future robot missions to Mars, and for drones on Earth flying without a human pilot, that decision has to be made by the machine itself. Ideally, it should rely only on a single onboard camera, which is light, uses little power and is already carried by most drones.

That is the problem RAMSES (Rover Aerial Mars Support and Exploration System) set out to address during AMADEE-24. Its first module, the RAMSES AI Network (RAN), is a single joint-learning model: one network with two branches, trained to do two things together, semantic segmentation and optical flow estimation. Semantic segmentation classifies what the ground is made of (rock, gravel, bare soil, bushes), and optical flow estimation compares consecutive video frames to work out how each part of the image moves. Because raised objects such as rocks are closer to the camera, they appear to move slightly more than the flat ground around them as the drone moves. This can help reveal height differences that a single image can miss.

A second module, the Landing Site Assessment Module (LSAM), then divides each image into a grid of tiles, each one a possible landing spot. Every tile receives two scores. The segmentation score depends on what the tile contains: open ground raises it, rocks lower it, and the neighbouring tiles are taken into account as well, so a clear spot surrounded by rocks is rated lower than one in the middle of open ground. The optical flow score depends on how much the tile appears to move compared with the rest of the image: areas that move more, which usually points to raised objects such as rocks, receive a lower score.

The two scores are brought to the same scale and averaged into one safety score per tile. Together, these scores form a heatmap in which the colours show which parts of the terrain are safe and which are not. Only tiles that clear a safety threshold are kept, and the three best-scoring ones are returned as landing candidates.

Training a system like this needs thousands of examples of terrain with the correct labels attached, and no existing dataset provided both segmentation and motion labels for Mars-analog aerial imagery. So the model’s training data had to be built first: over 170,000 synchronized video frames, each paired with a segmentation map and a motion field, generated in NVIDIA’s Isaac Sim across eight different simulated Mars- and Earth-like environments.

The real test was aerial imagery recorded during AMADEE-24 in Armenia by the SAMPLE team. The system had never seen this imagery during training. Real terrain brought three problems that did not appear in simulation. First, hard shadows caused segmentation errors. Second, large jumps between consecutive frames weakened the optical flow estimates. Third, safe ground sometimes looked almost the same on camera as gravel or mid-sized rocks. Segmentation then mixed up safe ground with a hazard, the very problem this story opened with.

The three problems were then addressed with two targeted fixes. The first was a denser resampling of the flight path, including image rotation. It kept consecutive frames close together, which fixed the optical flow estimates. It also fixed misclassifications between rock sizes, such as gravel, mid-sized rocks and big rocks, and helped segmentation capture whole rocks instead of only parts of them. The second was stronger lighting variation in the simulated training data. This made segmentation less sensitive to hard shadows and helped it tell safe ground apart from gravel and rocks more reliably.

These fixes improved the results. Of all the tiles the system marked as safe, 95.4% were actually safe, up from 88.3% before the fixes. As expected, its overall accuracy is lower, at 69.6%, up from 63.5%. This is a result of the neighbourhood check in LSAM. A safe-looking tile next to rocks receives a lower score, so some tiles that are actually safe get rejected. This was intended, because a good landing site must also be clear of rocks and stones around it, not just within the tile itself. Landing right next to a rock is exactly what the design is meant to avoid. The real imagery has no hand-made labels, so these numbers were measured against automatically generated reference labels.

The two branches also compensate for each other. This was measured on the original imagery, where errors were most frequent. In 45.2% of the tiles where at least one branch was wrong, the combined decision was still correct, because the other branch had the right answer. This is exactly the fault tolerance the dual-branch design was meant to provide.

RAMSES is a perception system for drones in which an AI model reads the terrain and the motion between frames in the camera images, and safe landing sites are then selected from these results. It was tested offline on aerial imagery of the AMADEE-24 mission site. Even outside simulation, the results held: combining what the terrain looks like with how it shifts between consecutive frames as the drone moves gave the system a measurable fault tolerance that neither branch has on its own. That is the broader point, beyond this one drone: in safety-critical settings where only a single camera is available, extracting two complementary types of information from the same images can make landing decisions more reliable.

SCIENTIFIC OUTPUT · DELIVERED MSC THESIS

“Precise Landing Site Detection for UAVs using Deep Learning,” by Hüseyin Burak Ünal, University of Klagenfurt, 2026.
netlibrary.aau.at · urn:nbn:at:at-ubk:1-63611
oewf.org · OeWF academic theses

Source: Hüseyin Burak Ünal¹, “Precise Landing Site Detection for UAVs using Deep Learning”, MSc thesis, University of Klagenfurt, 2026

¹ Control of Networked Systems Group, Institute of Smart Systems Technologies, University of Klagenfurt, Austria

Author: Hüseyin Burak Ünal

Dieser Artikel ist auch verfügbar auf: