EarthScape: A Multimodal Benchmark for Geologic Mapping and Surface-Aware Learning under Geographic Domain Shift

Matthew A. Massey; Nusrat Munia; Abdullah-Al-Zubaer Imran

cs.CVMar 19, 2025

EarthScape: A Multimodal Benchmark for Geologic Mapping and Surface-Aware Learning under Geographic Domain Shift

Authors: Matthew A. Massey, Nusrat Munia, Abdullah-Al-Zubaer Imran

Organizations: Kentucky Geological Survey University of Kentucky Lexington, KY 40506, USA · Department of Computer Science University of Kentucky Lexington, KY 40506, USA

Abstract

Understanding natural Earth surface processes requires integrating topography, imagery, hydrology, and human infrastructure while remaining robust to geographic variation, yet few benchmarks support both challenges. We introduce EarthScape, a multimodal benchmark for geologic mapping and surface-aware learning under geographic domain shift. EarthScape combines aerial imagery, LiDAR-derived elevation, multi-scale terrain derivatives, and contextual vector data into 37 co-registered input channels per sample, together with pixel-level geologic masks. Its geographically separated regions share a common label space, enabling controlled evaluation of in-domain and cross-region generalization under realistic class imbalance and spatial complexity. Benchmark experiments show that engineered terrain representations transfer more robustly than raw elevation, while spatial support and multimodal terrain inputs strongly influence performance. Fusion effectiveness depends on backbone and evaluation domain, with more complex cross-attention providing no consistent advantage over simpler strategies. These representation trends extend across geospatial foundation models and semantic segmentation, while class-level results show that no single terrain feature or scale is optimal across geologic classes. Together, EarthScape provides a benchmark for studying representation design, multimodal fusion, and geographic robustness in surface-aware geospatial learning. Our source code is available at https://github.com/masseygeo/earthscape and the dataset at https://doi.org/10.13023/kgs.data.05.01.2025.

Figures & tables

Figure 1: EarthScape study regions and selected modalities. Left: in-domain (ID) Warren and cross-domain (CD) Hardin subsets with latitude–longitude grid lines and centroid distance between areas. Right: selected channels from one $256\times 256$ patch (location shown in red on map), including the geologic mask, raw DEM and RGB inputs, hydrography (NHD), and DEM-derived terrain features at multiple spatial supports. Each patch is overlaid on a hillshade image for visualizing surface morphology, and geologic class boundaries (black lines) are shown for reference.

Figure 2: EarthScape dataset complexity. Top left: class frequency ( $n_{c}$ ) with percentages indicating the relative proportion of patches containing each class. Bottom left: sample frequency ( $n$ ) of label cardinality ( $|Y_{i}|$ ). Right: spatial distribution of $|Y_{i}|$ illustrating geospatial variation in patch complexity.

Class	$n_{c}$	Mean Area	MAR	ENS	IRLbl
Qr	29271	0.651	0.702	9464.6	1.0
Qal	18801	0.089	0.058	8474.4	1.6
Qc	13768	0.142	0.148	7476.3	2.1
af1	10910	0.051	0.035	6641.4	2.7
Qca	7669	0.061	0.054	5355.7	3.8
Qat	1435	0.006	0.004	1336.9	20.4

Table 1: Label statistics and imbalance metrics for EarthScape, including global frequency (

n_{c}

), mean class-area proportion, majority area rate (MAR), effective number of samples (ENS) [ 19 ] , and the imbalance ratio per label (IRLbl) [ 11 ] .

Figure 3: SGMap-Net and fusion baselines. Left: SGMap-Net accepts an arbitrary set of modality inputs, adapts each branch to three channels, encodes them with a shared backbone, and fuses modality features using multi-head attention and attention-weighted aggregation before multilabel classification. Right: simplified baselines used to compare fusion strategy, including early channel stacking before encoding and mid-level concatenation after modality-specific encoding.

Fusion	ResNeXt					ViT
Fusion	Input	$\bar{\text{F1}}_{\text{ID}}\pm\text{SD}$	$\bar{\text{F1}}_{\text{CD}}\pm\text{SD}$	$\Delta$	$\Delta\%$	Input	$\bar{\text{F1}}_{\text{ID}}\pm\text{SD}$	$\bar{\text{F1}}_{\text{CD}}\pm\text{SD}$	$\Delta$	$\Delta\%$
-	DEM	0.682 $\pm$ 0.003	0.526 $\pm$ 0.026	0.156	22.9	DEM	0.655 $\pm$ 0.015	0.524 $\pm$ 0.007	0.131	20.0
-	S-10	0.653 $\pm$ 0.015	0.587 $\pm$ 0.014	0.066	10.1	S-5	0.642 $\pm$ 0.010	0.587 $\pm$ 0.015	0.055	8.6
Stacking	PrC-ms	0.654 $\pm$ 0.014	0.593 $\pm$ 0.018	0.061	9.3	S-ms	0.650 $\pm$ 0.008	0.583 $\pm$ 0.021	0.067	10.3
Concat.	PrC-ms	0.606 $\pm$ 0.011	0.509 $\pm$ 0.027	0.097	16.0	S-ms	0.659 $\pm$ 0.010	0.569 $\pm$ 0.016	0.090	13.7
CA	PrC-ms	0.612 $\pm$ 0.020	0.536 $\pm$ 0.024	0.076	12.4	S-ms	0.636 $\pm$ 0.010	0.577 $\pm$ 0.016	0.059	9.3

Table 2: Representative SGMap-Net classification results for ResNeXt and ViT backbones. Inputs were selected by the highest combined ID

+

CD F1 within the single modality, multi-scale, and multimodal experiments; all fusion strategies are shown for the multi-scale and multimodal selections. Values report mean and standard deviation (SD) across five random seeds;

\Delta=\bar{\mathrm{F1}}_{\mathrm{ID}}-\bar{\mathrm{F1}}_{\mathrm{CD}}

and

\Delta\%=100(\bar{\mathrm{F1}}_{\mathrm{ID}}-\bar{\mathrm{F1}}_{\mathrm{CD}})/\bar{\mathrm{F1}}_{\mathrm{ID}}

denote absolute and relative cross-domain degradation, respectively.

