Continuous Thought Segmentation with Expert-Oriented Uncertainty Priors for Lunar Terrain Understanding

Semantic understanding of lunar terrain presents a fundamentally ill-posed representation learning challenge: absolute ground truth is prohibitively scarce, and morphological features are deeply entangled with varying spatial resolutions and transient observational conditions. Conventional segmentation networks, bounded by single-pass feed-forward mappings and standard confidence calibration, fail to resolve these observation-dependent ambiguities and systematically misrepresent the epistemic uncertainty inherent in expert interpretation. To address this, we introduce a novel dual-branch architecture that couples a Continuous Thought Machine (CTM) with a decoupled, expert-oriented uncertainty modeling module. To dynamically adapt to resolution and observational shifts, the CTM employs an internal temporal reasoning mechanism. By iteratively refining visual representations across multiple latent "thought steps," it constructs robust, scale-aware feature manifolds rather than rigid, static mappings. Concurrently, to navigate severe label ambiguity, our decoupled branch learns fine-grained spatial uncertainty maps without degrading the primary segmentation objective. Bootstrapped from model-derived epistemic cues, this branch establishes a robust machine-derived uncertainty prior. This decoupled design lays the architectural groundwork to accommodate future expert annotations, explicitly paving the way to bridge conventional prediction entropy with human perceptual priors. Ultimately, our framework establishes an interpretable, multi-step reasoning paradigm that jointly achieves high-fidelity segmentation and expert-aware uncertainty representation in heavily observation-dependent, label-ambiguous regimes.