AI-Enhanced Vision-Based Hazard Detection Operations in Lunar Landing Scenario

Future lunar exploration, driven by scientific and commercial objectives, requires advanced platforms capable of enabling soft landing of both crewed and uncrewed missions with extremely high precision and autonomy. Accurate hazard assessment during descent is critical to mission success and often relies on highresolution digital elevation models (DEMs) data processing. State-of-the-art techniques perform pixel-wise terrain features analysis accounting for sensor uncertainty, lander shape, and tolerances associated with its trajectory evolution. Although robust, these approaches are computationally intensive and struggle to meet realtime constraints as the DEM's spatial resolution increases and the hazard detection target size decreases. To address these limitations, we present the design, implementation, and validation of a realtime vision-based hazard detection (HD) module for lunar landing missions enhanced by deep learning. This system reframes hazard map generation as a pixel-wise (dense) regression task, using Convolutional Neural Networks to infer low-frequency (i.e., slope) and high-frequency (i.e., roughness, rocks) terrain feature maps from monocular grayscale images. The application is deployed in a hybrid fashion, employing the neural block to facilitate topography information extraction. The resulting pixel-wise distribution of these parameters is provided to a deterministic processing module, which estimates the landing safety probability. Training is supported by a custom Blender-based dataset generation pipeline developed by our research group, which leverages open-source DEM datasets to generate high-resolution, feature-rich lunar environments, and enables the acquisition of photorealistic, automatically labeled renders of the environment. The output landing safety map can be subsequently employed by the onboard guidance system to perform retargeting of the original landing site, ranking new candidates based on computed metrics. By combining deep learning with wellestablished processing techniques, the proposed framework aims to develop a more explainable and transparent tool to improve the safety and reliability of future lunar landing missions.