CETUS: How Far Do Representations Trained on Earth Transfer to Cassini SAR of Titan?

Kevin Lee

cs.CVOct 6, 2026

CETUS: How Far Do Representations Trained on Earth Transfer to Cassini SAR of Titan?

Authors: Kevin Lee

Organizations: University of California at Los Angeles, Los Angeles, CA · NASA Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA

Abstract

Cassini synthetic aperture radar (SAR) images reveal the dunes, plains, and lake basins of Titan, providing an instance of representations learned from Earth imagery for planetary terrain classification. Cross-domain Evaluation of Earth-to-Titan Transfer Using SAR (CETUS) compares features from DINOv2, DOFA and CROMA with classical image measurements and features from an untrained vision transformer on the U.S. Geological Survey's Cassini SAR mosaic. The classifiers learn terrain labels from an expert geomorphological map and predict those labels in geographically separate Titan regions. Under logistic regression settings, pretrained encoders achieve higher mean macro F1 than the combined classical features. Encoder rankings change when feature scaling, optimization, and regularization change together. Further training on Titan improves DINOv2 performance, degrades DOFA performance, and leads to mixed results for CROMA under the tested settings. Architectural and input processing differences prevent these comparisons from isolating the effect of pretraining. Classifier fitting and performance on individual terrain classes matter when assessing representation transfer for planetary mapping. Since the map draws partly on the same radar observations, the scores measure agreement with expert interpretation.

Figures & tables

Figure 1: Titan test images and their map labels. For each class, the example is the test tile whose modal map fraction is closest to that class’s median across the test set; tile ID breaks ties. The mixed example follows the same rule among test tiles whose modal class occupies less than half the valid map samples. Each image covers 45 km by 45 km. Map and SAR percentages use separate validity masks. Brightness spans the 2nd to 98th percentile of each image’s finite positive display DN. Because each image has its own stretch, brightness differences between subplots have no common radiometric scale. The examples illustrate the images and label mixtures produced by this sampling rule; a geological assessment of their representativeness would require expert interpretation.

Figure 2: Geographic coverage and numbers of test tiles. Left: original and additional tile centers on a longitude and latitude grid. Vertical lines delimit the five sectors; markers show image centers. Right: tile counts by map class and fold, with a logarithmic color scale. The Craters class has one tile in fold 2 and 61 in fold 4, although the two folds receive equal weight in the primary scores. Tile counts measure image sampling and may include several images of the same landform.

Fitting choice	Common logistic regression	SGD classifier	Partial fine-tuning
Model parameters fitted	Linear classifier	Linear classifier	Final two transformer layers and a new linear classifier
Standardize classifier inputs	Training mean and population SD	None	None
Optimizer and regularization	LBFGS with L2, $C=1$	SGD, zero weight decay	AdamW, weight decay 0.01
Training duration	At most 2,000 iterations	100 epochs	All 50 epochs
Model evaluated	One fit; validation data unused	Classifier from the final epoch	Earliest epoch attaining the highest validation macro recall among epochs with optimizer updates
Fits per fold	One per feature family	Five classifier seeds per encoder	Five optimization seeds per pretrained encoder

Table 1: Settings for classifier fitting and partial encoder training. Standardization centers and scales features using the training mean and population SD.

Feature family	Precision	Recall	F1
DINOv2	$35.49\pm 2.94$	$43.05\pm 3.85$	$36.81\pm 2.76$
DOFA	$35.24\pm 2.97$	$43.12\pm 5.40$	$35.91\pm 3.52$
CROMA	$\mathbf{38.46}\pm 2.62$	$\mathbf{47.23}\pm 5.77$	$\mathbf{39.34}\pm 2.48$
Untrained ViT	$30.80\pm 4.98$	$34.87\pm 3.93$	$27.28\pm 3.96$
Intensity statistics	$30.46\pm 2.59$	$39.29\pm 6.83$	$23.62\pm 1.76$
Intensity histogram	$25.60\pm 2.12$	$32.89\pm 6.99$	$18.67\pm 1.58$

Table 2: Logistic regression applied to eight feature families, with encoder weights unchanged. Training features determine standardization; inverse class frequency weights give every class equal total weight. Precision, recall and F1 are macro averages in percent, followed by their sample SD across five folds. All 40 fits finish in fewer than 2,000 iterations without warnings. Bold indicates the highest mean in each metric column across all eight feature families.

Figure 3: Logistic regression with each feature family. Left: five geographic fold scores and their mean. Right: mean F1 for each class, with the number of distinct test tiles above each column. All families use the same training and test images and classifier settings. They share the initial DN scaling, but subsequent input transformations differ. Numbers beside family names give feature dimensions. A common color scale spans 0 to 80% F1. Training sets overlap between folds. The Craters class contributes between one and 61 tiles per fold, which limits comparisons among regional class scores.

Figure 4: Change in macro F1 when logistic regression replaces SGD, with identical encoder features and image sets. Colored markers show the difference for each geographic fold in percentage points; black marks give the mean. The SGD score first averages five training seeds. Logistic regression changes feature scaling, optimization, regularization and the training schedule together, so the comparison assesses the fitting procedures as a whole.

