TerraNostra-Bench: Benchmarking Spatial Downscaling and Temporal Rollout for Global 1-km Land-Use Reconstruction
Authors: Amirpasha Mozaffari, Marina Castaño, Stefano Materia, Joan Vedrí, Jordi Varela-Agrelo, Dario Garcia-Gasulla, Etienne Tourigny, Mario Acosta, Amanda Duarte
Organizations: Earth Science Department, Barcelona Supercomputing Center (BSC), Barcelona, Spain · Foundation Euro-Mediterranean Center on Climate Change, Bologna, Italy · AI Institute, Barcelona Supercomputing Center (BSC), Barcelona, Spain
Uncertainty in the terrestrial carbon cycle remains a major constraint in climate projections, partly due to imperfect land surface representations in Earth system models. We propose a data-driven framework for high-resolution land-use reconstruction, treated as a semantic-segmentation task, that combines coarse-resolution scenario data and climate classification with static geophysical features to produce annual land-use maps at 1 km resolution. We introduce TerraNostra-Bench, a change-aware evaluation protocol that separates spatial downscaling from autoregressive temporal rollout and scores models on the rare land-use-change pixels rather than the dominant static background. Using it, we compare convolutional (U-Net) and transformer (Swin) backbones in autoregressive and non-autoregressive schemes against persistence, linear, and shallow convolutional baselines. The autoregressive prior is the most informative input yet accumulates error over long rollouts: autoregressive models lose 5.7-7.2 accuracy points over 13 years, against 2.2-2.8 for non-autoregressive models. This trade-off is governed by the autoregressive choice rather than the backbone. It can be tuned by an inference-time prior-trust threshold that requires no retraining and by the training-time weighting of reconstruction and change objectives; the two levers trace complementary frontiers, with calibration recovering up to 10 accuracy points while keeping change sensitivity above the persistence floor. Extrapolating beyond the training window, backward to 1955-1959 and forward to 2015-2019 under scenario forcing, the models retain most of their in-window skill and show the same trade-off, indicating it is intrinsic to the design. This work provides the methodological foundation and evaluation surface for a forthcoming temporally continuous land-use product.
Figures & tables
Figure 1: Annual global carbon emissions and their partitioning from 1850 to present. The data highlights the significant and persistent contribution of net land-use change emissions (ELUC, orange) to the atmospheric carbon burden, alongside fossil emissions. Source: Global Carbon Budget 2025 ( Friedlingstein et al., 2025 )
HILDA+ (Original)
TerraNostra (Target)
LUH2 (Predictors)
Code
Class Name
Code
Class Name
Abbr.
Full Class Name
11
Urban
1
Urban
primf
Primary Forest
22
Cropland
2
Cropland
primn
Primary Non-Forest
33
Pasture
3
Pasture
secdf
Secondary Forest
40
Forest (Unknown/Other)
4
Forest
secdn
Secondary Non-Forest
41
Forest (Evergreen, needle)
5
Grass/shrubland
pastr
Managed Pasture
Table 1: Overview of Land Use/Land Cover taxonomies and predictor variables. The table displays the original HILDA+ classes, the consolidated TerraNostra target classes, and the LUH2 predictor variable definitions.
Figure 2: Temporal Data Scope. An overview of the dynamic and static predictors along with the target variable over the historical and future timelines.
Figure 3: Architecture & Autoregressive Prior. A schematic of the neural network architecture and the autoregressive training strategy.
Figure 4: Validation-loss convergence for the non-autoregressive runs and the learnable baselines. The linear model (top panel) at a much higher validation loss ( ≈0.57 ) than the deep models ( 0.32 – 0.39 , bottom panel), quantifying the value of spatial context.
Overall
Land-use change
Model
Acc.
mIoU
Acc.
mIoU
Rolling persistence
0.9947
0.9890
0.0000
0.0000
Frozen persistence
0.9704
0.9419
0.1630
0.0883
Linear
0.8058
0.3988
0.3510
0.1458
Simple CNN
0.8193
0.4879
0.3603
0.1466
AR U-Net
0.8072
0.4218
0.3931
0.1661
Table 2: Overall versus land-use-change performance (5-seed means, 2001–2014 average). Overall accuracy is dominated by the pixels that stay in the same class between consecutive years, so persistence baselines are unbeatable on this axis; the land-use-change regime—the scientifically meaningful signal—measures accuracy only on the pixels that actually transitioned. The two regimes separate the models cleanly: the non-autoregressive variants lead on overall accuracy and mIoU, the autoregressive variants—trained with a change-weighted objective—lead on every change metric. The Linear and Simple CNN baselines obtain generally worse results across the metrics. mIoU is the macro-average of per-class IoU over the 7 land classes (Ocean excluded, as it is trivially separable at IoU ≈ 0.999 for every model). Accuracy columns are pixel-level micro averages. The two non-autoregressive overall-mIoU scores (0.5276, 0.5264) are within seed noise of each other.
