Título: AURORAS: Automated Remote Observation for Recognition and Analysis of Safe Landing Sites via Deep Learning
Autores: Luísa Mirelle Costa dos Santos, Matheus Corrêa Domingos, Hercules Carlos dos Santos Pereira, Rafael Santos, Luis Vieira, Valdivino Alexandre de Santiago Junior & Elcio H. Shiguemori
Resumo: The autonomous identification of safe landing sites on small celestial bodies with highly irregular terrains is a challenge for deep space missions. Traditional methods often rely on active sensors (e.g., LiDAR) and 3D reconstructions based on Digital Elevation Models (DEMs), which impose mass, power, and complexity constraints that are impractical for many small-body missions. To address this limitation, this study proposes AURORAS (Autonomous Remote Observation for Recognition and Analysis of Safe Sites), a supervised semantic segmentation framework designed to detect safe and unsafe regions directly from high-resolution 2D RGB images (2048 ×2048 pixels, 2–5 cm/pixel) acquired by the Rosetta OSIRIS Narrow Angle Camera. A dataset of 17 manually labeled images was prepared and divided into non-overlapping 256 ×256 patches, initially totaling 1,088 samples. To improve model generalization, on-the-fly data augmentation was applied during training, including random horizontal and vertical flips, and rotations of 90°, 180°, and 270°. Three convolutional neural network (CNN) architectures were evaluated: U-NetResNet18, SegNet, and UNetResNet18 + SegNet with an Attention Mechanism. All models were trained in two stages, pretraining on augmented patches and fine-tuning on the original data, using a systematic split into training, validation, and test sets. Evaluation metrics included Pixel Accuracy, Intersection over Union (IoU), Precision, Recall, and F1-Score. Among the tested architectures, SegNet achieved the best overall performance, making it particularly suitable for minimizing false positives in critical space applications. The proposed method demonstrates the feasibility of using passive optical sensing and deep learning for autonomous landing site selection without relying on DEMs or active sensors.
Palavras-chave: Automated Remote Observation; Deep Learning; Semantic Segmentation; Comet; Landing; Sites Selection.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191923
Artigo em PDF: CBIC_2025_paper1191923.pdf
Arquivo BibTeX:
CBIC_2025_1191923.bib