Encoder	Fusion	$n$	PCC		SRCC
Encoder	Fusion	$n$	$r$	$p_{\text{Holm}}$	$\rho$	$p_{\text{Holm}}$
ResNeXt	-	35	0.248	0.451	0.368	0.090
	Stacking	15	0.369	0.451	0.639	0.041
	Concat.	15	0.497	0.237	0.454	0.179
	CA	15	0.286	0.451	0.318	0.248
ViT	-	35	0.598	0.001	0.513	0.010

Table 3: Statistical association between cross-domain F1 degradation (

\Delta=\mathrm{ID}-\mathrm{CD}

) and maximum mean discrepancy (MMD) for unimodal, and multiscale

+

multimodal experiments stratified by encoder and fusion method. PCC (

r

): Pearson’s correlation coefficient; SRCC (

\rho

): Spearman’s rank correlation coefficient. All

p

-values were Holm-corrected for multiple comparisons. Significant correlations marked in bold .

Encoder	Domain	Comparison	$n$	$t$	$p_{\mathrm{Holm}}$
ResNeXt	ID	Stacking–Concat.	15	8.045	$<$ 0.001
		Stacking–CA	15	10.209	$<$ 0.001
		Concat.–CA	15	-1.019	0.326
	CD	Stacking–Concat.	15	2.167	0.144
		Stacking–CA	15	0.939	0.364
		Concat.–CA	15	-1.871	0.165

Table 4: Statistical comparisons of stacking, concatenation, and cross-attention (CA) fusion paired by input configurations, stratified by encoder and evaluation domain (ID, in-domain; CD, cross-domain). All

p

-values were Holm-corrected within each encoder–domain family. Significant comparisons marked in bold .

Model	Input	${\text{F1}}_{\text{ID}}$	${\text{F1}}_{\text{CD}}$	$\Delta$	$\Delta\%$
Copernicus-FM	RGB+NIR	0.592	0.341	0.251	42.4
DOFA	RGB+NIR	0.602	0.386	0.216	35.9
Panopticon	RGB+NIR	0.632	0.507	0.125	19.8
Galileo	RGB+NIR	0.548	0.286	0.262	47.8
	DEM+S-5	0.593	0.508	0.085	14.3
	RGB+NIR+DEM+S-5	0.548	0.272	0.276	50.4

Table 5: Foundation model classification performance using compatible EarthScape inputs. In-domain (ID) and cross-domain (CD) macro-averaged F1 for the best-performing learning rates, with absolute (

\Delta

) and relative (

\Delta\%

) cross-domain degradation.

Appendix figures & tables43 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Example surficial geologic maps overlaid on hillshade images to emphasize topographic context. The spatial correspondence between geologic classes and landscape features is visually apparent. The black grid indicates the layout of EarthScape patches, each measuring $256\times 256$ pixels ( $390.14\times 390.14$ m) with 50% overlap. Red squares in the upper left corners highlight a single patch.

Geologic Class	Notation	Encoding	Geomorphic Process	Characteristic Topographic Expression and Visual Cues
Artificial fill	af1	0	Anthropogenic	Rectilinear shapes with sharp edges; surface patterns inconsistent with natural terrain.
Alluvium	Qal	1	Fluvial (active)	Relatively broad, low-relief valleys, channels, and basins.
Alluvial fans	Qaf	2	Debris Flow	Small, lobate deposits at tributary mouths; wedge-shaped morphology; often isolated.
Terrace deposits	Qat	3	Fluvial (relict)	Flat benches elevated above active floodplains; pronounced scarps; isolated or paired.
Colluvium	Qc	4	Gravity-driven (active)	Steep slopes ( $\geq 12^{\circ}$ ); irregular, rough terrain.
Colluvial aprons	Qca	5	Gravity-driven (stable)	Wedge-shaped deposits along slope bases; concave profiles.

Appendix

Table 6: Descriptions of surficial geologic units represented in EarthScape.

Figure 5: Map of the central United States showing the publicly available 1:24,000-scale surficial geologic maps. Red lines show boundaries of major geologic provinces, which provide geological constraints for generalizability. The Interior Low Plateaus and adjacent Appalachian Plateaus may provide suitable settings for geographic transfer based on shared terrain, bedrock, and geomorphic processes. In contrast, the glaciated Central Lowlands and Coastal Plain are characterized by fundamentally different surficial processes and materials.

Terrain Feature	Manuscript Notation	Computational Input		Spatial Support
Terrain Feature	Manuscript Notation	Source DEM GSD	Kernel Size	Spatial Support
Elevation Percentile	EP-5	5 ft ( $\approx$ 1.52 m)	$5\times 5$	625 ft 2 ( $\approx$ 58 m 2 )
	EP-11	5 ft ( $\approx$ 1.52 m)	$11\times 11$	3025 ft 2 ( $\approx$ 281 m 2 )
	EP-21	5 ft ( $\approx$ 1.52 m)	$21\times 21$	11025 ft 2 ( $\approx$ 1024 m 2 )
	EP-51	5 ft ( $\approx$ 1.52 m)	$51\times 51$	65025 ft 2 ( $\approx$ 6041 m 2 )
	EP-101	5 ft ( $\approx$ 1.52 m)	$101\times 101$	255025 ft 2 ( $\approx$ 23693 m 2 )

Appendix

Table 7: Terrain-feature notation used throughout the manuscript, with the source DEM resolution and kernel size used for computation. Spatial support is provided as an interpretive measure of the effective spatial area. Native processing units are feet, but metric equivalents are reported throughout the manuscript. S, PrC, and PlC are computed from native or resampled DEMs, whereas EP and SDS are computed on the native DEM using square neighborhood windows of varying size.

Figure 6: Example patch from the Warren County area showcasing the target mask and 37 predictive channels available in EarthScape. Channels are displayed from top left to bottom right: target mask, RGB aerial imagery, NIR aerial imagery, DEM, NHD hydrologic features, OSM infrastructure, six spatial scales of S, PrC, and PlC derived from downsampled DEMs, and multiple scales of SDS and EP calculated using six kernel sizes with the original DEM.

Figure 7: Example patch from the Hardin County area showcasing the target mask and 37 predictive channels available in EarthScape. Channels are displayed from top left to bottom right: target mask, RGB aerial imagery, NIR aerial imagery, DEM, NHD hydrologic features, OSM infrastructure, six spatial scales of S, PrC, and PlC derived from downsampled DEMs, and multiple scales of SDS and EP calculated using six kernel sizes with the original DEM.