Δ(Padding16−Padding0)
Δ(Padding32−Padding0)
Model
Acc.
mIoU
Acc.
mIoU
Linear
-0.17%
+0.15%
-0.16%
-0.18%
Simple CNN
+0.02%
+0.20%
-0.13%
-0.24%
AR U-Net
-0.13%
-0.53 %
-0.19%
-0.38%
non-AR U-Net
-0.01%
-0.39%
-0.00%
-0.18%
AR Swin-ViT
+0.17%
+0.20%
+0.31 %
+0.08%
Table 3: Padding effect on inference overall 14-Year rollout mean (3-seed means, 2001-2014 average).
Figure 5: Overall accuracy across the 14-year autoregressive rollout (5-seed mean, shaded ± 1 s.d.). All models decay, but autoregressive models show a much steeper decline than non-autoregressive models.
Model
Year 0
Year 13
Δ (Y0 → Y13)
AR U-Net
0.8476±0.0013
0.7903±0.0018
−0.057
AR Swin-ViT
0.8421±0.0016
0.7701±0.0018
−0.072
non-AR U-Net
0.8503±0.0012
0.8280±0.0012
−0.022
non-AR Swin-ViT
0.8597±0.0010
0.8314±0.0014
−0.028
Simple CNN
0.8323±0.0010
0.8073±0.0012
−0.025
Linear
0.8105±0.0012
0.8004±0.0012
−0.010
Table 4: Rollout endpoints (5-seed means ± s.d.). Overall accuracy at the first (2001) and last (2014) rollout year, and the change across the horizon. All models decline, including the Linear baseline, which carries no temporal state at all and therefore isolates the component of the decline attributable to the test years themselves rather than to error accumulation ( −0.010 ). The autoregressive models lose a further 4 – 6 points beyond that floor, whereas the non-autoregressive models lose only 1 – 2 . Bold marks the best value among the four deep models.
Figure 6: Land-use-change accuracy across the rollout (5-seed mean, shaded ± 1 s.d.). All learned models far exceed the frozen-persistence floor (dotted); autoregressive variants lead on change detection.
Removed input
Val. loss
Δ
% change
Autoregressive HILDA+ prior
0.6854
+0.0983
+16.74
Secondary non-forested ( secdn )
0.6322
+0.0451
+7.68
Primary non-forested ( primn )
0.6210
+0.0339
+5.77
C4 annual crops ( c4ann )
0.6181
+0.0310
+5.28
Köppen-Geiger class (dynamic)
0.6151
+0.0280
+4.77
Secondary deciduous forest ( secdf )
0.6096
+0.0225
+3.83
Table 5: Leave-one-out feature importance. Change in validation loss when each input (or group) is removed, relative to the full model (loss 0.5871 ); a positive Δ means the feature is helpful. Köppen-Geiger and population are supplied as dynamic annual inputs. The autoregressive prior dominates.
Removed input
Val. loss
Δ
% change
Köppen-Geiger class (dynamic)
0.8237
+0.4027
+95.65
Primary non-forested ( primn )
0.4291
+0.0081
+1.92
Secondary deciduous forest ( secdf )
0.4232
+0.0022
+0.52
Population (dynamic)
0.4211
+0.0001
+0.02
Secondary non-forested ( secdn )
0.4131
−0.0079
−1.88
C4 annual crops ( c4ann )
0.4123
−0.0087
−2.07
Table 6: Leave-one-out feature importance (non-autoregressive U-Net). Change in validation loss when each input (or group) is removed, relative to the full non-AR model (loss 0.4210 ); a positive Δ means the feature is helpful. Without the autoregressive HILDA+ prior, the dynamic Köppen-Geiger class dominates, and removing whole feature groups is far more damaging than removing any single LUH2 channel. Absolute losses are not comparable to Table 5 , as the non-AR model is trained without the change-loss term.
Class
AR U-Net
non-AR U-Net
AR Swin-ViT
non-AR Swin-ViT
Urban
0.1864
0.2068
0.1994
0.2043
Cropland
0.5687
0.6410
0.6159
0.6381
Pasture
0.4785
0.5401
0.4676
0.5668
Forest
0.6760
0.7186
0.6512
0.7255
Grass/shrubland
0.2376
0.3605
0.3843
0.3568
Other land
0.7447
0.7430
0.5581
0.7763
Table 7: Per-class IoU (5-seed mean, 2001–2014 average) for the four learned variants. Non-autoregressive models lead on six of the seven classes; the AR Water collapse reflects rollout error accumulation on the one class for which the coarse forcing carries no direct evidence. The ocean class is excluded.
