Contact-Straddling Hard Negatives for Mapping Young Lunar Lava Flows

Over the past decade, returned samples and orbital mapping have shown that the Moon kept erupting basalt long after it was supposed to be geologically dead. Chang'e-5 basalts date to about 2.0 billion years ago, Apollo-basin units to about 1.68 billion years, and some Orientale margin basalts appear as young as 0.67 billion years. Testing the heat-source models that explain this late activity depends on a dull-sounding task: drawing the boundary of each individual lava unit, so that it can be dated by crater counting and characterized chemically. Today those boundaries are drawn by hand or by fixed thresholds on coarse oxide maps. The hard cases are exactly the scientifically interesting ones. Adjacent basalt units differ only subtly, impact ejecta smears highland material across them, and the same patch of ground looks different in every orbital pass because the Sun angle changes.

We propose to train a dense feature encoder, meaning one that outputs a feature vector for every small image patch rather than one vector per image, using a contrastive objective. Contrastive learning pulls matched examples together in feature space and pushes mismatched ones apart. Our version is built from two kinds of pairs. Positives are the same ground location seen in different orbital passes, where only the lighting differs. Hard negatives are patches taken from just across a mapped geological contact, where the terrain is similar but the rock is different. The central bet is that drawing negatives from within the same basin solves a known failure of vision-language segmenters at the same time as it teaches fine discrimination. That failure is the regional "this is all mare basalt" signal swamping local detail. A sampling rule reserves likely-unmapped units so they are not pushed away as negatives.

Core Claim

Fine-grained geological contacts on the Moon can be recovered by a patch-level contrastive encoder if two conditions hold. Its negatives must come from the immediate neighborhood across a known contact, and its positives must span illumination changes across repeated passes of the same static terrain. This choice of negatives does more than make the task harder. When the anchor and all its hard negatives share the same regional context, the context carries no information for the loss. The encoder is therefore pushed to discard the dominant regional component of each patch feature while keeping it available elsewhere. That gives a learned, data-dependent alternative to the fixed subtraction used by SegEarth-OV, which removes a scaled copy of the global image token from every patch token ($\hat O = O_{patch} - \lambda O_{[CLS]}$). On Earth scenes this works because the dominant class is clutter. In a mare basin the dominant class is the petrological baseline, and the research file flags that removing it with a single $\lambda$ risks stripping out the context that distinguishes basalt families.

The four directions play different roles. Long-lived volcanism is the application and supplies the scientific payoff. Hard negative mining is the core mechanism. The global-local alignment gap is the failure mode the mechanism is designed to fix. The image-to-video shift appears as a constraint: we treat repeated orbital passes over unchanging terrain as a slow "video" in which only lighting moves.

Relation to Prior Work

Three lines of work motivate the gap. Chen et al. delineate 17 high-alumina basalt units in Mare Imbrium using multi-ratio thresholds on Diviner oxide and Lunar Prospector thorium maps. They report that about 26% of the identified pixels are potentially polluted by highland material, which is precisely where fixed thresholds fail. Yang et al. train a one-dimensional convolutional network on per-pixel Kaguya spectra, calibrated on 115 returned samples from 55 sites. They report oxide regression $R^2 > 0.90$ and identify 26 young units (U1–U26). Their unit outlines were drawn semi-heuristically from the resulting titanium maps, and the model has no spatial context. Liu et al. date multiphase Orientale basalts by crater counting on 7 m/pixel Chang'e-2 imagery and note that unit delineation and secondary-crater filtering dominate the age error. None of these studies learns a representation.

Two papers supply the method. SLIP-RS shows that counterfactual hard negatives force a remote-sensing encoder onto fine sub-structures. It does this by swapping one attribute of an otherwise identical aircraft description. On the "plane purpose" attribute it reports 72.04% mAP, against roughly 9% for prior models. Its negatives are discrete text edits on engineered objects. Ours are spatial neighbors across a continuous, contaminated geological contact, and they are image patches rather than text. ViC-MAE treats video frames about one second apart as positives for a global contrastive loss, combined with masked-patch reconstruction. We reuse the idea of temporal views as positives but make three changes. Our "frames" are revisits under different Sun angles rather than object motion. The contrast is applied per patch, not to a mean-pooled vector. And the negatives are hard rather than drawn from other videos.