Input	MMD	SD	95% CI
DEM	0.273	0.003	0.001
RGB	0.136	0.006	0.002
NIR	0.287	0.002	0.001
NHD	0.036	0.006	0.002
OSM	0.028	0.004	0.001
EP-5	0.056	0.006	0.002

Appendix

Table 8: Mean maximum mean discrepancy (MMD), standard deviation (SD), and 95% confidence interval half widths (CI) for raw feature input representations. Lower MMD values indicate greater distributional similarity between regions, whereas higher values indicate stronger geographic divergence.

Figure 8: EarthScape benchmark split locations together with class frequency ( $n_{c}$ ) and intra-patch spatial support ( $a_{c}$ ) distributions for the Warren (in-domain, ID) and Hardin (cross-domain, CD) regions. Geographic locations are shown in Fig. 1 .

Region	Domain	Candidate	Split	Split	ID Split
Region	Domain	Patches	Split	Count	Proportion
Warren	ID	23566	Training	8416	78.5%
			Validation	768	7.2%
			Testing	1536	14.3%
Hardin	CD	7452	Testing	1536	-

Appendix

Table 9: Total candidate patches for the in-domain (ID) Warren and cross-domain (CD) Hardin subsets, and patch counts and proportions for the benchmark splits. Proportions are relative to the total number of selected samples in the ID splits.

$w_{c}^{+}$	$\gamma$	$\alpha$	Input	ID	CD	$\Delta$	$\Delta\%$	ID $+$ CD
1	0	-	DEM	0.683	0.493	0.191	27.9	1.176
			S-5	0.639	0.601	0.038	5.9	1.240
			S-ms	0.644	0.580	0.064	9.9	1.225
			DEM+S-ms	0.655	0.585	0.071	10.8	1.240
								1.220
ICF	0	-	DEM	0.658	0.489	0.169	25.7	1.147

Appendix

Table 10: Macro-averaged F1 for in-domain (ID) and cross-domain (CD) performance for a loss-function ablation across four representative input configurations. Loss settings are parameterized by positive-class weight (

w_{c}^{+}

), focal focusing parameter (

\gamma

), and focal balancing parameter (

\alpha

); binary cross-entropy corresponds to

w_{c}^{+}=1

and

\gamma=0

. Transfer degradation was computed as

\Delta=\mathrm{ID}-\mathrm{CD}

and reported both in raw F1 scores and percentages (

\Delta\%

). Italicized values indicate mean ID

+

CD across input configurations; the mean ID

+

CD in bold denotes the selected loss setting.

Panel of Figure 9 — Figure 9: Representative training behavior for single-modality and complex multimodal SGMap-Net configurations. Loss and accuracy are shown for the training and validation sets across epochs. Dashed vertical lines indicate the checkpoint selected according to minimum validation loss.

Encoder	Input	Fusion	Run 1*		Run 2		Run 3		Run 4		Run 5		Mean $\pm$ SD
Encoder	Input	Fusion	ID	CD	ID	CD	ID	CD	ID	CD	ID	CD	ID	CD
ResNeXt	DEM	-	0.687	0.508	0.680	0.543	0.680	0.530	0.681	0.555	0.680	0.492	0.682 $\pm$ 0.003	0.526 $\pm$ 0.026
	S-10	-	0.674	0.599	0.639	0.577	0.651	0.569	0.663	0.592	0.639	0.600	0.653 $\pm$ 0.015	0.587 $\pm$ 0.014
	PrC-ms	Stacking	0.664	0.614	0.631	0.582	0.652	0.604	0.658	0.594	0.664	0.569	0.654 $\pm$ 0.014	0.593 $\pm$ 0.018
	PrC-ms	Concat.	0.598	0.500	0.614	0.504	0.592	0.473	0.609	0.520	0.619	0.547	0.606 $\pm$ 0.011	0.509 $\pm$ 0.027
	PrC-ms	CA	0.599	0.504	0.645	0.542	0.614	0.570	0.592	0.525	0.608	0.539	0.612 $\pm$ 0.020	0.536 $\pm$ 0.024

Appendix

Table 11: Macro-averaged F1 for in-domain (ID) and cross-domain (CD) performance across five training runs with different random seeds for representative benchmark configurations. Run 1* corresponds to the fixed-seed run used for the main experiments, while Runs 2–5 used additional random seeds to quantify performance variability. Mean and standard deviation (SD) were computed across all five runs. Best and second-best scores within each run and summary column are indicated in bold and underlined , respectively.

Encoder	Input	Fusion	Split 1*		Split 2		Split 3		Split 4		Split 5		Mean $\pm$ SD
Encoder	Input	Fusion	ID	CD	ID	CD	ID	CD	ID	CD	ID	CD	ID	CD
ResNeXt	DEM	-	0.687	0.508	0.669	0.535	0.664	0.506	0.693	0.562	0.704	0.521	0.683 $\pm$ 0.017	0.526 $\pm$ 0.023
	S-10	-	0.674	0.599	0.656	0.585	0.660	0.590	0.668	0.615	0.657	0.599	0.663 $\pm$ 0.008	0.598 $\pm$ 0.011
	PrC-ms	Stacking	0.664	0.614	0.665	0.611	0.663	0.578	0.670	0.607	0.643	0.597	0.661 $\pm$ 0.010	0.601 $\pm$ 0.015
	PrC-ms	Concat.	0.598	0.500	0.617	0.515	0.614	0.506	0.576	0.469	0.616	0.503	0.604 $\pm$ 0.018	0.499 $\pm$ 0.017
	PrC-ms	CA	0.599	0.504	0.619	0.543	0.635	0.578	0.649	0.555	0.635	0.527	0.627 $\pm$ 0.019	0.541 $\pm$ 0.028

Appendix

Table 12: Macro-averaged F1 for in-domain (ID) and cross-domain (CD) performance across five geographic data splits for representative benchmark configurations. Split 1* corresponds to the official EarthScape benchmark split, while Splits 2–5 were generated with different random seeds to quantify performance variability across geographic splits. Mean and standard deviation (SD) were computed across all five runs. Best and second-best scores within each run and summary column are indicated in bold and underlined , respectively.

