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

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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 256×256256\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 ( ncn_{c} ) with percentages indicating the relative proportion of patches containing each class. Bottom left: sample frequency ( nn ) of label cardinality ( ∣Yi∣|Y_{i}| ). Right: spatial distribution of ∣Yi∣|Y_{i}| illustrating geospatial variation in patch complexity.
Classncn_{c}Mean AreaMARENSIRLbl
Qr292710.6510.7029464.61.0
Qal188010.0890.0588474.41.6
Qc137680.1420.1487476.32.1
af1109100.0510.0356641.42.7
Qca76690.0610.0545355.73.8
Qat14350.0060.0041336.920.4
Table 1: Label statistics and imbalance metrics for EarthScape, including global frequency ( ncn_{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.
FusionResNeXt ViT
InputF1ˉID±SD\bar{\text{F1}}_{\text{ID}}\pm\text{SD}F1ˉCD±SD\bar{\text{F1}}_{\text{CD}}\pm\text{SD}Δ\DeltaΔ%\Delta\% InputF1ˉID±SD\bar{\text{F1}}_{\text{ID}}\pm\text{SD}F1ˉCD±SD\bar{\text{F1}}_{\text{CD}}\pm\text{SD}Δ\DeltaΔ%\Delta\%
-DEM0.682 ±\pm 0.0030.526 ±\pm 0.0260.15622.9 DEM0.655 ±\pm 0.0150.524 ±\pm 0.0070.13120.0
-S-100.653 ±\pm 0.0150.587 ±\pm 0.0140.06610.1 S-50.642 ±\pm 0.0100.587 ±\pm 0.0150.0558.6
StackingPrC-ms0.654 ±\pm 0.0140.593 ±\pm 0.0180.0619.3 S-ms0.650 ±\pm 0.0080.583 ±\pm 0.0210.06710.3
Concat.PrC-ms0.606 ±\pm 0.0110.509 ±\pm 0.0270.09716.0 S-ms0.659 ±\pm 0.0100.569 ±\pm 0.0160.09013.7
CAPrC-ms0.612 ±\pm 0.0200.536 ±\pm 0.0240.07612.4 S-ms0.636 ±\pm 0.0100.577 ±\pm 0.0160.0599.3
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; Δ=F1ˉID−F1ˉCD\Delta=\bar{\mathrm{F1}}_{\mathrm{ID}}-\bar{\mathrm{F1}}_{\mathrm{CD}} and Δ%=100(F1ˉID−F1ˉCD)/F1ˉID\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.
EncoderFusionnnPCC SRCC
rrpHolmp_{\text{Holm}} ρ\rhopHolmp_{\text{Holm}}
ResNeXt-350.2480.451 0.3680.090
Stacking150.3690.451 0.6390.041
Concat.150.4970.237 0.4540.179
CA150.2860.451 0.3180.248
ViT-350.5980.001 0.5130.010
Table 3: Statistical association between cross-domain F1 degradation ( Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} ) and maximum mean discrepancy (MMD) for unimodal, and multiscale ++ multimodal experiments stratified by encoder and fusion method. PCC ( rr ): Pearson’s correlation coefficient; SRCC ( ρ\rho ): Spearman’s rank correlation coefficient. All pp -values were Holm-corrected for multiple comparisons. Significant correlations marked in bold .
EncoderDomainComparisonnnttpHolmp_{\mathrm{Holm}}
ResNeXtIDStacking–Concat.158.045<< 0.001
Stacking–CA1510.209<< 0.001
Concat.–CA15-1.0190.326
CDStacking–Concat.152.1670.144
Stacking–CA150.9390.364
Concat.–CA15-1.8710.165
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 pp -values were Holm-corrected within each encoder–domain family. Significant comparisons marked in bold .
ModelInputF1ID{\text{F1}}_{\text{ID}}F1CD{\text{F1}}_{\text{CD}}Δ\DeltaΔ%\Delta\%
Copernicus-FMRGB+NIR0.5920.3410.25142.4
DOFARGB+NIR0.6020.3860.21635.9
PanopticonRGB+NIR0.6320.5070.12519.8
GalileoRGB+NIR0.5480.2860.26247.8
DEM+S-50.5930.5080.08514.3
RGB+NIR+DEM+S-50.5480.2720.27650.4
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 ( Δ\Delta ) and relative ( Δ%\Delta\% ) cross-domain degradation.
Appendix figures & tables43 assets

Supplementary material from the paper’s appendix.

