Remote-sensing relative density (RD) for the conterminous United States: direct prediction from AlphaEarth satellite embeddings

Version 1.1.0 adds a comprehensive stress-test and red-team review and corrects the framing with honest limitations. Forest relative density (RD = SDI/SDImax) is predicted directly from Google AlphaEarth annual satellite embeddings at FIA plot locations. The deposit provides the reusable mosvr R package, the full pipeline, anonymous plot-level inputs and out-of-fold outputs, the trained model, cross-validated skill, the head-to-head against the TreeMap 2022 RD product (Chivhenge et al. 2025), the dual-role disturbance substitution, and the red-team report. What holds: within every region, direct RS-RD tracks observed plot RD far better than the imputed TreeMap RD (per-region Pearson r2 0.08-0.16 vs 0.000-0.004); the disturbance dual-role occurrence and severity legs are preserved under the RD swap. Honest limitations (new in 1.1): (1) the surface is interpolative, not transferable - a model trained on some regions does not generalize out-of-region (leave-region-out R2 near or below zero), so it is reliable only where FIA plots inform it; (2) honest national spatial-block cross-validated R2 is ~0.18, modest because RD is observed only indirectly and public FIA coordinate fuzzing sets the ceiling; (3) the temporal change signal does not validate (RS-RD change is uncorrelated with FIA remeasurement change), so this product does NOT support spatiotemporal trend analysis; (4) the disturbance agent-share leg is model-dependent and is not claimed. FIA plot coordinates are never included. The full-resolution 30 m surface is planned for a later version.