Figure 10: Relationship between distribution shift (MMD) and cross-domain performance degradation ( $\Delta$ F1) across encoders and fusion strategies. Red lines show ordinary least-squares fits. Corresponding statistical analyses are reported in Table 13 .

Figure 11: Comparison of in-domain (ID) and cross-domain (CD) macro-F1 across stacking, concatenation, and cross-attention fusion strategies for ResNeXt and ViT encoders. Points show individual input configurations and boxplots summarize distributions.

Encoder	Fusion	$n$	$R^{2}$	PCC		SRCC		KRCC
Encoder	Fusion	$n$	$R^{2}$	$r$	$p_{\text{Holm}}$	$\rho$	$p_{\text{Holm}}$	$\tau$	$p_{\text{Holm}}$
ResNeXt	-	35	0.062	0.248	0.451	0.368	0.090	0.234	0.145
	Stacking	15	0.136	0.369	0.451	0.639	0.041	0.505	0.033
	Concat.	15	0.247	0.497	0.237	0.454	0.179	0.333	0.185
	CA	15	0.082	0.286	0.451	0.318	0.248	0.257	0.202
ViT	-	35	0.358	0.598	0.001	0.513	0.010	0.371	0.010

Appendix

Table 13: Statistical association between cross-domain F1 degradation,

\Delta=\mathrm{ID}-\mathrm{CD}

, and maximum mean discrepancy (MMD) for unimodal, and multiscale and multimodal experiments, stratified by encoder and fusion method. Linear association was summarized by coefficient of determination (

R^{2}

) and Pearson’s correlation coefficient (PCC,

r

), with monotonic association assessed using Spearman’s rank correlation coefficient (SRCC,

\rho

) and Kendall’s rank correlation coefficient (KRCC,

\tau

). All reported

p

-values were Holm-corrected for multiple comparisons.

Encoder	Domain	$A$	$\overline{\text{F1}}_{A}$	$B$	$\overline{\text{F1}}_{B}$	$n$	$t$	$p_{\mathrm{Holm}}$	$W$	$p_{\mathrm{Holm}}$	$d_{z}$
ResNeXt	ID	Stacking	0.653	Concat.	0.595	15	8.045	$<$ 0.001	0.000	$<$ 0.001	2.077
		Stacking	0.653	CA	0.605	15	10.209	$<$ 0.001	0.000	$<$ 0.001	2.636
		Concat.	0.595	CA	0.605	15	-1.019	0.326	56.000	0.847	-0.263
	CD	Stacking	0.559	Concat.	0.521	15	2.167	0.144	22.000	0.077	0.559
		Stacking	0.559	CA	0.544	15	0.939	0.364	41.000	0.303	0.242
		Concat.	0.521	CA	0.544	15	-1.871	0.165	21.000	0.077	-0.483

Appendix

Table 14: Paired statistical comparisons of stacking, concatenation, and cross-attention (CA) fusion across matched input configurations, stratified by encoder and evaluation domain (ID, in-domain; CD, cross-domain).

A

and

B

denote the paired fusion methods, with their corresponding mean macro-F1 (

\overline{\text{F1}}_{A}

and

\overline{\text{F1}}_{B}

) and number of matched pairs (

n

). Differences were assessed using paired

t

-tests (

t

) and Wilcoxon signed-rank tests (

W

), with paired-sample effect size reported as Cohen’s

d_{z}

. All

p

-values were Holm-corrected within each encoder–domain family.

Model	Input	$\alpha=0.001$		$\alpha=0.0001$		$\alpha=0.00001$
Model	Input	ID	CD	ID	CD	ID	CD
Copernicus-FM	RGB+NIR	0.558	0.316	0.592	0.341	0.591	0.357
DOFA	RGB+NIR	0.573	0.374	0.602	0.386	0.601	0.359
Galileo	RGB+NIR	0.545	0.260	0.548	0.286	0.542	0.249
Panopticon	RGB+NIR	0.587	0.408	0.632	0.507	0.617	0.461
TerraMind	RGB	0.563	0.394	0.590	0.414	0.582	0.376

Appendix

Table 15: Macro-averaged F1 for in-domain (ID) and cross-domain (CD) performance of Copernicus-FM, DOFA, Galileo, Panopticon, and TerraMind foundation models across three learning rates (

\alpha\in\{10^{-3},10^{-4},10^{-5}\}

) and supported input configurations.

Figure 12: Summary of in-domain (ID) and cross-domain (CD) performance across classification and semantic segmentation experiments. Classification results are shown for single-input, multi-scale, and multimodal configurations using ResNeXt and ViT, with segmentation results shown for DeepLabV3+ and SegFormer. The dashed black line indicates equivalent ID and CD performance, and shaded bands denote successive 5% intervals of relative performance degradation, $(\mathrm{ID}-\mathrm{CD})/\mathrm{ID}\times 100$ . Point colors distinguish input families.

Input	F1 (ResNeXt)				F1 (ViT)				mAP (ResNeXt)			mAP (ViT)
Input	ID	CD	$\Delta$	ID $+$ CD	ID	CD	$\Delta$	ID $+$ CD	ID	CD	$\Delta$	ID	CD	$\Delta$
DEM	0.687	0.508	0.179	1.195	0.674	0.517	0.157	1.191	0.736	0.565	0.171	0.720	0.579	0.141
RGB	0.608	0.504	0.104	1.112	0.599	0.473	0.127	1.072	0.653	0.483	0.170	0.638	0.483	0.155
NIR	0.611	0.520	0.091	1.131	0.606	0.481	0.125	1.087	0.647	0.534	0.113	0.616	0.464	0.151
NHD	0.520	0.437	0.084	0.957	0.512	0.438	0.075	0.950	0.481	0.375	0.106	0.465	0.377	0.088
OSM	0.540	0.466	0.074	1.006	0.539	0.458	0.081	0.997	0.495	0.411	0.084	0.488	0.401	0.087

Appendix

Table 16: Macro-averaged F1 and mAP for single input configurations with ResNeXt and ViT backbones. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. ID

