PEFT
Safetensors
lora

You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

DIVE V3: Lunar DINOv3 LoRA with Matryoshka Embeddings & Epistemic Uncertainty

DIVE V3 (Deep In-Strip Variable Embeddings) is a foundation model for lunar collapse pit and cave skylight detection from LROC Narrow Angle Camera (NAC) orbital imagery. The model adapts a DINOv3 ViT-S/16 backbone using LoRA and a 4-scale Matryoshka projection head (64d, 128d, 256d, 384d) with deterministic multi-scale epistemic uncertainty.

Technical Specifications

  • Backbone: DINOv3 ViT-S/16 (facebook/dinov3-vits16-pretrain-lvd1689m)
  • LoRA Adapter: rank 32, alpha 32 on qkv, proj, fc1, fc2 (3.78M trainable parameters)
  • Input Geometry: 256 ร— 256 px at 1.0 m/px, SPICE pushbroom projection with SLDEM2015 topographic relief, rotated to top-down solar illumination
  • Matryoshka Projections: 4 concentric normalized heads (64d, 128d, 256d, 384d) trained via MatryoshkaDIVELoss (margin 0.70)
  • Dataset Partition: 100 km Haversine geodesic clustering (1,311 train, 336 val in King crater, 283 held-out test across Aristarchus, Marius Hills, Ohm, Highland)

Matryoshka-BALD Analysis

Top Epistemic Anomalies

Calibration Thresholds

Calibrated decision thresholds tau_d computed from training split positive prototypes:

Dimension Decision Threshold (tau) Pit Mean (s_pit) Neg Mean (s_neg) Separation (ฮ”)
64d 0.3287 +0.9702 -0.3129 +1.2831
128d 0.3558 +0.9816 -0.2700 +1.2517
256d 0.2908 +0.9731 -0.3914 +1.3645
384d 0.2475 +0.9742 -0.4793 +1.4535

Benchmark Results

Validation Split (King Crater)

Dim Model Rank-1 (%) MRR AUC-ROC PR-AUC FPR@95 (%) F1-Score Acc (%) ECE Sep (ฮ”) s_pit s_neg
64d Legacy HF 42.1% 0.5265 0.8917 0.5708 56.71% 0.5253 86.0% 0.1662 +0.2182 +0.5738 +0.3557
DIVE V3 97.4% 0.9868 0.9994 0.9956 0.34% 0.9620 99.1% 0.0135 +1.2464 +0.9575 -0.2889
128d Legacy HF 28.9% 0.4246 0.8469 0.4765 64.43% 0.4161 74.1% 0.2683 +0.1595 +0.6539 +0.4943
DIVE V3 100.0% 1.0000 0.9999 0.9993 0.00% 0.9620 99.1% 0.0132 +1.2318 +0.9580 -0.2738
256d Legacy HF 47.4% 0.6045 0.8773 0.6233 46.98% 0.4167 75.0% 0.2707 +0.1745 +0.6584 +0.4839
DIVE V3 100.0% 1.0000 0.9996 0.9973 0.34% 0.9500 98.8% 0.0166 +1.2901 +0.9421 -0.3480
384d Legacy HF 73.7% 0.8298 0.9855 0.9047 8.05% 0.6387 87.2% 0.1496 +0.6598 +0.7962 +0.1364
DIVE V3 100.0% 1.0000 0.9995 0.9960 0.67% 0.9500 98.8% 0.0128 +1.3854 +0.9440 -0.4414

Held-Out Test Split (Aristarchus, Marius Hills, Ohm, Highland)

Dim Model Rank-1 (%) MRR AUC-ROC PR-AUC FPR@95 (%) F1-Score Acc (%) ECE Sep (ฮ”) s_pit s_neg
64d Legacy HF 39.4% 0.5610 0.6566 0.3617 94.40% 0.2623 68.2% 0.2253 +0.1124 +0.4695 +0.3571
DIVE V3 97.0% 0.9798 0.9937 0.9807 0.40% 0.9697 99.3% 0.0067 +1.1842 +0.8987 -0.2856
128d Legacy HF 39.4% 0.5932 0.6616 0.3105 78.80% 0.2602 67.8% 0.2392 +0.1076 +0.5478 +0.4403
DIVE V3 97.0% 0.9848 0.9956 0.9842 0.00% 0.9697 99.3% 0.0058 +1.1710 +0.9242 -0.2468
256d Legacy HF 57.6% 0.7306 0.6937 0.4199 75.20% 0.3232 76.3% 0.1947 +0.1365 +0.5459 +0.4095
DIVE V3 97.0% 0.9773 0.9880 0.9773 0.00% 0.9697 99.3% 0.0075 +1.2563 +0.8978 -0.3585
384d Legacy HF 93.9% 0.9621 0.9910 0.9098 3.60% 0.6000 84.5% 0.1651 +0.7228 +0.8220 +0.0991
DIVE V3 97.0% 0.9848 0.9945 0.9825 0.00% 0.9697 99.3% 0.0068 +1.3491 +0.8989 -0.4501