Figure 7: Per-class IoU across the four learned variants (5-seed mean, error bars ± 1 s.d., 2001–2014 aggregate). The autoregressive Water collapse and the difficulty of the rare Urban and Grass/Shrubland classes are visible, while the non-autoregressive variants lead on most classes.
Figure 8: Prior-trust calibration traces a stability–sensitivity frontier. Overall accuracy (stability) versus land-use-change macro-F1 (sensitivity) over the 14-year rollout, as the inference-time trust threshold is swept from 0.70 to 0.95 (circles: AR U-Net; diamonds: AR Swin-ViT; colour: threshold). Stars mark the vanilla autoregressive models (5-seed mean) and triangles the persistence baselines. Raising the threshold moves each model from the change-sensitive vanilla corner toward the stable persistence corner; the two architectures trace nearly the same curve.
Figure 9: Training-time loss balancing traces a stability–sensitivity frontier. Overall 14-year rollout accuracy (stability) versus land-use-change macro-F1 (sensitivity) as the loss weight α is swept from 0 to 1 (circles: U-Net; diamonds: Swin-ViT; colour: α ), five seeds per point. Increasing α shifts weight onto the overall-reconstruction term, moving each model from the change-sensitive corner toward the stable-reconstruction corner; the canonical default α=0.5 (red edge) sits near the knee. Triangles mark the persistence baselines.
Figure 10: The two control levers trace overlapping but non-identical frontiers. Overall rollout accuracy (stability) versus land-use-change macro-F1 (sensitivity), overlaying the inference-time trust-threshold sweep (solid) and the training-time loss-balance sweep (dashed); circles denote U-Net and diamonds Swin-ViT. Stars mark the vanilla autoregressive models ( α=0.5 , T=0 ) and triangles the persistence baselines. The two levers coincide in the middle of the frontier but each reaches a region the other cannot: the trust threshold extends further toward the stable persistence corner, loss-balancing further toward the change-sensitive corner.
Backward (1955–1959)
Forward (2015–2019)
Model
Overall Acc.
Change Acc.
Overall Acc.
Change Acc.
AR U-Net
0.8246
0.3207
0.8140
0.4212
non-AR U-Net
0.8425
0.3559
0.8259
0.3923
AR Swin-ViT
0.8130
0.3010
0.8032
0.4452
non-AR Swin-ViT
0.8583
0.3312
0.8303
0.4131
Simple CNN
0.8280
0.2987
0.8082
0.4188
Table 8: Out-of-window extrapolation (5-seed means, 5-year rollout averages). Backward : 1955–1959, conditioned on 1960, historical forcing. Forward : 2015–2019, conditioned on 2014, SSP2-4.5 forcing (SSP3-7.0 and SSP4-6.0 agree to within 0.001 ). Both windows are strictly outside the 1960–2000 training period. All learned models exceed the frozen-persistence change floor ( 0.163 ).
Figure 11: Extrapolation in both temporal directions. Per-year overall accuracy for the backward (1955–1959) and forward (2015–2019, SSP2-4.5) rollouts, framing the in-window 2001–2014 period. Autoregressive models drift away from the conditioning anchor in both directions; non-autoregressive models remain flat.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12: Spatial five-fold split. Top: FPS patch centres coloured by their exclusive 40×40 macro-cell fold assignment. Bottom: the five cross-validation rounds; each round trains on four folds (grey) and holds out one fold (colour) for testing. The canonical 70/15/15 split used in the main results is drawn from the same macro-cell machinery.
Overall acc.
Macro F1
Mean IoU
Change acc.
Change F1
Fold 0
0.793
0.688
0.562
0.429
0.281
Fold 1
0.789
0.708
0.581
0.477
0.317
Fold 2
0.764
0.712
0.582
0.452
0.311
Fold 3
0.785
0.692
0.566
0.437
0.297
Fold 4
0.757
0.688
0.558
0.434
0.285
Fold mean ± std
0.778±0.016
0.698±0.011
0.570±0.011
0.446±0.019
0.298±0.016
Appendix
Table 9: Split sensitivity of the non-AR U-Net. Fourteen-year rollout means on the held-out fold of each cross-validation round, against the canonical-split reference (epoch-9 checkpoint; ± std over five evaluation seeds, i.e. pure sampling noise). Class-balanced metrics (macro F1, mean IoU) are insensitive to the split; overall accuracy varies with the regional class composition of the held-out area, as expected for spatial cross-validation.
Overall acc.
Macro F1
Mean IoU
Change acc.