SGD classifier
Encoder	Precision	Recall	F1
DINOv2	$\mathbf{36.02}\pm 3.51$	$44.70\pm 4.30$	$\mathbf{37.71}\pm 3.56$
DOFA	$32.57\pm 3.76$	$45.25\pm 7.67$	$29.04\pm 2.94$
CROMA	$35.85\pm 3.73$	$\mathbf{48.45}\pm 7.40$	$35.26\pm 4.05$
Untrained ViT	$6.01\pm 2.17$	$16.66\pm 0.004$	$7.97\pm 2.47$
Nearest neighbors

Table 3: SGD classifiers and nearest neighbors on the same geographic test sets and encoder features. Values give mean macro precision, recall and F1 in percent, with sample SD across five folds. SGD metrics first average the training seeds in each fold. Bold indicates the highest mean in each metric column, separately for SGD and nearest neighbors.

Figure 5: Training the last two transformer layers, compared with training only a classifier on the same geographic divisions. FT denotes partial fine-tuning. Left: differences in test macro F1. Right: differences in test macro recall. Differences subtract the indicated classifier scores from the FT scores and are expressed in percentage points. Colored symbols show five fold values; black marks give their mean. FT and SGD first average five training seeds per fold, with corresponding seeds compared for the difference. Logistic regression has one fit per fold. Validation macro recall determines the FT epoch. Optimizer and training settings also differ, so these comparisons assess the full training configurations.

Encoder	FT F1, %	FT recall, %	FT minus LR F1, pp	FT minus SGD F1, pp
DINOv2	$\mathbf{41.16}\pm 2.60$	$\mathbf{52.28}\pm 7.05$	$4.35\pm 0.34$	$3.45\pm 1.10$
DOFA	$15.68\pm 3.35$	$23.76\pm 3.61$	$-20.23\pm 4.45$	$-13.36\pm 4.17$
CROMA	$39.17\pm 2.31$	$48.93\pm 4.52$	$-0.17\pm 1.12$	$3.91\pm 3.03$

Table 4: Partial fine-tuning (FT) compared with classifier training alone. LR denotes logistic regression. Scores are percentages; differences are FT less the comparison score, in percentage points. Cells give the mean and sample SD of five folds after averaging five FT seeds. SGD differences compare corresponding seeds; LR has one fit per fold. Validation macro recall determines the FT epoch. Bold indicates the highest mean fine-tuning score in each metric column. Differences from earlier fits remain in ordinary type.

Model	Primary $FT-LR$	All tiles LR	All tiles FT	All tiles $FT-LR$
DINOv2	$4.35$	$38.43$	$42.99$	$4.55$
DOFA	$-20.23$	$37.68$	$16.21$	$-21.47$
CROMA	$-0.17$	$40.98$	$41.82$	$0.85$

Table 5: Effect of regional weighting on the comparison of FT with LR. FT denotes partial fine-tuning and LR logistic regression. F1 scores are percentages; differences are percentage points. The primary calculation averages the five regional scores equally. The second calculation combines confusion counts across test folds for each seed before calculating F1, then averages seeds.

Figure 6: Change in class F1 after training the last two transformer layers, relative to logistic regression on their original features. Each colored symbol gives one fold’s difference after averaging five training seeds. Black marks give the mean across folds. Positive values indicate closer agreement with the modal map class. The common horizontal scale places DOFA’s broad decline beside the smaller changes for DINOv2 and CROMA. Several images can represent the same physical landform.

Encoder	Class	Tiles	FT P, %	FT R, %	FT F1, %	LR P, %	LR R, %	LR F1, %
DINOv2	Plains	15406	$88.05$ $\pm 2.69$	$68.03$ $\pm 7.48$	$\mathbf{76.24}$ $\pm 4.32$	$85.22$ $\pm 1.36$	$68.92$ $\pm 5.70$	$76.10$ $\pm 3.65$
DINOv2	Dunes	3958	$61.80$ $\pm 7.60$	$71.52$ $\pm 9.70$	$\mathbf{65.76}$ $\pm 7.47$	$53.23$ $\pm 11.12$	$68.04$ $\pm 8.21$	$59.43$ $\pm 9.37$
DINOv2	Hummocky	2757	$46.21$ $\pm 3.87$	$57.96$ $\pm 18.87$	$\mathbf{49.34}$ $\pm 7.55$	$37.36$ $\pm 5.34$	$57.65$ $\pm 17.63$	$44.43$ $\pm 8.22$
DINOv2	Labyrinths	498	$30.10$ $\pm 10.64$	$63.81$ $\pm 12.99$	$\mathbf{38.23}$ $\pm 7.48$	$25.24$ $\pm 9.92$	$38.88$ $\pm 18.40$	$27.85$ $\pm 5.32$
DINOv2	Lakes	495	$12.78$ $\pm 9.88$	$32.84$ $\pm 23.53$	$\mathbf{14.10}$ $\pm 11.58$	$10.51$ $\pm 8.82$	$20.87$ $\pm 20.69$	$11.09$ $\pm 12.33$
DINOv2	Craters	132	$4.64$ $\pm 6.43$	$19.54$ $\pm 23.64$	$\mathbf{3.27}$ $\pm 3.13$	$1.37$ $\pm 1.95$	$3.93$ $\pm 5.70$	$1.94$ $\pm 2.67$

Table 6: Class precision (P), recall (R) and F1 for partial fine-tuning (FT) and logistic regression (LR), in percent. Class ratios are calculated for each seed and fold; FT seeds are averaged in each fold, then the five folds receive equal weight. LR has one fit per fold. SD measures variation among folds; the tile count includes each test image once. Undefined ratios are zero. Several tiles may depict portions of the same landform. Bold indicates the higher mean F1 between FT and LR for each encoder and class.