Appendix

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 256×256256\times 256 pixels ( 390.14×390.14390.14\times 390.14 m) with 50% overlap. Red squares in the upper left corners highlight a single patch.
Geologic ClassNotationEncodingGeomorphic ProcessCharacteristic Topographic Expression and Visual Cues
Artificial fillaf10AnthropogenicRectilinear shapes with sharp edges; surface patterns inconsistent with natural terrain.
AlluviumQal1Fluvial (active)Relatively broad, low-relief valleys, channels, and basins.
Alluvial fansQaf2Debris FlowSmall, lobate deposits at tributary mouths; wedge-shaped morphology; often isolated.
Terrace depositsQat3Fluvial (relict)Flat benches elevated above active floodplains; pronounced scarps; isolated or paired.
ColluviumQc4Gravity-driven (active)Steep slopes ( ≥12∘\geq 12^{\circ} ); irregular, rough terrain.
Colluvial apronsQca5Gravity-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 FeatureManuscript NotationComputational InputSpatial Support
Source DEM GSDKernel Size
Elevation PercentileEP-55 ft ( ≈\approx 1.52 m)5×55\times 5625 ft 2 ( ≈\approx 58 m 2 )
EP-115 ft ( ≈\approx 1.52 m)11×1111\times 113025 ft 2 ( ≈\approx 281 m 2 )
EP-215 ft ( ≈\approx 1.52 m)21×2121\times 2111025 ft 2 ( ≈\approx 1024 m 2 )
EP-515 ft ( ≈\approx 1.52 m)51×5151\times 5165025 ft 2 ( ≈\approx 6041 m 2 )
EP-1015 ft ( ≈\approx 1.52 m)101×101101\times 101255025 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.
InputMMDSD95% CI
DEM0.2730.0030.001
RGB0.1360.0060.002
NIR0.2870.0020.001
NHD0.0360.0060.002
OSM0.0280.0040.001
EP-50.0560.0060.002
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 ( ncn_{c} ) and intra-patch spatial support ( aca_{c} ) distributions for the Warren (in-domain, ID) and Hardin (cross-domain, CD) regions. Geographic locations are shown in Fig. 1 .
RegionDomainCandidateSplitSplitID Split
PatchesCountProportion
WarrenID23566Training841678.5%
Validation7687.2%
Testing153614.3%
HardinCD7452Testing1536-
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.
wc+w_{c}^{+}γ\gammaα\alphaInputIDCDΔ\DeltaΔ%\Delta\%ID ++ CD
10-DEM0.6830.4930.19127.91.176
S-50.6390.6010.0385.91.240
S-ms0.6440.5800.0649.91.225
DEM+S-ms0.6550.5850.07110.81.240
        1.220
ICF0-DEM0.6580.4890.16925.71.147
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 ( wc+w_{c}^{+} ), focal focusing parameter ( γ\gamma ), and focal balancing parameter ( α\alpha ); binary cross-entropy corresponds to wc+=1w_{c}^{+}=1 and γ=0\gamma=0 . Transfer degradation was computed as Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} and reported both in raw F1 scores and percentages ( Δ%\Delta\% ). Italicized values indicate mean ID ++ CD across input configurations; the mean ID ++ CD in bold denotes the selected loss setting.
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.
EncoderInputFusionRun 1* Run 2 Run 3 Run 4 Run 5 Mean ±\pm SD
IDCD IDCD IDCD IDCD IDCD IDCD
ResNeXtDEM-0.6870.508 0.6800.543 0.6800.530 0.6810.555 0.6800.492 0.682 ±\pm 0.0030.526 ±\pm 0.026
S-10-0.6740.599 0.6390.577 0.6510.569 0.6630.592 0.6390.600 0.653 ±\pm 0.0150.587 ±\pm 0.014
PrC-msStacking0.6640.614 0.6310.582 0.6520.604 0.6580.594 0.6640.569 0.654 ±\pm 0.0140.593 ±\pm 0.018