+

CD was additionally reported for F1 as a measure of combined performance across domains. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Precision (ResNeXt)			Precision (ViT)			Recall (ResNeXt)			Recall (ViT)
Input	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$
DEM	0.666	0.474	0.192	0.639	0.439	0.200	0.719	0.585	0.134	0.740	0.649	0.090
RGB	0.569	0.436	0.134	0.578	0.423	0.155	0.654	0.615	0.040	0.668	0.649	0.019
NIR	0.566	0.492	0.074	0.549	0.467	0.082	0.683	0.586	0.097	0.713	0.510	0.203
NHD	0.418	0.347	0.072	0.419	0.362	0.057	0.741	0.708	0.033	0.767	0.725	0.042
OSM	0.454	0.378	0.076	0.449	0.373	0.076	0.837	0.829	0.008	0.841	0.820	0.021

Appendix

Table 17: Macro-averaged precision and recall for single input configurations with ResNeXt and ViT backbones. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	-	0.703	0.813	0.229	0.457	0.883	0.750	0.972	0.659	0.808	0.250	0.354	0.901	0.778	0.966
RGB	-	0.719	0.790	0.000	0.262	0.837	0.680	0.966	0.695	0.777	0.088	0.209	0.813	0.652	0.961
NIR	-	0.694	0.778	0.029	0.303	0.817	0.693	0.963	0.681	0.762	0.080	0.270	0.831	0.653	0.964
NHD	-	0.542	0.762	0.043	0.209	0.657	0.464	0.964	0.541	0.764	0.036	0.165	0.655	0.461	0.964
OSM	-	0.677	0.773	0.030	0.111	0.721	0.502	0.965	0.671	0.770	0.033	0.128	0.711	0.494	0.966

Appendix

Table 18: Class-wise F1 for single input configurations with ResNeXt and ViT backbones under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	-	0.558	0.683	0.000	0.023	0.688	0.611	0.993	0.523	0.694	0.000	0.023	0.761	0.623	0.992
RGB	-	0.629	0.702	0.000	0.237	0.507	0.459	0.991	0.467	0.669	0.000	0.131	0.637	0.451	0.954
NIR	-	0.583	0.601	0.007	0.408	0.578	0.474	0.988	0.624	0.680	0.000	0.260	0.360	0.448	0.992
NHD	-	0.440	0.709	0.000	0.142	0.471	0.304	0.991	0.436	0.710	0.008	0.155	0.459	0.306	0.991
OSM	-	0.650	0.709	0.009	0.106	0.495	0.302	0.991	0.624	0.708	0.009	0.103	0.474	0.296	0.992

Appendix

Table 19: Class-wise F1 for single input configurations with ResNeXt and ViT backbones under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	-	0.849	0.839	0.893	0.925	0.961	0.914	0.939	0.800	0.836	0.930	0.905	0.972	0.931	0.936
NIR	-	0.837	0.742	0.748	0.824	0.914	0.869	0.869	0.831	0.694	0.743	0.769	0.910	0.841	0.819
NHD	-	0.553	0.664	0.567	0.784	0.635	0.625	0.733	0.504	0.651	0.587	0.770	0.610	0.617	0.710
OSM	-	0.823	0.561	0.673	0.635	0.723	0.661	0.649	0.806	0.582	0.733	0.688	0.712	0.654	0.594
RGB	-	0.857	0.763	0.707	0.828	0.924	0.870	0.884	0.846	0.718	0.806	0.860	0.911	0.842	0.845

Appendix

Table 20: Class-wise AUC for single input configurations with ResNeXt and ViT backbones under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	-	0.774	0.660	0.629	0.722	0.953	0.893	0.636	0.724	0.644	0.703	0.497	0.954	0.924	0.814
NIR	-	0.779	0.554	0.573	0.828	0.777	0.786	0.808	0.822	0.591	0.565	0.678	0.596	0.792	0.723
NHD	-	0.524	0.633	0.693	0.725	0.487	0.518	0.543	0.500	0.641	0.592	0.701	0.494	0.518	0.538
OSM	-	0.839	0.510	0.467	0.568	0.656	0.587	0.668	0.824	0.495	0.551	0.558	0.615	0.597	0.648
RGB	-	0.843	0.596	0.461	0.724	0.691	0.784	0.770	0.739	0.571	0.311	0.624	0.921	0.730	0.687

Appendix

Table 21: Class-wise AUC for single input configurations with ResNeXt and ViT backbones under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	F1 (ResNeXt)				F1 (ViT)				mAP (ResNeXt)			mAP (ViT)
Input	Fusion	ID	CD	$\Delta$	ID $+$ CD	ID	CD	$\Delta$	ID $+$ CD	ID	CD	$\Delta$	ID	CD	$\Delta$
EP-ms	Stacking	0.670	0.394	0.276	1.064	0.656	0.558	0.098	1.214	0.710	0.465	0.245	0.703	0.575	0.128
	Concat.	0.605	0.523	0.082	1.128	0.660	0.546	0.114	1.206	0.610	0.511	0.099	0.702	0.579	0.123
	CA	0.612	0.524	0.088	1.136	0.632	0.542	0.090	1.174	0.655	0.537	0.118	0.674	0.560	0.114
PlC-ms	Stacking	0.620	0.542	0.078	1.162	0.622	0.508	0.114	1.130	0.664	0.57	0.094	0.661	0.531	0.130
	Concat.	0.541	0.455	0.086	0.996	0.625	0.532	0.093	1.157	0.539	0.435	0.104	0.678	0.549	0.129

Appendix

Table 22: Macro-averaged F1 and mAP for multi-scale input configurations with ResNeXt and ViT backbones. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. ID

+

CD was additionally reported for F1 as a measure of combined performance across domains. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	Precision (ResNeXt)			Precision (ViT)			Recall (ResNeXt)			Recall (ViT)
Input	Fusion	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$
EP-ms	Stacking	0.644	0.457	0.187	0.606	0.497	0.109	0.706	0.377	-0.329	0.650	0.734	-0.084
	Concat.	0.540	0.472	0.068	0.629	0.519	0.110	0.602	0.711	-0.109	0.586	0.713	-0.127
	CA	0.565	0.496	0.069	0.591	0.496	0.095	0.571	0.677	-0.106	0.607	0.699	-0.092
PlC-ms	Stacking	0.625	0.486	0.139	0.574	0.439	0.135	0.653	0.632	0.021	0.632	0.695	-0.063
	Concat.	0.461	0.370	0.091	0.584	0.465	0.119	0.650	0.680	-0.030	0.642	0.677	-0.035