	Predicted map class
Map class	Plains	Dunes	Hummocky	Labyrinths	Lakes	Craters	Test tiles
Plains	10,689	1,687	1,771	656	364	239	15,406
Dunes	875	2,660	141	110	117	55	3,958
Hummocky	425	87	1,965	76	42	162	2,757
Labyrinths	143	57	70	203	14	11	498
Lakes	98	157	76	29	125	10	495

Table 7: CROMA predictions from logistic regression across all five test folds. Each of 23,246 test tiles appears once. Rows give modal map classes and columns give predictions. These counts weight the test population by tile count, whereas the primary class scores weight the five regions equally. CROMA is the encoder with the highest mean F1 in the common classifier comparison.

Fold	Craters tiles	Components	Largest component	Tiles in largest
1	8	2	18	7
2	1	1	17	1
3	14	5	14	8
4	61	4	58	52
5	48	5	53	30

Table 8: Craters test tiles and the map polygon components they represent. The largest component column identifies the map component with the most test tiles; the last column gives that count. A separate geological inventory is needed to relate these components to distinct impact structures.

Figure 7: Effect of replacing training tiles while keeping class counts and test sets constant. Left: logistic regression. Right: SGD. Colored points give the change in macro F1 for each fold, in percentage points; black marks and printed values give the mean. LR averages five replacement sets, each compared with the original fit. Each set reinstates all eligible excluded tiles; the sets vary which retained tiles are removed. SGD compares corresponding training seeds, averages five training seeds per set, then averages five sets. Initial DN scaling stays the same; LR recalculates feature standardization. All fits contribute, and every LR fit finishes without a warning or reaching its iteration limit. Locations and image content both change, and training sets overlap between folds.

Fold	Training tiles	Replaced tiles	Training %	Reinstated Craters	Validation Craters	Test Craters
1	14,603	486	3.33	0	1	8
2	12,542	486	3.87	2	14	1
3	12,348	429	3.47	0	59	14
4	13,350	528	3.96	0	48	61
5	14,644	322	2.20	0	8	48

Table 9: Training, validation and test counts for the encoder training and tile replacement experiments. Replaced tiles are those reinstated near another region; in each class, the same number of tiles is removed as reinstated. Percentages divide the replacement count by the original training size. The last three columns count Craters tiles. Validation and test sets stay the same. Every count refers to tiles, independent of the number of seeds or replacement sets.

Fold	Reinstated	Test	Nearby	Nearby (%)
1	486	3660	207	5.66
2	486	4497	155	3.45
3	429	5721	301	5.26
4	528	4748	319	6.72
5	322	4620	154	3.33

Table 10: Test images near reinstated training images. The training column counts reinstated tiles. The nearby test column counts distinct test centers within the original minimum separation of at least one reinstated training center. Percentages use the full test count of each fold.

Fold	DINOv2 (E/R)	DOFA (E/R)	CROMA (E/R)	Untrained ViT (E/R)
1	$-2.51/+0.00$	$-2.42/-0.91$	$-0.48/-0.19$	$+0.97/-1.22$
2	$-1.94/-0.12$	$-3.10/-0.49$	$-2.45/-0.71$	$+0.26/+1.07$
3	$-1.33/+0.24$	$+1.06/-0.01$	$-4.32/-0.44$	$-2.19/-1.04$
4	$+1.69/-0.56$	$-1.50/-0.42$	$-2.51/-0.05$	$-0.94/-0.56$
5	$-4.94/+0.04$	$-4.94/+0.66$	$-4.03/+0.43$	$+1.04/+0.44$

Table 11: Change in test error rate after replacing training tiles for logistic regression, in percentage points. Each cell gives the change among nearby test tiles followed by the change among remaining tiles. Positive values mean more disagreement with the map. Values average five replacement sets, and the table includes every encoder and fold.

Figure 8: SGD performance by region, training repeat and map mixture. Left: F1 for five training seeds in each fold, with connected means and the full seed ranges. The dashed baseline always predicts the most numerous training class. Right: error rates for modal map fractions below 0.75 and at least 0.90, across 20 eligible class and fold combinations. Each group contains at least 20 tiles per combination, and combinations receive equal weight. Craters has too few examples to enter this comparison. Untrained ViT’s markers almost coincide because its predictions are nearly constant. The plotted differences may include effects of both terrain mixture and image conditions.

Property	Current geographic evaluation	Earlier Earth transfer experiments
Titan catalog	Expanded catalog, 23,380 tiles	Original catalog, 11,371 tiles
Geography	Five longitude sectors with separation between training, validation and testing	Spatial groups without separation zones
Prediction task	Six modal Titan map classes	Six classes within Titan; five for Earth transfer
Fitting	Logistic regression, SGD and training of the final two encoder layers	Original classifier settings and additional comparisons
Other domains	Earth and Venus considered separately	Earth source labels require curation; Venus is unlabeled
Interpretation	Geographic map agreement and sensitivity to fitting	Results for the original sample and additional comparisons

Table 12: The two datasets and experimental designs. Each score refers to the images and training settings in its column. Several aspects change between the experiments, so a difference in score alone cannot identify the effect of any one change.