PrC-msConcat.0.5980.500 0.6140.504 0.5920.473 0.6090.520 0.6190.547 0.606 ±\pm 0.0110.509 ±\pm 0.027
PrC-msCA0.5990.504 0.6450.542 0.6140.570 0.5920.525 0.6080.539 0.612 ±\pm 0.0200.536 ±\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.
EncoderInputFusionSplit 1* Split 2 Split 3 Split 4 Split 5 Mean ±\pm SD
IDCD IDCD IDCD IDCD IDCD IDCD
ResNeXtDEM-0.6870.508 0.6690.535 0.6640.506 0.6930.562 0.7040.521 0.683 ±\pm 0.0170.526 ±\pm 0.023
S-10-0.6740.599 0.6560.585 0.6600.590 0.6680.615 0.6570.599 0.663 ±\pm 0.0080.598 ±\pm 0.011
PrC-msStacking0.6640.614 0.6650.611 0.6630.578 0.6700.607 0.6430.597 0.661 ±\pm 0.0100.601 ±\pm 0.015
PrC-msConcat.0.5980.500 0.6170.515 0.6140.506 0.5760.469 0.6160.503 0.604 ±\pm 0.0180.499 ±\pm 0.017
PrC-msCA0.5990.504 0.6190.543 0.6350.578 0.6490.555 0.6350.527 0.627 ±\pm 0.0190.541 ±\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 ( Δ\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.
EncoderFusionnnR2R^{2}PCC SRCC KRCC
rrpHolmp_{\text{Holm}} ρ\rhopHolmp_{\text{Holm}} τ\taupHolmp_{\text{Holm}}
ResNeXt-350.0620.2480.451 0.3680.090 0.2340.145
Stacking150.1360.3690.451 0.6390.041 0.5050.033
Concat.150.2470.4970.237 0.4540.179 0.3330.185
CA150.0820.2860.451 0.3180.248 0.2570.202
ViT-350.3580.5980.001 0.5130.010 0.3710.010
Appendix
Table 13: Statistical association between cross-domain F1 degradation, Δ=ID−CD\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 ( R2R^{2} ) and Pearson’s correlation coefficient (PCC, rr ), with monotonic association assessed using Spearman’s rank correlation coefficient (SRCC, ρ\rho ) and Kendall’s rank correlation coefficient (KRCC, τ\tau ). All reported pp -values were Holm-corrected for multiple comparisons.
EncoderDomainAAF1‾A\overline{\text{F1}}_{A}BBF1‾B\overline{\text{F1}}_{B}nnttpHolmp_{\mathrm{Holm}}WWpHolmp_{\mathrm{Holm}}dzd_{z}
ResNeXtIDStacking0.653Concat.0.595158.045<< 0.0010.000<< 0.0012.077
Stacking0.653CA0.6051510.209<< 0.0010.000<< 0.0012.636
Concat.0.595CA0.60515-1.0190.32656.0000.847-0.263
CDStacking0.559Concat.0.521152.1670.14422.0000.0770.559
Stacking0.559CA0.544150.9390.36441.0000.3030.242
Concat.0.521CA0.54415-1.8710.16521.0000.077-0.483
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). AA and BB denote the paired fusion methods, with their corresponding mean macro-F1 ( F1‾A\overline{\text{F1}}_{A} and F1‾B\overline{\text{F1}}_{B} ) and number of matched pairs ( nn ). Differences were assessed using paired tt -tests ( tt ) and Wilcoxon signed-rank tests ( WW ), with paired-sample effect size reported as Cohen’s dzd_{z} . All pp -values were Holm-corrected within each encoder–domain family.
ModelInputα=0.001\alpha=0.001 α=0.0001\alpha=0.0001 α=0.00001\alpha=0.00001
IDCD IDCD IDCD
Copernicus-FMRGB+NIR0.5580.316 0.5920.341 0.5910.357
DOFARGB+NIR0.5730.374 0.6020.386 0.6010.359
GalileoRGB+NIR0.5450.260 0.5480.286 0.5420.249
PanopticonRGB+NIR0.5870.408 0.6320.507 0.6170.461
TerraMindRGB0.5630.394 0.5900.414 0.5820.376
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 ( α∈{10−3,10−4,10−5}\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, (ID−CD)/ID×100(\mathrm{ID}-\mathrm{CD})/\mathrm{ID}\times 100 . Point colors distinguish input families.
InputF1 (ResNeXt) F1 (ViT) mAP (ResNeXt) mAP (ViT)
IDCDΔ\DeltaID ++ CD IDCDΔ\DeltaID ++ CD IDCDΔ\Delta IDCDΔ\Delta