Per-Region Held-Out Test Breakdown (DIVE V3)

Target Region Head Rank-1 (%) MRR AUC-ROC PR-AUC Acc (%) Sep (ฮ”) s_pit s_neg
Aristarchus (14 pos, 106 neg) 64d 92.9% 0.9524 0.9757 0.9439 98.3% +1.0962 +0.8533 -0.2429
128d 92.9% 0.9643 0.9865 0.9580 98.3% +1.0901 +0.8710 -0.2192
256d 92.9% 0.9464 0.9704 0.9458 98.3% +1.1964 +0.8501 -0.3463
384d 92.9% 0.9643 0.9838 0.9549 98.3% +1.2764 +0.8505 -0.4259
Marius Hills (7 pos, 49 neg) 64d 100.0% 1.0000 1.0000 1.0000 100.0% +1.2450 +0.9614 -0.2836
128d 100.0% 1.0000 1.0000 1.0000 100.0% +1.2039 +0.9726 -0.2313
256d 100.0% 1.0000 1.0000 1.0000 100.0% +1.2872 +0.9496 -0.3376
384d 100.0% 1.0000 1.0000 1.0000 100.0% +1.4085 +0.9543 -0.4542
Highland (10 pos, 80 neg) 384d 100.0% 1.0000 1.0000 1.0000 100.0% +1.3887 +0.9103 -0.4784
Ohm (2 pos, 15 neg) 384d 100.0% 1.0000 1.0000 1.0000 100.0% +1.4274 +0.9833 -0.4441

Matryoshka-BALD Epistemic Uncertainty Cohorts

Cohort N Mean s_64 Mean s_384 Mean p_mean H_alea (bits) I_MBALD (bits) Mean U_veto Max U_veto
Pits (Valid) 71 +0.9220 +0.9230 0.9824 0.0278 0.0007 0.0022 0.0500
Background (Clean) 538 -0.2862 -0.4589 0.0040 0.0273 0.0013 0.0030 0.1330
Noon (Washed-Out) 45 +0.0706 -0.0426 0.2708 0.1488 0.0418 0.0157 0.2356
Boundary (Ambiguous) 676 +0.8422 +0.8010 0.9179 0.1285 0.0191 0.0418 0.8735

Usage Example

import json
import numpy as np
import torch

# Load calibration metadata and positive prototypes
with open("calibration_metadata.json", "r") as f:
    calib = json.load(f)

temperature = calib["temperature"]
thresholds = {d: calib["heads"][d]["tau"] for d in ["64d", "128d", "256d", "384d"]}
prototypes = torch.load("prototypes.pt", map_location="cpu", weights_only=True)

def evaluate_tile_uncertainty(projections: dict) -> dict:
    """Evaluate multi-scale probabilities, BALD epistemic uncertainty, and coarse-to-fine veto."""
    probs = {}
    for d in ["64d", "128d", "256d", "384d"]:
        z = projections[d].squeeze().numpy()
        z = z / max(float(np.linalg.norm(z)), 1e-12)
        mu = prototypes[d].numpy()
        s = float(np.dot(z, mu))
        tau = thresholds[d]
        p = 1.0 / (1.0 + np.exp(-np.clip((s - tau) / temperature, -50.0, 50.0)))
        probs[d] = float(p)

    p_bar = float(np.mean(list(probs.values())))
    p_clamped = np.clip(p_bar, 1e-12, 1.0 - 1e-12)
    h_total = float(-p_clamped * np.log2(p_clamped) - (1.0 - p_clamped) * np.log2(1.0 - p_clamped))

    h_heads = []
    for p in probs.values():
        pc = np.clip(p, 1e-12, 1.0 - 1e-12)
        h_heads.append(-pc * np.log2(pc) - (1.0 - pc) * np.log2(1.0 - pc))
    h_alea = float(np.mean(h_heads))

    i_mbald = float(max(0.0, h_total - h_alea))
    u_veto = float(abs(probs["64d"] - probs["384d"]))

    return {
        "is_pit": bool(probs["384d"] >= 0.5),
        "p_mean": p_bar,
        "probs": probs,
        "m_bald_bits": i_mbald,
        "u_veto": u_veto,
    }

Framework Versions

  • PEFT 0.19.1
Downloads last month
2
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for F1nnSBK/lunar-dinov3-lora