Encoder	Classifier training	Precision	Recall	F1
DINOv2	Earth	$24.76\pm 1.70$	$25.14\pm 0.46$	$18.31\pm 0.87$
	Titan	$\mathbf{41.54}\pm 1.53$	$\mathbf{53.25}\pm 3.95$	$\mathbf{43.29}\pm 1.84$
DOFA	Earth	$5.89\pm 0.49$	$22.99\pm 1.44$	$9.33\pm 0.71$
	Titan	$\mathbf{32.50}\pm 1.66$	$\mathbf{57.65}\pm 7.39$	$\mathbf{25.83}\pm 2.91$
CROMA	Earth	$24.04\pm 2.09$	$31.03\pm 7.45$	$11.48\pm 2.04$
	Titan	$\mathbf{37.03}\pm 2.72$	$\mathbf{66.23}\pm 6.36$	$\mathbf{35.99}\pm 3.35$

Table 13: Equal training counts by domain and class in the original dataset. Macro precision, recall and F1 are percentages with sample SD across five fold means. Each fold first averages five training seeds for each subset, then gives the five subsets equal weight. Domains use identical class counts and numbers of training updates. Undefined class scores are zero. Training sets overlap among folds and subsets, limiting inferences that require independent repetitions. Bold indicates the higher mean between Earth and Titan training for each encoder and metric.

Scaling limits	Quantization	Smoothing	DINOv2	DOFA	CROMA
Moments	Continuous	Off	$17.65\pm 0.92$	$9.40\pm 0.84$	$11.42\pm 2.02$
Moments	Continuous	On	$12.71\pm 0.68$	$\mathbf{9.56}\pm 1.94$	$8.50\pm 0.43$
Moments	256	Off	$17.62\pm 0.89$	$9.40\pm 0.84$	$11.48\pm 2.02$
Moments	256	On	$12.73\pm 0.69$	$9.55\pm 1.93$	$8.51\pm 0.43$
Percentiles	Continuous	Off	$\mathbf{18.94}\pm 0.93$	$8.67\pm 1.71$	$11.62\pm 2.74$
Percentiles	Continuous	On	$13.26\pm 0.75$	$8.90\pm 1.90$	$\mathbf{12.84}\pm 1.27$

Table 14: Eight image preprocessing conditions for the original dataset. Macro F1 is in percent, with sample SD of five fold means after averaging five training seeds. Moment scaling uses limits three population SDs below and above the training dB mean; percentile scaling uses the 1st and 99th percentiles. The 256 setting quantizes values before optional smoothing. Continuous retains values without additional quantization. These transformations alter image statistics; physical calibration and averaging of independent radar looks require information unavailable in the Titan DN. Bold indicates the highest mean F1 in each encoder column across the displayed preprocessing conditions.

Model	Macro precision	Macro recall	Macro F1
DINOv2	23.84	25.08	18.61 $\pm$ 0.81
DOFA	26.11	25.45	18.76 $\pm$ 0.83
CROMA	24.77	32.32	10.17 $\pm$ 2.02

Table 15: Classifiers trained on Earth and evaluated on Titan for the five classes present in the original Earth training set. Precision, recall and F1 are percentages from the original confusion matrices. Values after

\pm

give the sample SD of five fold means. DOFA uses identifier 13.78 for both domains in these results. Tables 13 and 14 provide additional experiments with different training samples and separately estimated image normalization. Bold indicates the highest mean in each metric column.

Domains	DINOv2	DOFA	CROMA
Earth (C-band) and Venus (S-band)	0.216	0.261	0.744
Venus (S-band) and Titan (Ku-band)	0.297	0.497	0.509
Earth (C-band) and Titan (Ku-band)	0.112	0.165	0.412

Table 16: MMD 2 between planetary image domains, averaged over 100 repeated samples with one RBF bandwidth per encoder. Each repeat uses 398 tiles per domain when Earth enters the comparison and 2,000 for Venus and Titan. Each column uses a different feature space, so numerical comparisons across encoders have no common scale. Earth and Titan give the smallest MMD 2 for DINOv2, DOFA and CROMA. The images differ in radar products, frequency, surface materials, atmosphere, acquisition and preprocessing, which prevents assigning this ordering to one physical factor.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Encoder	Epoch range used for evaluation	Attempted updates	Successful updates	AMP skips	Example presentations
DINOv2	1–14	527,750	527,475	275	16,871,750
DOFA	1–21	527,750	527,453	297	16,871,750
CROMA	1–13	527,750	527,487	263	16,871,750

Appendix

Table 17: Training totals when fitting the final two transformer blocks. Epochs start at one. The ranges identify epochs used for evaluation; the computation totals cover all 50 epochs in 25 fits per encoder. Example presentations count each use of an image, including repetitions across epochs and fits. Successful updates and updates skipped by automatic mixed precision (AMP) sum to attempted updates. Comparing computational costs with experiments that keep encoder weights unchanged requires the corresponding totals for those experiments.