DEM0.6870.5080.1791.195 0.6740.5170.1571.191 0.7360.5650.171 0.7200.5790.141
RGB0.6080.5040.1041.112 0.5990.4730.1271.072 0.6530.4830.170 0.6380.4830.155
NIR0.6110.5200.0911.131 0.6060.4810.1251.087 0.6470.5340.113 0.6160.4640.151
NHD0.5200.4370.0840.957 0.5120.4380.0750.950 0.4810.3750.106 0.4650.3770.088
OSM0.5400.4660.0741.006 0.5390.4580.0810.997 0.4950.4110.084 0.4880.4010.087
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 Δ=ID−CD\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.
InputPrecision (ResNeXt) Precision (ViT) Recall (ResNeXt) Recall (ViT)
IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta
DEM0.6660.4740.192 0.6390.4390.200 0.7190.5850.134 0.7400.6490.090
RGB0.5690.4360.134 0.5780.4230.155 0.6540.6150.040 0.6680.6490.019
NIR0.5660.4920.074 0.5490.4670.082 0.6830.5860.097 0.7130.5100.203
NHD0.4180.3470.072 0.4190.3620.057 0.7410.7080.033 0.7670.7250.042
OSM0.4540.3780.076 0.4490.3730.076 0.8370.8290.008 0.8410.8200.021
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 Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} . Best and second-best scores in each column are indicated in bold and underlined , respectively.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM-0.7030.8130.2290.4570.8830.7500.972 0.6590.8080.2500.3540.9010.7780.966
RGB-0.7190.7900.0000.2620.8370.6800.966 0.6950.7770.0880.2090.8130.6520.961
NIR-0.6940.7780.0290.3030.8170.6930.963 0.6810.7620.0800.2700.8310.6530.964
NHD-0.5420.7620.0430.2090.6570.4640.964 0.5410.7640.0360.1650.6550.4610.964
OSM-0.6770.7730.0300.1110.7210.5020.965 0.6710.7700.0330.1280.7110.4940.966
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM-0.5580.6830.0000.0230.6880.6110.993 0.5230.6940.0000.0230.7610.6230.992
RGB-0.6290.7020.0000.2370.5070.4590.991 0.4670.6690.0000.1310.6370.4510.954
NIR-0.5830.6010.0070.4080.5780.4740.988 0.6240.6800.0000.2600.3600.4480.992
NHD-0.4400.7090.0000.1420.4710.3040.991 0.4360.7100.0080.1550.4590.3060.991
OSM-0.6500.7090.0090.1060.4950.3020.991 0.6240.7080.0090.1030.4740.2960.992
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM-0.8490.8390.8930.9250.9610.9140.939 0.8000.8360.9300.9050.9720.9310.936
NIR-0.8370.7420.7480.8240.9140.8690.869 0.8310.6940.7430.7690.9100.8410.819
NHD-0.5530.6640.5670.7840.6350.6250.733 0.5040.6510.5870.7700.6100.6170.710
OSM-0.8230.5610.6730.6350.7230.6610.649 0.8060.5820.7330.6880.7120.6540.594
RGB-0.8570.7630.7070.8280.9240.8700.884 0.8460.7180.8060.8600.9110.8420.845
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM-0.7740.6600.6290.7220.9530.8930.636 0.7240.6440.7030.4970.9540.9240.814
NIR-0.7790.5540.5730.8280.7770.7860.808 0.8220.5910.5650.6780.5960.7920.723
NHD-0.5240.6330.6930.7250.4870.5180.543 0.5000.6410.5920.7010.4940.5180.538
OSM-0.8390.5100.4670.5680.6560.5870.668 0.8240.4950.5510.5580.6150.5970.648
RGB-0.8430.5960.4610.7240.6910.7840.770 0.7390.5710.3110.6240.9210.7300.687
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.
InputFusionF1 (ResNeXt) F1 (ViT) mAP (ResNeXt) mAP (ViT)
IDCDΔ\DeltaID ++ CD IDCDΔ\DeltaID ++ CD IDCDΔ\Delta IDCDΔ\Delta
EP-msStacking0.6700.3940.2761.064 0.6560.5580.0981.214 0.7100.4650.245 0.7030.5750.128
Concat.0.6050.5230.0821.128 0.6600.5460.1141.206 0.6100.5110.099 0.7020.5790.123
