Abstract
Identifying ancient life on Earth and other planets requires detecting subtle biosignatures within complex geological records. This paper introduces SPACESeg, a hybrid machine learning model designed to automatically detect Double-Rippled Bedforms (DRBs)—millimeter-scale structures from the Ediacaran Period that indicate the presence of stabilizing microbial mats. By integrating an unsupervised rule-based module (CORA) with a supervised saliency-guided deep learning encoder (SGE), SPACESeg effectively distinguishes biological markers from non-biological artifacts like shadows and cracks. Experimental results on datasets from the Northern Flinders region of Australia demonstrate that SPACESeg achieves 89% accuracy, significantly outperforming standard segmentation models and offering a robust tool for future planetary exploration missions.
Introduction
The Ediacaran Period (approximately 574–539 million years ago) represents a pivotal era in Earth's history, marking the emergence of complex multicellular life. A primary geological indicator of this life is the Double-Rippled Bedform (DRB). Unlike typical seafloor ripples that are transient and easily eroded, Ediacaran ripples were often stabilized by organic microbial mats. When sediment-laden storm events occurred, these mats preserved the ripple morphology, creating distinct "stacked" cross-sections that serve as definitive biosignatures.
Detecting these features is exceptionally challenging due to their minute scale (often ~1mm thin) and the presence of significant geological "noise," such as shadows and varying rock textures that mimic biological structures. Furthermore, the scarcity of high-quality labeled geological data hinders the training of conventional deep learning models.
Methodology: The SPACESeg Architecture
To address these challenges, the researchers developed SPACESeg (Scene-aware Perception Automation using Composite Embedding for Segmentation), a hybrid framework comprising two synergistic components:
1. CORA: Composite-image Optimized Region Activation
CORA is an unsupervised, rule-based module designed to provide geological context. It standardizes lighting using a "dark channel prior" and employs Principal Component Analysis (PCA) on image attributes like texture and orientation to generate a "Composite Heatmap." This heatmap is then processed via Fuzzy Energy Active Contours to isolate potential ripple boundaries while filtering out low-entropy artifacts like shadows.
2. SGE: Saliency Guided Encoder
SGE is a supervised deep learning network that builds upon the U-Net architecture. Unlike standard models, SGE utilizes a dual-encoder system: one for the raw input image and another for the saliency map produced by CORA. These streams are multiplexed at every network level, ensuring the deep learning model remains "guided" by the rule-based geological constraints, preventing the "over-segmentation" common in purely supervised approaches.
Experimental Results
The model was evaluated using a custom dataset from the Northern Flinders region in Australia, consisting of 2,304 sub-images derived from high-resolution geological cross-sections. The dataset featured a wide range of scales (2cm to 1m field of view) and resolutions (3 to 265 pixels/cm) to simulate the variable conditions of remote rover imagery.
SPACESeg demonstrated superior performance across all key metrics:
- Accuracy: 89% (compared to 72% for U-Net and 80% for PSPNet).
- Structural Similarity (SSIM): 85%, representing a 14% improvement over using either the rule-based or deep learning components in isolation.
- Robustness: Qualitatively, SPACESeg successfully ignored large rock artifacts and complex shadow patterns that typically cause false positives in other models, accurately identifying ripples as thin as 1mm.
Conclusion and Future Work
SPACESeg represents a significant advancement in automated biosignature detection. By combining the interpretability of rule-based systems with the representational power of deep learning, the model overcomes the "shadow trap" and data scarcity issues inherent in geological imaging. Future work will focus on incorporating incremental learning capabilities and expanding the model's application to diverse planetary surfaces, potentially supporting the Mars 2020 Perseverance rover in identifying high-priority sites for sample collection.
Summary updated