Class	DINOv2	DOFA	CROMA
Plains	85.93 / 67.60 / 75.54	86.57 / 42.92 / 56.61	88.74 / 55.25 / 67.78
Dunes	53.86 / 69.11 / 60.21	48.77 / 59.99 / 52.67	48.03 / 59.81 / 51.78
Hummocky	39.06 / 59.61 / 46.03	43.49 / 47.75 / 39.30	48.22 / 58.61 / 49.37
Labyrinths	24.78 / 45.62 / 29.85	6.70 / 48.51 / 11.45	12.12 / 57.02 / 19.12
Lakes	9.94 / 20.01 / 11.29	7.90 / 31.51 / 10.67	16.03 / 42.12 / 19.97
Craters	2.56 / 6.24 / 3.33	1.99 / 40.85 / 3.57	1.98 / 17.90 / 3.53

Appendix

Table 18: Precision, recall and F1 by map class for SGD classifiers on the enlarged catalog, in percent. Entries are precision / recall / F1. Each ratio is calculated for each classifier and fold, then averaged over classifier seeds in the fold and equally over folds. Figure 2 gives the class counts. Bold indicates the highest mean F1 across encoders for each class.

Figure 9: SGD classification in the current geographic evaluation. Rows give the map class; columns give the predicted class. Percentages first average classifier seeds in each fold, then give all folds equal weight. The count line includes each distinct test tile once across folds, irrespective of the number of classifier seeds. Diagonal values are mean class recall. Precision and F1 also depend on predictions for other map classes; their primary estimates average ratios calculated before combining folds. Craters counts range from one to 61 tiles per fold. The labels describe broad mapped units.

Figure 10: Examples from the original Earth, Titan and Venus collections. Columns show Earth site labels, whose landform assignments still need review, and Titan classes from the expert map. Earth contains no Hummocky examples. Venus images have no class labels or implied class correspondence across columns. Each image spans a nominal $45\text{\,}\mathrm{km}$ by $45\text{\,}\mathrm{km}$ . Earth linear $\sigma^{0}$ is shown as $10\log_{10}(\sigma^{0})$ in dB; Titan and Venus use display digital numbers. Each display stretches the 2nd to 98th percentiles of valid pixels, so brightness is relative to that image. Magenta and diagonal hatching identify invalid SAR pixels, including Earth pixels filled during preprocessing. The examples illustrate image appearance; geological equivalence requires independent evidence.

Model	Linear classifier		$k$ -NN
	Macro recall	$\kappa$	Macro recall	$\kappa$
DINOv2 ViT-B/14	56.0 $\pm$ 2.7	0.502 $\pm$ 0.039	32.4 $\pm$ 1.4	0.388 $\pm$ 0.016
DOFA ViT-B/16	58.1 $\pm$ 8.0	0.296 $\pm$ 0.030	33.5 $\pm$ 0.8	0.400 $\pm$ 0.020
CROMA ViT-B/8	66.1 $\pm$ 2.4	0.432 $\pm$ 0.038	38.7 $\pm$ 0.7	0.440 $\pm$ 0.034
Untrained ViT-B/14	17.0 $\pm$ 0.7	0.003 $\pm$ 0.006	19.0 $\pm$ 0.9	0.022 $\pm$ 0.010

Appendix

Table 19: Training and testing on Titan in the original catalog. Values give means and sample standard deviations over five geographic folds. For linear classifiers, each metric is first averaged over the five training seeds in each fold;

k

-NN gives one reproducible result per fold. Among linear classifiers, CROMA has the highest six-class macro recall, 66.1%, and DINOv2 the highest

\kappa

, 0.502. Untrained ViT uses an untrained encoder whose original architecture remains incompletely verified, as described in Section B.2.1 . Bold indicates the highest mean in each metric column for its classifier.

Comparison	Titan $\to$ Titan	Earth $\to$ Titan
Majority class	16.7 $\pm$ 0.0	20.0 $\pm$ 0.0
Class-prior random	16.7 $\pm$ 0.2	20.0 $\pm$ 0.9
Raw intensity statistics	54.8 $\pm$ 7.3	28.1 $\pm$ 4.9
Intensity histogram	42.8 $\pm$ 4.8	19.7 $\pm$ 1.0
Texture descriptors	39.3 $\pm$ 2.5	19.7 $\pm$ 0.8
Combined classical	44.8 $\pm$ 1.9	20.2 $\pm$ 0.8

Appendix

Table 20: Additional classification comparisons on the original dataset. Macro recall is in percent, with mean

\pm

sample standard deviation across five geographic folds after averaging seeds in each fold. Classical-feature classifiers give identical results across seeds; only class-prior random predictions and CNN fits vary with seed. Transfer excludes Hummocky. Classical classifiers fit the five Earth classes, and the CNN’s Hummocky logit is masked before softmax. All CNN predictions before masking already belong to a represented class, so masking leaves the test predictions unchanged. Bold indicates the highest mean recall in each training-domain column among the methods shown.

Figure 11: Linear classifiers with encoder weights unchanged in the original dataset. Left: macro recall after Titan or Earth training, evaluated on the five classes present in Earth. Transfer using pretrained models stays above the 20% five-class benchmark and is 29.8 to 32.4 percentage points below Titan training. Right: transfer recall by class, excluding Hummocky because Earth supplies no examples. Dunes has the highest mean transfer recall for DINOv2 and DOFA, at 89.6% and 83.9%. CROMA’s mean Craters recall is 90.0%, with only 17 distinct test tiles across all folds. Error bars show sample SD across five fold means after averaging classifier seeds. Identifying a physical property that transfers between Earth and Titan would require evidence beyond recall.