CA0.6120.5240.0881.136 0.6320.5420.0901.174 0.6550.5370.118 0.6740.5600.114
PlC-msStacking0.6200.5420.0781.162 0.6220.5080.1141.130 0.6640.570.094 0.6610.5310.130
Concat.0.5410.4550.0860.996 0.6250.5320.0931.157 0.5390.4350.104 0.6780.5490.129
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 Δ=ID−CD\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.
InputFusionPrecision (ResNeXt) Precision (ViT) Recall (ResNeXt) Recall (ViT)
IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta
EP-msStacking0.6440.4570.187 0.6060.4970.109 0.7060.377-0.329 0.6500.734-0.084
Concat.0.5400.4720.068 0.6290.5190.110 0.6020.711-0.109 0.5860.713-0.127
CA0.5650.4960.069 0.5910.4960.095 0.5710.677-0.106 0.6070.699-0.092
PlC-msStacking0.6250.4860.139 0.5740.4390.135 0.6530.6320.021 0.6320.695-0.063
Concat.0.4610.3700.091 0.5840.4650.119 0.6500.680-0.030 0.6420.677-0.035
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 Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} . Best and second-best scores in each column are indicated in bold and underlined , respectively.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
EP-msStacking0.7240.8100.1500.4200.8690.7510.966 0.6960.8070.1220.3800.8590.7590.967
Concat.0.7070.7830.0780.2650.7610.6770.966 0.7400.8180.1510.3510.8510.7450.968
CA0.7000.7790.1330.2430.7590.7030.965 0.7000.8050.1600.2180.8470.7290.965
PlC-msStacking0.6660.7990.0710.2760.8310.7320.966 0.6430.7990.1070.2950.8240.7230.965
Concat.0.5070.7710.0000.2050.7110.6260.966 0.6310.8060.1140.3010.8360.7260.958
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
EP-msStacking0.6230.6920.0000.3010.1140.0330.991 0.6470.7020.0000.2860.6440.6350.991
Concat.0.6720.7390.0000.3430.5630.3480.992 0.7060.7090.0000.1590.6180.6400.992
CA0.6990.7580.0000.3600.3270.5360.991 0.6650.7030.0000.1200.6890.6250.991
PlC-msStacking0.6260.7100.0000.2090.6740.5840.991 0.5620.7020.0000.1150.6240.5610.989
Concat.0.4330.7070.0000.0570.5310.4660.992 0.5000.6850.0000.3140.6360.5980.991
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
EP-msStacking0.8650.8400.7720.8940.9510.9170.923 0.8410.8420.8340.9160.9440.9250.915
Concat.0.8590.7540.7170.7550.8380.8560.778 0.8690.8520.8540.8990.9420.9220.919
CA0.8430.7820.8720.8460.8370.8940.853 0.8300.8260.8810.8570.9310.9140.882
PlC-msStacking0.8060.8100.6910.8800.9140.8960.902 0.7970.8040.8250.9020.9150.9070.900
Concat.0.5450.6690.7220.7970.7990.8180.780 0.7750.8160.8660.8840.9260.9120.900
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
EP-msStacking0.8190.7410.5030.7730.5040.6060.613 0.8520.6930.7200.7450.8450.9180.821
Concat.0.8590.7360.5860.7630.7350.8160.832 0.8740.7220.7090.7190.8260.9200.859
CA0.8600.8050.7750.7770.6460.8680.816 0.8530.6310.6510.5700.8610.9170.884
PlC-msStacking0.8130.7110.7090.6940.8780.9010.829 0.7680.6530.7830.6080.8440.8890.832
Concat.0.5210.6620.7600.6270.7620.8020.691 0.6930.6550.8310.7730.8640.8970.814
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.
InputFusionF1 (ResNeXt) F1 (ViT) mAP (ResNeXt) mAP (ViT)
IDCDΔ\DeltaID ++ CD IDCDΔ\DeltaID ++ CD IDCDΔ\Delta IDCDΔ\Delta
DEM+RGBStacking0.6420.5190.1231.160 0.6620.5070.1551.169 0.6660.5160.150 0.6860.5390.147
Concat.0.5950.5250.0701.119 0.6510.4460.2051.097 0.6020.5340.068 0.7070.5130.194