Two papers constrain the design. LCDNet found that lunar crater catalogs are incomplete, so treating confident uncatalogued detections as negatives suppresses real craters. Its reserved negative sampling holds such detections out instead, and with it the model recovers 24,504 confirmed new craters. This is a direct warning for us: young basalt units are also under-catalogued, and aggressive negative pushing could erase them. The selectively contrastive triplet loss of Stylianou et al. shows that very hard negatives destabilize standard triplet and InfoNCE losses unless the push is applied only inside a calibrated similarity band. We adopt that gating. DenseCLIP and STSUN provide the dense-prediction and multi-temporal machinery we compare against; neither mines negatives.

Research Program

The input is a co-registered stack over three regions with published unit maps: Mare Imbrium (Chen et al.), Oceanus Procellarum (Yang et al.), and Orientale (Liu et al.). Each grid cell carries Kaguya multiband reflectance, Diviner oxide and rock-abundance layers, Chang'e-2 CCD panchromatic imagery and terrain model, and, where available, multiple acquisitions of the same footprint under different illumination. Each acquisition sequence is treated as a slow video of static ground.

The encoder is a vision transformer with a masked-reconstruction auxiliary loss, following ViC-MAE, so that features keep texture and morphology. For an anchor patch $i$ with feature $z_i$, the positive $z_i^+$ is the same ground cell in a different acquisition. The negative set $\mathcal N_i$ has three parts. Hard negatives are cells within distance $d$ across a mapped contact. Easy negatives are random cells from other basins. Reserved cells are excluded entirely. The loss is

$$ \mathcal L_i = -\log \frac{e^{s(z_i, z_i^+)/\tau}}{e^{s(z_i, z_i^+)/\tau} + \sum_{j \in \mathcal N_i} w_{ij}\, e^{s(z_i, z_j)/\tau}} $$

Here $s$ is cosine similarity and $\tau$ is a temperature. The weight $w_{ij}$ is 1 for easy negatives. For hard negatives, $w_{ij}$ is nonzero only when $s(z_i,z_j)$ falls inside a danger band, following Stylianou et al., which prevents gradient blow-up on near-identical pairs. A cell is reserved, rather than used as a negative, when two conditions hold. It must lie outside every mapped unit, and its similarity to some mapped young unit must exceed a threshold that is recalibrated each epoch, as in LCDNet's reserved negative sampling. Contamination is handled through the negatives too. Cells inside Chen et al.'s flagged contaminated zone are paired against pristine cells of the adjacent unit, so the encoder must separate lithology from ejecta veneer using morphology and texture as well as spectra.

For segmentation, patch features feed a lightweight decoder, and boundaries are read from feature discontinuities. We also report open-vocabulary pixel-text scoring in the DenseCLIP style, using text descriptors of unit chemistry. This is secondary; the core claim does not depend on language.

Proposed Experiments

Evaluation uses leave-region-out splits. We train with Imbrium contacts and test on Procellarum units U1–U26 and the Orientale units. This avoids circularity: Chen et al.'s units were themselves derived from Diviner thresholds, so they cannot fairly score a model that ingests Diviner. The metrics are boundary F-score at tolerances of 1, 2, and 5 grid cells, unit-level IoU, and recall of held-out units. A further independent check uses crater ages. On Orientale, predicted boundaries should separate regions whose crater-count model ages differ in Liu et al., and we test whether predicted units are more internally age-consistent than the published polygons. Illumination stability is measured as the ratio of feature distance across passes of the same cell to feature distance across a contact. No reported result found on Lacuna for boundary F-score or unit IoU on any lunar mare unit benchmark, so all boundary numbers in this study are new and every baseline is rerun.