Model	Linear classifier		$\Delta$ recall	$k$ -NN
	Macro recall	$\kappa$		Macro recall	$\kappa$
DINOv2 ViT-B/14	25.1 $\pm$ 0.4	0.135 $\pm$ 0.017	$-$ 29.8	25.9 $\pm$ 1.0	0.146 $\pm$ 0.024
DOFA ViT-B/16	25.5 $\pm$ 1.1	0.121 $\pm$ 0.014	$-$ 32.1	26.1 $\pm$ 0.6	0.182 $\pm$ 0.018
CROMA ViT-B/8	32.3 $\pm$ 3.7	0.034 $\pm$ 0.022	$-$ 32.4	24.0 $\pm$ 2.5	0.094 $\pm$ 0.032
Untrained ViT-B/14	20.0 $\pm$ 0.0	0.000 $\pm$ 0.000	$+$ 3.6	20.2 $\pm$ 0.9	0.004 $\pm$ 0.020

Appendix

Table 21: Classifiers trained on Earth and evaluated on the five Titan classes present in Earth training. Values are mean

\pm

sample standard deviation across five folds. Linear-classifier values are first averaged over seeds in each fold;

k

-NN gives one reproducible result per fold.

\Delta

is the difference from the corresponding classifier trained on Titan and evaluated on the same classes. The linear classifiers trained on Earth and on Titan both retain six output logits. Untrained ViT remains near the class-frequency benchmark despite its positive difference, providing no evidence of useful transfer. CROMA has the highest transfer mean at 32.3%, followed by DOFA at 25.5% and DINOv2 at 25.1%. Comparisons require identical normalization, catalogs and image sampling. Bold indicates the highest mean in each recall or kappa column for its classifier. Transfer differences remain in ordinary type.

Model	Earth	Titan stratified	Titan uniform
DINOv2	25.1 $\pm$ 0.4	43.9 $\pm$ 1.1	41.3 $\pm$ 2.1
DOFA	25.5 $\pm$ 1.1	45.9 $\pm$ 5.2	40.6 $\pm$ 3.8
CROMA	32.3 $\pm$ 3.7	51.1 $\pm$ 4.3	45.4 $\pm$ 4.4

Appendix

Table 22: Effect of training-set size on recall for the five Titan classes present in Earth. Every training subset contains 352 tiles. Values are macro recall in percent, with sample standard deviation across five fold means after averaging five seeds. The experiment equalizes total count while allowing class composition and site diversity to differ. The values reproduce the workshop training-size calculations. Bold indicates the highest mean recall across encoders for each training setting.

Figure 12: Feature distributions in the original dataset. Top: Earth and Titan MMD 2 from repeated equal-size samples, shown as mean and sample standard deviation across five fold means for pretrained encoders. The random encoder is omitted from this subplot. Stable arithmetic confirms its repeated estimate, but its original permutation statistic is invalid. Bottom: MMD 2 for Earth, Venus and Titan, shown separately for each encoder. Each encoder uses one bandwidth across domain pairs and equal sample counts in each pair. Error bars give sample SD across 100 draws and describe subsampling variation. Earth and Titan have the smallest estimates for the pretrained encoders under this estimator. Interpretation is restricted to each model’s feature space; physical causes and comparisons of magnitude between models require further evidence.

Model	MMD 2	Original $\boldsymbol{d_{\mathcal{A}}}$	Centroid Dist.
DINOv2 ViT-B/14	0.113	1.948	25.93
DOFA ViT-B/16	0.183	1.918	2.32
CROMA ViT-B/8	0.430	1.990	6.18
Untrained ViT-B/14	0.112	1.817	0.002

Appendix

Table 23: Earth and Titan feature comparisons for the original dataset, averaged over five geographic folds. MMD 2 is the biased RBF-kernel V-statistic averaged over 100 equal-size resamples in each fold. The Untrained ViT estimate is recalculated from its original arrays using stable arithmetic; its fold standard deviation is given in the text. The proxy

\mathcal{A}

-distance column retains the original estimates with unequal domain counts. The text gives the later estimates using equal domain counts under a different fitting design. Centroid distances use the unnormalized coordinates of each model. These comparisons characterize feature distributions; identifying physical causes of transfer error requires additional evidence.

Figure 13: t-distributed stochastic neighbor embedding (t-SNE) of original fold 1 features: DINOv2 at left and CROMA at right. Each model uses all 352 Earth training tiles and 500 Titan test tiles. A common random stream with seed 42 draws different Titan subsets for the models. The same coordinates appear twice, colored by domain above and by Earth site label or Titan map class below. Both fits use perplexity 30 and seed 42. The plotted coordinate systems are arbitrary. Separation or overlap in the original feature spaces, and any relation to transfer performance, require measurements beyond these projections.

Model	Biased	Unbiased	Omit each Earth site	Resample Titan regions	Max perm. $p$
DINOv2	0.113	0.110	0.115	0.113	0.001
DOFA	0.183	0.180	0.182	0.183	0.001
CROMA	0.430	0.427	0.430	0.431	0.001

Appendix

Table 24: Sensitivity of Earth and Titan MMD 2 to estimation and resampling choices. MMD 2 estimates are means across five Titan geographic folds. The final column gives the largest permutation

p

-value. Each row uses that model’s feature space, which sets the scale and prevents direct comparisons of magnitude across models.