CA0.5860.4840.1021.071 0.6690.4250.2441.094 0.6210.5450.076 0.7110.5050.205
DEM+NHD+ OSM+RGBStacking0.6490.4840.1661.133 0.6320.4650.1671.096 0.6820.5010.181 0.6880.5140.174
Concat.0.5830.4660.1161.049 0.6520.5130.1391.165 0.5860.5430.044 0.6980.5460.152
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 Δ=ID−CD\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.
InputFusionPrecision (ResNeXt) Precision (ViT) Recall (ResNeXt) Recall (ViT)
IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta
DEM+RGBStacking0.5980.4520.146 0.6060.4760.129 0.7720.6280.143 0.7610.5480.213
Concat.0.5620.5280.034 0.6260.5040.122 0.6340.5370.097 0.6840.4460.237
CA0.5460.4880.058 0.6280.4410.187 0.6400.6230.017 0.7380.5030.235
DEM+NHD+OSM+RGBStacking0.6130.4480.165 0.6140.4390.175 0.7180.5370.181 0.6540.5400.115
Concat.0.5460.3960.150 0.6260.4480.178 0.6400.693-0.053 0.6990.6180.080
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 Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} . Best and second-best scores in each column are indicated in bold and underlined , respectively.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM+RGBStacking0.7140.7790.0870.4130.8230.7070.968 0.7080.7950.2420.3100.8580.7530.966
Concat.0.6680.7710.0000.2760.7970.6850.964 0.7120.8010.0000.4030.8840.7870.971
CA0.6620.7780.0000.2190.8190.6620.965 0.7140.8060.1360.4340.8810.7460.969
DEM+NHD+OSM+RGBStacking0.7150.7710.1630.3510.8590.7190.967 0.6750.7980.1080.2400.8680.7660.965
Concat.0.6770.7710.0890.2000.7970.5790.966 0.7230.7970.1400.2880.8770.7690.971
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM+RGBStacking0.5840.6920.0000.1810.6520.5310.992 0.5140.5980.0000.0000.7710.6720.992
Concat.0.6490.6960.0000.2870.5850.4640.991 0.4180.3230.0000.0240.7050.6590.992
CA0.5170.6060.0000.3270.5300.4170.992 0.4580.1690.0000.0000.7900.5650.992
DEM+NHD+OSM+RGBStacking0.5840.6850.0000.0270.6410.4600.988 0.3710.6460.0000.0000.6220.6210.992
Concat.0.5470.7100.0000.1760.4790.3610.991 0.6010.6280.0000.0300.7110.6290.992
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM+RGBStacking0.8500.7520.8610.8880.9130.8820.918 0.8490.7860.9070.8960.9390.9200.931
Concat.0.7850.7020.5330.7040.8810.8650.726 0.8520.8320.8680.9190.9630.9390.942
CA0.8030.7000.7760.8610.9090.8350.867 0.8540.8220.9250.9190.9660.9190.939
DEM+NHD+OSM+RGBStacking0.8550.7590.8320.8860.9340.8950.924 0.8280.8000.9330.8910.9520.9210.931
Concat.0.8130.6440.7370.7810.8870.8020.818 0.8640.8200.8660.9090.9600.9290.947
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.
InputFusionResNeXt ViT
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM+RGBStacking0.8030.5800.4180.6150.8740.8500.720 0.7420.5460.4000.3570.9530.9220.722
Concat.0.8280.6970.2480.7280.8470.8320.814 0.6590.5570.5240.4250.8890.9220.746
CA0.7330.5380.5600.7810.8880.8570.692 0.6340.5340.7860.4360.9410.8860.792
DEM+NHD+OSM+RGBStacking0.7720.6030.7120.5410.8570.7750.617 0.6240.4860.3340.5490.9050.9000.818
Concat.0.7760.6330.7790.7520.8660.8970.653 0.8030.5990.5020.4920.9160.9030.692
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.
InputmIoU (DeepLabV3+) mIoU (Segformer) Dice (DeepLabV3+) Dice (Segformer)
IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta
DEM0.4860.3560.130 0.5120.4120.100 0.5480.4150.133 0.5760.4740.101