Appendix

Table 23: Macro-averaged precision and recall for multi-scale input configurations with ResNeXt and ViT backbones. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
EP-ms	Stacking	0.724	0.810	0.150	0.420	0.869	0.751	0.966	0.696	0.807	0.122	0.380	0.859	0.759	0.967
	Concat.	0.707	0.783	0.078	0.265	0.761	0.677	0.966	0.740	0.818	0.151	0.351	0.851	0.745	0.968
	CA	0.700	0.779	0.133	0.243	0.759	0.703	0.965	0.700	0.805	0.160	0.218	0.847	0.729	0.965
PlC-ms	Stacking	0.666	0.799	0.071	0.276	0.831	0.732	0.966	0.643	0.799	0.107	0.295	0.824	0.723	0.965
	Concat.	0.507	0.771	0.000	0.205	0.711	0.626	0.966	0.631	0.806	0.114	0.301	0.836	0.726	0.958

Appendix

Table 24: Class-wise F1 for multi-scale input configurations with ResNeXt and ViT backbones under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
EP-ms	Stacking	0.623	0.692	0.000	0.301	0.114	0.033	0.991	0.647	0.702	0.000	0.286	0.644	0.635	0.991
	Concat.	0.672	0.739	0.000	0.343	0.563	0.348	0.992	0.706	0.709	0.000	0.159	0.618	0.640	0.992
	CA	0.699	0.758	0.000	0.360	0.327	0.536	0.991	0.665	0.703	0.000	0.120	0.689	0.625	0.991
PlC-ms	Stacking	0.626	0.710	0.000	0.209	0.674	0.584	0.991	0.562	0.702	0.000	0.115	0.624	0.561	0.989
	Concat.	0.433	0.707	0.000	0.057	0.531	0.466	0.992	0.500	0.685	0.000	0.314	0.636	0.598	0.991

Appendix

Table 25: Class-wise F1 for multi-scale input configurations with ResNeXt and ViT backbones under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
EP-ms	Stacking	0.865	0.840	0.772	0.894	0.951	0.917	0.923	0.841	0.842	0.834	0.916	0.944	0.925	0.915
	Concat.	0.859	0.754	0.717	0.755	0.838	0.856	0.778	0.869	0.852	0.854	0.899	0.942	0.922	0.919
	CA	0.843	0.782	0.872	0.846	0.837	0.894	0.853	0.830	0.826	0.881	0.857	0.931	0.914	0.882
PlC-ms	Stacking	0.806	0.810	0.691	0.880	0.914	0.896	0.902	0.797	0.804	0.825	0.902	0.915	0.907	0.900
	Concat.	0.545	0.669	0.722	0.797	0.799	0.818	0.780	0.775	0.816	0.866	0.884	0.926	0.912	0.900

Appendix

Table 26: Class-wise AUC for multi-scale input configurations with ResNeXt and ViT backbones under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
EP-ms	Stacking	0.819	0.741	0.503	0.773	0.504	0.606	0.613	0.852	0.693	0.720	0.745	0.845	0.918	0.821
	Concat.	0.859	0.736	0.586	0.763	0.735	0.816	0.832	0.874	0.722	0.709	0.719	0.826	0.920	0.859
	CA	0.860	0.805	0.775	0.777	0.646	0.868	0.816	0.853	0.631	0.651	0.570	0.861	0.917	0.884
PlC-ms	Stacking	0.813	0.711	0.709	0.694	0.878	0.901	0.829	0.768	0.653	0.783	0.608	0.844	0.889	0.832
	Concat.	0.521	0.662	0.760	0.627	0.762	0.802	0.691	0.693	0.655	0.831	0.773	0.864	0.897	0.814

Appendix

Table 27: Class-wise AUC for multi-scale input configurations with ResNeXt and ViT backbones under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	F1 (ResNeXt)				F1 (ViT)				mAP (ResNeXt)			mAP (ViT)
Input	Fusion	ID	CD	$\Delta$	ID $+$ CD	ID	CD	$\Delta$	ID $+$ CD	ID	CD	$\Delta$	ID	CD	$\Delta$
DEM+RGB	Stacking	0.642	0.519	0.123	1.160	0.662	0.507	0.155	1.169	0.666	0.516	0.150	0.686	0.539	0.147
	Concat.	0.595	0.525	0.070	1.119	0.651	0.446	0.205	1.097	0.602	0.534	0.068	0.707	0.513	0.194
	CA	0.586	0.484	0.102	1.071	0.669	0.425	0.244	1.094	0.621	0.545	0.076	0.711	0.505	0.205
DEM+NHD+ OSM+RGB	Stacking	0.649	0.484	0.166	1.133	0.632	0.465	0.167	1.096	0.682	0.501	0.181	0.688	0.514	0.174
	Concat.	0.583	0.466	0.116	1.049	0.652	0.513	0.139	1.165	0.586	0.543	0.044	0.698	0.546	0.152

Appendix

Table 28: Macro-averaged F1 and mAP for multimodal input configurations with ResNeXt and ViT backbones. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. ID

+

CD was additionally reported for F1 as a measure of combined performance across domains. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	Precision (ResNeXt)			Precision (ViT)			Recall (ResNeXt)			Recall (ViT)
Input	Fusion	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$
DEM+RGB	Stacking	0.598	0.452	0.146	0.606	0.476	0.129	0.772	0.628	0.143	0.761	0.548	0.213
	Concat.	0.562	0.528	0.034	0.626	0.504	0.122	0.634	0.537	0.097	0.684	0.446	0.237
	CA	0.546	0.488	0.058	0.628	0.441	0.187	0.640	0.623	0.017	0.738	0.503	0.235
DEM+NHD+OSM+RGB	Stacking	0.613	0.448	0.165	0.614	0.439	0.175	0.718	0.537	0.181	0.654	0.540	0.115
	Concat.	0.546	0.396	0.150	0.626	0.448	0.178	0.640	0.693	-0.053	0.699	0.618	0.080