	Plains (7441)		Dunes (2215)		Hummocky (1014)		Labyrinths (257)		Lakes (362)		Craters (17)
Model	T $\to$ T	E $\to$ T	T $\to$ T	E $\to$ T	T $\to$ T	E $\to$ T	T $\to$ T	E $\to$ T	T $\to$ T	E $\to$ T	T $\to$ T	E $\to$ T
DINOv2	71.2	33.5	76.0	89.6	61.8	N/A	50.1	2.2	75.3	0.0	1.6	0.0
DOFA	39.7	27.9	63.5	83.9	60.6	N/A	67.1	12.5	81.7	3.0	36.0	0.0
CROMA	57.6	4.7	70.8	36.0	72.8	N/A	67.0	27.7	83.4	3.2	44.7	90.0
Untrained ViT	44.0	0.0	36.0	20.0	20.0	N/A	0.0	40.0	1.9	20.0	0.0	20.0

Appendix

Table 25: Recall by class for linear classifiers, in %, averaged over seeds in each fold and then over five geographic folds. Parentheses give the total number of distinct test tiles outside Selk; each tile is tested once. N/A marks Hummocky because Earth supplies no training examples. Small class counts make estimates unstable, especially for Craters, with

n=17

. Bold indicates the highest mean recall among all four encoders for each class and training domain.

Figure 14: Model uncertainty and disagreement for the original 65 Selk tiles. The star is the center at 6°N, 161°E; the dashed curve lies 5° from it on a sphere. Color shows the mean of entropy and disagreement after each is scaled separately to $[0,1]$ using its minimum and maximum over these 65 tiles. Numbered rings mark the three tiles with the highest scores in rank order. The first is the tile at 6.64°N, 164.83°E. Scores rank only this set of images and models. Interpreting them as calibrated probabilities of anomaly, novelty, scientific value, risk or mission priority would require independent validation.

Model	Before CORAL	CORAL	MMD 2 before/after
DINOv2	25.1	23.1	0.113 / 0.009
DOFA	25.5	26.4	0.183 / 0.006
CROMA	32.3	35.3	0.430 / 0.009

Appendix

Table 26: Earth-to-Titan classification before and after CORAL, measured by macro recall in percent over five classes. Metrics average seeds in each fold and then the five folds. CORAL uses unlabeled Titan training features and evaluates classification on separate test regions. Both MMD 2 calculations use the fitting samples and estimate bandwidth separately before and after the transformation. Their ratio describes those changing kernels on training features; measuring a reduction for one kernel on independent data would require a separate calculation. Bold indicates the higher mean recall before or after CORAL for each encoder. MMD values remain in ordinary type.

	Encoder weights unchanged		Train final 2 blocks
Model	Recall	$\kappa$	Recall	$\kappa$
DINOv2	55.9 $\pm$ 2.9	0.502 $\pm$ 0.038	60.5 $\pm$ 1.9	0.531 $\pm$ 0.041
DOFA	57.8 $\pm$ 8.4	0.295 $\pm$ 0.031	30.6 $\pm$ 2.4	0.124 $\pm$ 0.012
CROMA	66.1 $\pm$ 2.5	0.432 $\pm$ 0.038	62.4 $\pm$ 2.2	0.532 $\pm$ 0.036

Appendix

Table 27: Training the final two transformer blocks on the original Titan dataset for 50 epochs with AdamW, cosine annealing and batch size 8. Values are mean

\pm

sample standard deviation across five geographic folds after averaging five training seeds per fold. These runs used the implementation before weighted gradient accumulation was corrected. Table 4 gives the current implementation on the enlarged catalog. Bold indicates the higher mean for each encoder and metric between classifier training alone and training the final two encoder blocks.

Figure 15: Results from the original implementation, before weighted gradient accumulation was corrected. Left: Cohen’s $\kappa$ for linear classifiers using encoders whose weights remain unchanged and for training the final two transformer blocks. Bars are means over five geographic folds; error bars give the sample standard deviation of fold means. Right: CROMA confusion matrix after training the two blocks, combining predictions over all folds and seeds before normalizing each row. For CROMA, higher $\kappa$ together with lower macro recall reflects a redistribution of errors. The combined counts describe the predictions, while uncertainty calculations use fold means.

Figure 16: Sensitivity to the number of labeled Titan training images in the original dataset, using a loss without class weights. Lines give six-class macro recall averaged over five seeds for sampling and training in each fold, then over five geographic folds. Bands show the sample standard deviation across fold means. At

n=10

, CROMA has the highest mean, 26.8% versus 25.1% for DINOv2; this difference is small relative to the fold spread. DINOv2 gives the highest result using all training tiles, 50.3%. All denotes each fold’s complete training set, plotted at the mean training size for DINOv2.

Terrain Class	Site Name	Location	Analog Rationale
Dunes	Namib Sand Sea	Namibia	Linear longitudinal dunes
	Grand Erg Oriental	Algeria	Large-scale erg morphology
	Rub’ al Khali	Saudi Arabia	Massive sand sea
Craters	Manicouagan	Canada	Large impact structure
Craters	Mistastin	Canada	Complex crater with melt sheet
Lakes	Arctic thermokarst	Siberia	Shallow depressions

Appendix

Table 28: Earth site labels and intended morphological analogies. Assignments still need verification for individual image areas, especially offshore Arctic and Icelandic areas and the site incorrectly named Wadi Rum. Sentinel-1 IW VV products are processed with ESA SNAP and sampled at

351.5625\text{\,}\mathrm{m}

/pixel.