EP-50.4480.3780.070 0.4600.4000.059 0.5130.4230.089 0.5250.4470.079
EP-510.4530.3780.075 0.4850.4030.083 0.5210.4260.095 0.5520.4510.102
S-50.4610.3840.078 0.4840.4300.054 0.5220.4420.080 0.5460.4910.055
S-500.4240.3940.031 0.4320.4070.025 0.4820.4420.040 0.4880.4550.034
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 Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} . Best and second-best scores in each column are indicated in bold and underlined , respectively.
InputPrecision (DeepLabV3+) Precision (Segformer) Recall (DeepLabV3+) Recall (Segformer)
IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta IDCDΔ\Delta
DEM0.6160.4980.118 0.6320.5350.096 0.5510.4410.110 0.5850.4960.089
EP-50.5820.4860.096 0.5940.4940.099 0.5280.4610.066 0.5360.4840.052
EP-510.5810.4710.111 0.6160.4970.119 0.5330.4650.068 0.5610.4900.071
S-50.5920.5080.084 0.6070.5410.066 0.5290.4540.075 0.5550.5110.044
S-500.5320.4900.042 0.5300.5000.030 0.5050.4610.044 0.5120.4740.037
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 Δ=ID−CD\Delta=\mathrm{ID}-\mathrm{CD} . Best and second-best scores in each column are indicated in bold and underlined , respectively.
InputDeepLabV3+ Segformer
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM0.2400.3430.0000.0570.5630.3160.750 0.2960.4070.0000.0760.5900.3390.766
EP-50.2390.3280.0000.0170.4690.2090.720 0.2660.3560.0000.0550.4940.2770.727
EP-510.2620.3670.0000.0780.4560.2570.728 0.2860.3840.0000.0580.5140.2800.750
S-50.2270.3030.0000.0480.5800.2960.732 0.2560.3140.0000.0550.6060.3190.761
S-500.0450.2430.0000.0430.5470.2990.711 0.0370.2490.0000.0520.5550.3230.717
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.
InputDeepLabV3+ Segformer
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM0.1260.1890.0000.0000.2640.0990.714 0.2010.2860.0000.0000.3380.1230.755
EP-50.1050.2600.0000.0270.0230.0070.774 0.1470.3060.0000.0780.0220.0010.783
EP-510.1790.2830.0000.0370.0170.0080.763 0.1740.3070.0000.0660.0160.0170.780
S-50.1380.2310.0000.0440.3650.1260.749 0.1730.2850.0000.0670.4050.1670.783
S-500.0060.1740.0000.0130.3760.1380.766 0.0060.1850.0000.0130.3740.1620.776
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.
InputDeepLabV3+ Segformer
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM0.2950.4150.0000.0790.6300.3920.802 0.3570.4840.0000.1030.6570.4140.819
EP-50.2940.3990.0000.0260.5440.2720.782 0.3290.4280.0000.0740.5680.3490.786
EP-510.3240.4420.0000.1060.5320.3310.790 0.3490.4590.0000.0830.5880.3560.808
S-50.2790.3700.0000.0710.6460.3660.788 0.3140.3830.0000.0750.6750.3920.814
S-500.0580.3010.0000.0610.6280.3730.771 0.0490.3030.0000.0700.6330.3990.776
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.
InputDeepLabV3+ Segformer
af1QalQafQatQcQcaQr af1QalQafQatQcQcaQr
DEM0.1740.2480.0000.0000.3320.1430.779 0.2650.3520.0000.0000.4080.1690.819
EP-50.1480.3220.0000.0410.0340.0100.831 0.1980.3710.0000.1090.0310.0020.837
EP-510.2430.3470.0000.0550.0250.0130.821 0.2350.3720.0000.0890.0230.0250.835
S-50.1920.2870.0000.0660.4350.1680.813 0.2350.3480.0000.0880.4800.2250.841
S-500.0100.2140.0000.0220.4600.1880.824 0.0100.2240.0000.0200.4570.2170.833
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.
More in this topic