Appendix

Table 29: Macro-averaged precision and recall for multimodal input configurations with ResNeXt and ViT backbones. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM+RGB	Stacking	0.714	0.779	0.087	0.413	0.823	0.707	0.968	0.708	0.795	0.242	0.310	0.858	0.753	0.966
	Concat.	0.668	0.771	0.000	0.276	0.797	0.685	0.964	0.712	0.801	0.000	0.403	0.884	0.787	0.971
	CA	0.662	0.778	0.000	0.219	0.819	0.662	0.965	0.714	0.806	0.136	0.434	0.881	0.746	0.969
DEM+NHD+OSM+RGB	Stacking	0.715	0.771	0.163	0.351	0.859	0.719	0.967	0.675	0.798	0.108	0.240	0.868	0.766	0.965
	Concat.	0.677	0.771	0.089	0.200	0.797	0.579	0.966	0.723	0.797	0.140	0.288	0.877	0.769	0.971

Appendix

Table 30: Class-wise F1 for multimodal input configurations with ResNeXt and ViT backbones under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM+RGB	Stacking	0.584	0.692	0.000	0.181	0.652	0.531	0.992	0.514	0.598	0.000	0.000	0.771	0.672	0.992
	Concat.	0.649	0.696	0.000	0.287	0.585	0.464	0.991	0.418	0.323	0.000	0.024	0.705	0.659	0.992
	CA	0.517	0.606	0.000	0.327	0.530	0.417	0.992	0.458	0.169	0.000	0.000	0.790	0.565	0.992
DEM+NHD+OSM+RGB	Stacking	0.584	0.685	0.000	0.027	0.641	0.460	0.988	0.371	0.646	0.000	0.000	0.622	0.621	0.992
	Concat.	0.547	0.710	0.000	0.176	0.479	0.361	0.991	0.601	0.628	0.000	0.030	0.711	0.629	0.992

Appendix

Table 31: Class-wise F1 for multimodal input configurations with ResNeXt and ViT backbones under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM+RGB	Stacking	0.850	0.752	0.861	0.888	0.913	0.882	0.918	0.849	0.786	0.907	0.896	0.939	0.920	0.931
	Concat.	0.785	0.702	0.533	0.704	0.881	0.865	0.726	0.852	0.832	0.868	0.919	0.963	0.939	0.942
	CA	0.803	0.700	0.776	0.861	0.909	0.835	0.867	0.854	0.822	0.925	0.919	0.966	0.919	0.939
DEM+NHD+OSM+RGB	Stacking	0.855	0.759	0.832	0.886	0.934	0.895	0.924	0.828	0.800	0.933	0.891	0.952	0.921	0.931
	Concat.	0.813	0.644	0.737	0.781	0.887	0.802	0.818	0.864	0.820	0.866	0.909	0.960	0.929	0.947

Appendix

Table 32: Class-wise AUC for multimodal input configurations with ResNeXt and ViT backbones under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Fusion	ResNeXt							ViT
Input	Fusion	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM+RGB	Stacking	0.803	0.580	0.418	0.615	0.874	0.850	0.720	0.742	0.546	0.400	0.357	0.953	0.922	0.722
	Concat.	0.828	0.697	0.248	0.728	0.847	0.832	0.814	0.659	0.557	0.524	0.425	0.889	0.922	0.746
	CA	0.733	0.538	0.560	0.781	0.888	0.857	0.692	0.634	0.534	0.786	0.436	0.941	0.886	0.792
DEM+NHD+OSM+RGB	Stacking	0.772	0.603	0.712	0.541	0.857	0.775	0.617	0.624	0.486	0.334	0.549	0.905	0.900	0.818
	Concat.	0.776	0.633	0.779	0.752	0.866	0.897	0.653	0.803	0.599	0.502	0.492	0.916	0.903	0.692

Appendix

Table 33: Class-wise AUC for multimodal input configurations with ResNeXt and ViT backbones under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	mIoU (DeepLabV3+)			mIoU (Segformer)			Dice (DeepLabV3+)			Dice (Segformer)
Input	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$
DEM	0.486	0.356	0.130	0.512	0.412	0.100	0.548	0.415	0.133	0.576	0.474	0.101
EP-5	0.448	0.378	0.070	0.460	0.400	0.059	0.513	0.423	0.089	0.525	0.447	0.079
EP-51	0.453	0.378	0.075	0.485	0.403	0.083	0.521	0.426	0.095	0.552	0.451	0.102
S-5	0.461	0.384	0.078	0.484	0.430	0.054	0.522	0.442	0.080	0.546	0.491	0.055
S-50	0.424	0.394	0.031	0.432	0.407	0.025	0.482	0.442	0.040	0.488	0.455	0.034

Appendix

Table 34: Mean IoU (mIoU) and macro-averaged Dice scores for semantic segmentation with DeepLabV3+ and Segformer architectures. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	Precision (DeepLabV3+)			Precision (Segformer)			Recall (DeepLabV3+)			Recall (Segformer)
Input	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$	ID	CD	$\Delta$
DEM	0.616	0.498	0.118	0.632	0.535	0.096	0.551	0.441	0.110	0.585	0.496	0.089
EP-5	0.582	0.486	0.096	0.594	0.494	0.099	0.528	0.461	0.066	0.536	0.484	0.052
EP-51	0.581	0.471	0.111	0.616	0.497	0.119	0.533	0.465	0.068	0.561	0.490	0.071
S-5	0.592	0.508	0.084	0.607	0.541	0.066	0.529	0.454	0.075	0.555	0.511	0.044
S-50	0.532	0.490	0.042	0.530	0.500	0.030	0.505	0.461	0.044	0.512	0.474	0.037

Appendix

Table 35: Macro-averaged precision and recall for semantic segmentation with DeepLabV3+ and Segformer architectures. Performance was evaluated in-domain (ID) and cross-domain (CD), with cross-domain degradation computed as

\Delta=\mathrm{ID}-\mathrm{CD}

. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	DeepLabV3+							Segformer
Input	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	0.240	0.343	0.000	0.057	0.563	0.316	0.750	0.296	0.407	0.000	0.076	0.590	0.339	0.766
EP-5	0.239	0.328	0.000	0.017	0.469	0.209	0.720	0.266	0.356	0.000	0.055	0.494	0.277	0.727
EP-51	0.262	0.367	0.000	0.078	0.456	0.257	0.728	0.286	0.384	0.000	0.058	0.514	0.280	0.750
S-5	0.227	0.303	0.000	0.048	0.580	0.296	0.732	0.256	0.314	0.000	0.055	0.606	0.319	0.761
S-50	0.045	0.243	0.000	0.043	0.547	0.299	0.711	0.037	0.249	0.000	0.052	0.555	0.323	0.717

