Abstract
<title>Abstract</title> <p>A geospatial point cloud has no stable identity: the same survey re-encoded from LAS to LAZ to COPC yields different bytes and different file hashes, leaving a verifier unable to distinguish a content-preserving re-encoding from a modified copy. We present PointCloud-ID, a deterministic content identity that survives re-encodings preserving a declared grid, and that detects and spatially localizes changes to in-scope geometry or attributes. The construction has three stages. First, canonicalization under a declared, versioned profile: coordinates quantize to the profile's grid following LAS scale/offset semantics, the coordinate reference system (CRS) is fixed, the multiset is ordered by Morton code, and the attribute scope is enumerated. Second, an attribute-partitioned Merkle tree over spatially chunked canonical data yields per-attribute roots, folded with a canonical digest of the profile into one identity. Because that digest enters the identity itself rather than an accompanying assertion, identical records declared under a different CRS or profile version cannot share an identity; changes localize to the affected chunks and attribute domains with logarithmic membership proofs. Third, a signed manifest (RFC 8785 canonical JSON, COSE/Ed25519) records tool provenance and maps onto a namespaced C2PA assertion. Lossy transforms intentionally break the identity and are recorded as explicit derived-from links. The scheme attests that signed canonical content is unchanged, not that the signed scene reflects reality. The accompanying artifact provides the profile specification, reference vectors, implementation, fixed experimental inputs, and deterministic evaluation protocol. Identities and test-vector outputs reproduce exactly, while timing and memory measurements remain environment-dependent.</p>