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

Explore similar work

CardsList
Feb 6, 2026

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

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.

Lucia Gordon, Serge Belongie, Christian Igel +1

Harvard University, USA · University of Copenhagen, Denmark

Test-Time AdaptationMultimodal RobustnessGeospatial Foundation ModelsRemote SensingTest-Time TrainingRemote Sensing Benchmarks+1
AskPDF
Mar 19, 2026

TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation

Vision-language models (VLMs) have shown promise in earth observation (EO), yet they struggle with tasks that require grounding complex spatial reasoning in precise pixel-level visual representations. To address this problem, we introduce TerraScope, a unified VLM that delivers pixel-grounded geospatial reasoning with two key capabilities: (1) modality-flexible reasoning: it handles single-modality inputs (optical or SAR) and adaptively fuses different modalities into the reasoning process when both are available; (2) multi-temporal reasoning: it integrates temporal sequences for change analysis across multiple time points. In addition, we curate Terra-CoT, a large-scale dataset containing 1 million samples with pixel-level masks embedded in reasoning chains across multiple sources. We also propose TerraScope-Bench, the first benchmark for pixel-grounded geospatial reasoning with six sub-tasks that evaluates both answer accuracy and mask quality to ensure authentic pixel-grounded reasoning. Experiments show that TerraScope significantly outperforms existing VLMs on pixel-grounded geospatial reasoning while providing interpretable visual evidence.

Yan Shu, Bin Ren, Zhitong Xiong +4

University of Trento · MBZUAI · Technical University of Munich +1

Remote Sensing VLMsGeospatial ReasoningRemote Sensing Visual GroundingGeospatial Foundation ModelsRemote Sensing BenchmarksRemote Sensing Change Detection
AskPDF
Mar 26, 2026

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

Current Large Multimodal Models (LMMs) in Earth observation are predominantly evaluated on planar optical tasks and often neglect the vertical dimension, although vertical geometric structure can be critical in applications such as disaster response and urban-morphology analysis. Progress on height-aware reasoning is also hindered by the absence of systematic evaluation: few benchmarks pair optical imagery with height products across pixel-, object-, and scene-level reasoning. To address this gap, we introduce GeoHeight-Bench, a large-scale benchmark for height-aware remote sensing understanding, together with a more challenging terrain-oriented extension, GeoHeight-Bench+. The benchmark is constructed through a scalable, VLM-driven generation pipeline that combines metadata extraction with prompt engineering, and its quality is assessed through a human-in-the-loop verification protocol. To examine whether height-aware reasoning can be learned from optical imagery, we further provide GeoHeightChat, a height-aware baseline that transfers implicit height-related geometric representations into an optical LMM. Evaluations of a broad range of closed- and open-source LMMs show that current models remain limited in their ability to reason about height information, while aligning implicit height priors improves most height-dependent tasks. However, several tasks, particularly slope reasoning and terrain-based flood-susceptibility mapping, remain largely unsolved, highlighting concrete open problems for height-aware GeoAI. Dataset and Code will be released here.

Xuran Hu, Zhitong Xiong, Zhongcheng Hong +3

KTH Royal Institute of Technology · Technical University of Munich · The Hong Kong University of Science and Technology (Guangzhou)

Remote Sensing Image UnderstandingLarge Vision-Language ModelsRemote SensingGeospatial ReasoningRemote Sensing BenchmarksMultimodal Model Evaluation
AskPDF