Authors:
Abhay Valiyaparambil, D. Hemavathi, Sayyed Khawar Abbas

Addresses:
Department of Data Science and Business Systems, SRM Institute of Science and Technology, Kattankulathur, Chennai, Tamil Nadu, India. Department of Information Systems, Corvinus University of Budapest, Budapest, Hungary.

Abstract:

Safe landing zone detection from a single downward-facing camera is a practical necessity for autonomous UAVs operating without GPS or prior maps. Existing approaches range from lightweight, training-free heuristics to large, fine-tuned deep networks, yet no study compares them under evaluation criteria relevant to the actual landing task. This paper presents a five-way comparison: A Laplacian flatness detector, an HSV color-consistency heuristic, our proposed Multi-Cue Saliency Safety Score (MC-SSS), SAM zero-shot region scoring, and SegFormer-B0 fine-tuned across three data regimes. Researchers also introduce Landing Zone Accuracy (LZA) and Largest Safe Region Ratio (LSRR), two metrics that assess whether a prediction results in a successful landing rather than simply overlapping with a ground-truth mask. Experiments on 3,939 labeled images from Aeroscapes and UAVid show that color-only methods lose up to 37.2 percentage points of accuracy when terrain changes from rural to urban. At the same time, MC-SSS holds the drop to 18.1 points. SAM, applied without any fine-tuning, achieves LZA below 38% on both benchmarks: it segments scenes correctly but selects the wrong region. SegFormer trained on data from both datasets reaches 98.0% and 78.4% LZA on Aeroscapes and UAVid, respectively. The results give practitioners a concrete, evidence-based basis for choosing a method under different operational constraints.

Keywords: Unmanned Aerial Vehicle (UAV); Semantic Segmentation; Landing Zone Accuracy (LZA); Segment Anything Model; Domain Generalization; Safe Landing Zone; Laplacian Edge Variance.

Received on: 27/05/2025, Revised on: 22/07/2025, Accepted on: 21/10/2025, Published on: 10/08/2026

DOI: 10.69888/FTSCS.2026.000719

FMDB Transactions on Sustainable Computing Systems, 2026 Vol. 4 No. 3, Pages: 165-173

• Views : 14
• Downloads : 6

Download PDF