Baselines

The naive reference point is the published unit polygons rasterized onto the grid, together with k-means clustering of the oxide stack. This shows how much any learned method adds over the inputs a geologist already has.

The strongest existing method is per-pixel spectral inversion following Yang et al., which reports $R^2 > 0.90$ for oxide regression. That figure measures chemistry, not boundaries. We retrain their one-dimensional network on the same sample set and leave-one-out protocol. We then segment its oxide maps with the best of three post-processors, chosen on the Imbrium validation set: thresholding, a conditional-random-field smoother, and a graph-cut. This gives the spectral approach spatial context at no learning cost. Beating it means a higher boundary F-score at the 2-cell tolerance on Procellarum and Orientale, with confidence intervals from unit-level bootstrap that exclude zero. The gap should be largest inside the contaminated zone.

The strongest training-free vision baseline is SegEarth-OV, which reports an average improvement of 5.8% over MaskCLIP and ClearCLIP on Earth multi-class segmentation. We run it on the CCD and false-color stack with $\lambda$ swept on Imbrium. We also run DenseCLIP, fine-tuned with the same labels and compute as our decoder.

The ablations isolate the key ingredients. The first replaces hard negatives with random in-batch negatives, which is effectively ViC-MAE adapted to patch level. The second keeps hard negatives but removes the reservation rule. The third replaces multi-pass positives with photometric augmentation of a single mosaic. The fourth adds SegEarth-OV's fixed subtraction on top of our features. All learned variants share the same backbone, number of epochs, and hyperparameter search budget, which we fix at twenty trials each.

Why not just…?

Why not just run per-pixel spectral inversion and smooth the map? This is the most serious threat. Chemistry may simply determine units, and a good smoother might recover contacts as well as any learned encoder. Our expectation that it falls short rests on one point. Chen et al. show that a quarter of high-alumina pixels carry ejecta contamination, and a mixed spectrum cannot distinguish "intrinsic high-alumina basalt" from "low-alumina basalt with highland dust." Morphology, such as flow fronts and texture, might. If the smoothed-inversion baseline matches us within bootstrap noise inside the contaminated zone, we will conclude that spatial representation learning is not needed for unit mapping and report that result.

Why not just photometrically normalize the images instead of using multi-pass positives? Physical photometric correction removes much of the Sun-angle effect and costs nothing to learn. The third ablation tests this directly: augmentation-only positives on normalized mosaics. If illumination stability and boundary F-score are unchanged, the video-style positives are unnecessary, and the proposal reduces to a hard-negative method on single mosaics. That would still be useful, but narrower.

Why not just use a training-free vision-language segmenter with global-token subtraction? SegEarth-OV's fixed subtraction is cheap and removes the regional signal. However, its CLIP features were learned on terrestrial images with no notion of basalt chemistry, and a single $\lambda$ either keeps the "all basalt" context or strips it everywhere. Our fourth ablation and the $\lambda$ sweep show whether the fixed subtraction ever matches in-region negatives. If it does, the "learned bias removal" part of our claim is wrong.

Why not just use random negatives, as in ViC-MAE? Random negatives from other basins can be separated using regional cues alone, which is the laziness pathology. The first ablation measures how much of the gain comes specifically from contact-straddling negatives.

Decisive Checks

Three outcomes would settle the proposal. First, the held-out unit test. We remove a third of the Procellarum young units from the training labels and compare their recall under hard negatives with and without reservation. If reservation does not protect these units, the LCDNet lesson does not transfer, and hard negative mining is unsafe for discovery-oriented mapping. Second, the boundary F-score inside Chen et al.'s contaminated zone against smoothed spectral inversion. This tests whether spatial learning adds anything beyond chemistry. Third, the Orientale crater-age consistency of predicted units against the published polygons. This is the only check independent of every label source and the one most relevant to dating late volcanism.