Enhancing Rotorcraft Safety: Zero-Shot Visual Language Model for Obstacle Detection around Helipads from Satellite Imagery

Rotorcraft continue to experience higher fatal accident rates compared to fixed-wing aircraft, primarily due to low altitude flight operations and reduced situational awareness in complex environments. A critical factor is the limited availability of accurate, up-to-date information on helipads and surrounding obstacles - such as trees, poles, and buildings - that pose significant risks during takeoff and landing. Existing resources, including the Federal Aviation Administration's heliport registry, are often outdated and incomplete, particularly for private or state-operated sites, and fail to report nearby obstacles. This lack of up-to-date data is largely due to privacy restrictions at certain locations and the high cost associated with comprehensive obstacle surveys. To address this challenge, we develop a deep learning (DL) framework that automatically detects helipads and nearby obstacles from high-resolution satellite imagery. Our approach combines Mask R-CNN for precise pixel-level helipad segmentation with Grounding DINO, a zero-shot vision-language model that identifies obstacles using flexible text prompts (e.g., "Pole", "Tree") without task-specific training. This text-guided, scalable detection method adapts to diverse and evolving operational settings. We validate our framework across helipads in the United States, and demonstrate strong performance in both helipad localization and obstacle detection. In addition, we build a web-based application that automates image processing, updates incorrect heliport coordinates, and provides obstacle reports. This work aims to enhance aviation safety, modernize infrastructure records, and deliver scalable tools to the aviation and machine learning communities.