Parameter	Value	Source
Cassini frequency	$13.78\text{\,}\mathrm{GHz}$	Elachi et al. [1]
Cassini wavelength	$2.176\text{\,}\mathrm{cm}$	$c/f$
Sentinel-1 frequency	$5.405\text{\,}\mathrm{GHz}$	ESA
Sentinel-1 wavelength	$5.547\text{\,}\mathrm{cm}$	$c/f$
Titan sphere radius	$2574.7\text{\,}\mathrm{km}$	Benchmark geometry
Image width	$45\text{\,}\mathrm{km}$	Benchmark geometry

Appendix

Table 29: Radar frequencies and benchmark geometry. Frequencies identify the instruments; the experiments vary several factors together and leave the individual frequency effects undetermined.

Model	Patch size	Encoder channels	Supplied SAR input	Weights source
DINOv2	14	3, RGB	One SAR channel repeated three times	Meta AI
DOFA	16	Variable	One SAR channel; original extraction used 13.78 for Earth and Titan	earthflow/DOFA
CROMA	8	2, VV and VH	One SAR channel repeated twice; $120\times 120$ pixels	CROMA authors
Untrained ViT	14	1	One SAR channel and a single-channel stem	Seed 42

Appendix

Table 30: Model configurations used with the original dataset. The architecture chosen by the Untrained ViT fallback and its exact initial weights remain unverified; see Section B.2.1 .

Explore similar work

Jun 22, 2026cs.CV

UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation

Vision Transformers (ViT) dominate computer vision. However, their reliance on rigid patch projectors hinders transfer to Earth Observation (EO), where input modalities, scales, and resolutions vary widely. We introduce UniverSat, a ViT-style backbone built around a Universal Patch Encoder that maps patches from arbitrary spatial, spectral, and temporal resolutions, and from both optical and non-optical sensors, into a shared embedding space with a shared set of weights. This enables training a single model on heterogeneous multimodal corpora via self-supervision, yielding robust, sensor-agnostic spatial features. We validate this approach with strong results across classification and segmentation on standard EO benchmarks from GeoBench, PANGEABench, and SpectralEarth. Our code and models are available at https://github.com/gastruc/UniverSat.

Yohann Perron, Guillaume Astruc, Nicolas Gonthier +2

LIGM, Ecole Nationale des Ponts et Chaussées, IP Paris, Univ Gustave Eiffel, CNRS · EFEO · LASTIG, Univ Gustave Eiffel, IGN, ENSG +2

Trustworthy Earth Observation Decisions Earth Observation Self-Supervised Vision Transformers Detection Transformer Vision Transformer Transformer Architectures+1

Ask PDF

May 20, 2026eess.IV

Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis

Field-scale crop maps support supply-chain forecasting and policy, yet statewide crop identification still often depends on retrospective surveys or remote-sensing workflows built around hand-engineered spectral features. Those pipelines can be accurate, but they require repeated preprocessing and often lose robustness across years. This study evaluated whether Google DeepMind's AlphaEarth geospatial embeddings can serve as an analysis-ready alternative for mapping processing tomato systems in California. LandIQ 2018 crop polygons were used to assemble a balanced reference dataset of 4,742 tomato and 4,742 non-tomato fields. For each polygon, 64-band AlphaEarth embedding chips were extracted and aligned with binary masks, then divided into spatially independent training (n = 6,638), validation (n = 1,422), and test (n = 1,424) sets. A U-Net segmentation model was trained on AWS SageMaker using a composite masked binary cross-entropy and soft Dice loss. To complement hard predictions, Monte Carlo dropout was retained at inference and repeated 100 times per chip to estimate predictive mean and variance. On the independent test set, the model achieved 99.19% pixel accuracy, 98.69% precision, 99.40% recall, 99.04% F1 score, 98.11% intersection over union, and 99.02% chip accuracy. Uncertainty maps were consistently highest near field edges and low within field interiors. The results show that AlphaEarth embeddings retain crop-relevant spatial and temporal structure and can support accurate, field-scale tomato mapping without manual feature engineering.

Mohammadreza Narimani, Alireza Pourreza, Parastoo Farajpoor

Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA

Agricultural Monitoring Deep Learning Chip Design

Ask PDF

Jun 25, 2026cs.CV

Tessellating The Earth

Geolocation encoders, which map geographic coordinates to learned representations, are emerging as an effective means of capturing visual and non-visual characteristics from a latitude-longitude pair alone. However, existing approaches project coordinates onto fixed bases (e.g., spherical harmonics), allocating representational capacity uniformly and devoting equal resources to the open ocean and to a developing city. We introduce Tessellating the Earth (TTE), a location encoder built from learnable Spherical Voronoi partitions that concentrates representational capacity where it is needed in a fully differentiable, end-to-end manner. Each Voronoi site carries its own embedding and migrates during training toward discriminative areas. To bridge the gap between local spatial structure and global semantic understanding, we introduce \emph{global semantic tokens}: a set of shared learnable concept tokens that distill semantic knowledge from the satellite imagery into a compact vocabulary the location encoder can reference at inference, enabling geographically distant sites covering similar environments to share semantics. TTE sets a new state of the art for location encoders across a suite of geospatial classification and regression tasks, and achieves the strongest results when used as a geographic prior for fine-grained species classification on iNaturalist-2018. Code, and weights are available at https://github.com/mvrl/TTE.

Daniel Cher, Hamza Iqbal, Eric Xing +2

Washington University in St. Louis

Semantic Geometry Voronoi-Type Losses Spatial Priors Earth

Ask PDF