Appendix

Table 36: Class-wise IoU for semantic segmentation with DeepLabV3+ and Segformer architectures under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	DeepLabV3+							Segformer
Input	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	0.126	0.189	0.000	0.000	0.264	0.099	0.714	0.201	0.286	0.000	0.000	0.338	0.123	0.755
EP-5	0.105	0.260	0.000	0.027	0.023	0.007	0.774	0.147	0.306	0.000	0.078	0.022	0.001	0.783
EP-51	0.179	0.283	0.000	0.037	0.017	0.008	0.763	0.174	0.307	0.000	0.066	0.016	0.017	0.780
S-5	0.138	0.231	0.000	0.044	0.365	0.126	0.749	0.173	0.285	0.000	0.067	0.405	0.167	0.783
S-50	0.006	0.174	0.000	0.013	0.376	0.138	0.766	0.006	0.185	0.000	0.013	0.374	0.162	0.776

Appendix

Table 37: Class-wise IoU for semantic segmentation with DeepLabV3+ and Segformer architectures under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	DeepLabV3+							Segformer
Input	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	0.295	0.415	0.000	0.079	0.630	0.392	0.802	0.357	0.484	0.000	0.103	0.657	0.414	0.819
EP-5	0.294	0.399	0.000	0.026	0.544	0.272	0.782	0.329	0.428	0.000	0.074	0.568	0.349	0.786
EP-51	0.324	0.442	0.000	0.106	0.532	0.331	0.790	0.349	0.459	0.000	0.083	0.588	0.356	0.808
S-5	0.279	0.370	0.000	0.071	0.646	0.366	0.788	0.314	0.383	0.000	0.075	0.675	0.392	0.814
S-50	0.058	0.301	0.000	0.061	0.628	0.373	0.771	0.049	0.303	0.000	0.070	0.633	0.399	0.776

Appendix

Table 38: Class-wise Dice scores for semantic segmentation with DeepLabV3+ and Segformer architectures under in-domain (ID) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Input	DeepLabV3+							Segformer
Input	af1	Qal	Qaf	Qat	Qc	Qca	Qr	af1	Qal	Qaf	Qat	Qc	Qca	Qr
DEM	0.174	0.248	0.000	0.000	0.332	0.143	0.779	0.265	0.352	0.000	0.000	0.408	0.169	0.819
EP-5	0.148	0.322	0.000	0.041	0.034	0.010	0.831	0.198	0.371	0.000	0.109	0.031	0.002	0.837
EP-51	0.243	0.347	0.000	0.055	0.025	0.013	0.821	0.235	0.372	0.000	0.089	0.023	0.025	0.835
S-5	0.192	0.287	0.000	0.066	0.435	0.168	0.813	0.235	0.348	0.000	0.088	0.480	0.225	0.841
S-50	0.010	0.214	0.000	0.022	0.460	0.188	0.824	0.010	0.224	0.000	0.020	0.457	0.217	0.833

Appendix

Table 39: Class-wise Dice scores for semantic segmentation with DeepLabV3+ and Segformer architectures under cross-domain (CD) evaluation. Best and second-best scores in each column are indicated in bold and underlined , respectively.

Remote Sensing Benchmarks

Remote sensing benchmarks measure how well models interpret Earth-observation data across sensors, regions, time, tasks, and operating conditions.

See all 195 papers

MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training
Feb 6, 2026Lucia Gordon, Serge Belongie, Christian Igel +1Test-Time Adaptation Multimodal Robustness
Harvard University, USA · University of Copenhagen, Denmark
TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation
Mar 19, 2026Yan Shu, Bin Ren, Zhitong Xiong +4Remote Sensing VLMs Geospatial Reasoning
University of Trento · MBZUAI · Technical University of Munich +1
GeoHeight-Bench: Towards Height-Aware Multimodal Reasoning in Remote Sensing
Mar 26, 2026Xuran Hu, Zhitong Xiong, Zhongcheng Hong +3Remote Sensing Image Understanding Large Vision-Language Models
KTH Royal Institute of Technology · Technical University of Munich · The Hong Kong University of Science and Technology (Guangzhou)

Feb 6, 2026J/K move · Enter open · S save

MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training

Lucia Gordon, Serge Belongie, Christian Igel, Nico Lang

Harvard University, USA · University of Copenhagen, Denmark

Test-Time Adaptation Multimodal Robustness Geospatial Foundation Models Remote Sensing Test-Time Training Remote Sensing Benchmarks+1

Open paper AskPDF arXiv

Recent research in geospatial machine learning demonstrates that models pretrained with self-supervised learning on Earth observation data can perform well on downstream tasks with limited labeled data. However, most benchmark datasets have few data modalities and poor global representation, limiting the ability to evaluate multimodal pretrained models at global scales. In order to fill this gap, we introduce MMEarth-Bench, a collection of five new environmental tasks with 12 modalities, globally distributed data, and both random and geographic test splits. We benchmark a diverse set of pretrained models and find that while (multimodal) pretraining tends to improve model robustness in limited data settings, geographic generalization abilities remain poor. Moreover, a simple randomly initialized multimodal model is competitive given enough labeled data. Although data is abundant, models can currently only make use of the modalities on which they were pretrained. To solve this problem, we propose using all the modalities available at test time as auxiliary tasks for test-time adaptation. Our model-agnostic method for test-time training with multimodal reconstruction (TTT-MMR) can improve performance across all models and tasks on both test splits. Furthermore, geographic batching leads to a good trade-off between regularization and specialization during TTT, which is especially beneficial for long-tail distributions. Our dataset, code, and visualization tool are linked on the project page: lgordon99.github.io/mmearth-bench.

Abstract

Remote Sensing Benchmarks

Explore similar work

MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training

TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation

GeoHeight-Bench: Towards Height-Aware Multimodal Reasoning in Remote Sensing