Change F1
Fold 0
0.738
0.569
0.458
0.420
0.269
Fold 1
0.777
0.667
0.545
0.457
0.267
Fold 2
0.643
0.577
0.448
0.423
0.240
Fold 3
0.746
0.601
0.484
0.413
0.254
Fold 4
0.614
0.553
0.424
0.419
0.262
Fold mean ± std
0.704±0.071
0.593±0.045
0.472±0.046
0.426±0.018
0.258±0.012
Appendix
Table 10: Split sensitivity of the AR U-Net. Fourteen-year rollout means on the held-out fold of each cross-validation round (best-validation checkpoints at epochs 2–3 under a fixed training budget), against the canonical-split model evaluated at matched maturity (epochs 2 and 3) and at its best-validation checkpoint (epoch 4; the setup reported in the main text). At matched budget the canonical scores are consistent with the fold distribution, but the fold spread is four times larger than for the non-AR model (Table 9 ) and orders by the land-use-change fraction of the held-out regions (folds 2 and 4, ∼46% change pixels, score lowest) — the HILDA prior that feeds the autoregressive design is less informative where landscapes change.
Figure 13: High-change patch, northwest India ( 28.8∘ N, 74.0∘ E; 23% of pixels change class between 2000 and 2013). Rows: ground truth and the four models; columns: rollout years 0 (2000), 6 (2006) and 13 (2013). The number in each panel is its pixel accuracy against ground truth. All four models stay within a few points of each other throughout (68–72% at year 13), dip at year 6 and partly recover; all smooth the grass/shrubland mosaic into cropland.
Figure 14: High-change patch, Kansas City, USA ( 39.1∘ N, 94.7∘ W; 21% change). Layout as in Figure 13 . Accuracy falls for all models, slightly more for the autoregressive variants (AR U-Net 60→50% , AR Swin-ViT 62→49% ) than for the non-autoregressive ones ( 63→54% for both). All models capture the urban core but smooth the surrounding pasture–cropland mosaic.
Figure 15: Moderate-change patch, US Great Plains ( 44.0∘ N, 99.5∘ W; 10% change). Layout as in Figure 13 . The non-autoregressive models remain nearly flat (U-Net 68→66% , Swin-ViT 70→68% ), as does the AR U-Net ( 65% throughout), whereas the AR Swin-ViT declines from 70% to 61% . All models reproduce the large-scale pasture–cropland pattern but miss the river and its riparian vegetation.
Figure 16: Moderate-change patch, northern Cameroon ( 11.3∘ N, 14.6∘ E; 10% change). Layout as in Figure 13 . The clearest example of autoregressive drift: the AR U-Net falls from 59% to 48% and the AR Swin-ViT from 72% to 49% as both progressively overwrite the eastern pasture with cropland, whereas the non-autoregressive U-Net stays largely stable ( 73→67% ) and the non-autoregressive Swin-ViT declines less ( 73→58% ).
Uncertainty in the terrestrial carbon cycle remains a major constraint in climate projections, partly driven by the uncertainties affecting the land surface representation and variability in Earth system models. To address this limitation, we present a data-driven framework AI4Land, for generating high-resolution historical reconstructions and future projections of key land surface variables. The framework follows a two-phase approach using a U-Net architecture. In the first phase, which is the focus of this work, it reconstructs annual land use and land cover by integrating coarse-resolution scenario data with static geophysical features. In a planned second phase, the resulting high-resolution maps will be used to predict dynamic biophysical variables, particularly leaf area index, at finer temporal scales. Trained on Earth observation data, the models learn to reproduce spatially explicit and physically consistent land surface patterns, extending temporal coverage to periods lacking direct observations. AI4Land was developed and trained on MareNostrum5, demonstrating how GPU-accelerated HPC infrastructure enables global-scale climate AI pipelines. The final product is a suite of open-source emulators designed for real-time coupling with digital twin platforms, such as those developed under the Destination Earth initiative. By delivering realistic and evolving land surface conditions on demand, this work aims to reduce critical uncertainties and improve the predictive power of next-generation climate simulations.
Earth Science Department, Barcelona Supercomputing Center, Barcelona, 08034, Spain · AI Institute, Barcelona Supercomputing Center, Barcelona, 08034, Spain
Many Earth Observation applications need land-use/land-cover maps that are both precise and frequently updated, yet the strongest Earth Observation foundation models build their embeddings from a full year of observations. We present a controlled study of the temporal sensitivity of Tessera, one of these leading foundation models, for land-use/land-cover mapping. Keeping the encoder frozen, we recompute its embeddings over varying observation windows, from a full year down to a single day. We use them as inputs to a linear probe and a UNet segmentation head, benchmarking both of them against from-scratch networks on LUCAS, DynamicEarthNet, and PASTIS-R datasets. We show that the value of the embeddings is task-dependent. Where classes are separated by phenology, as for the crop types of PASTIS-R, they reach a mean Intersection-over-Union of 58.3, about 46% above the best from-scratch model. Where classes are temporally stable (e.g., forests in DynamicEarthNet and LUCAS), embedding-based and from-scratch models match only under full supervision. On both datasets, Tessera embeddings remain markedly more label-efficient. Degradation under shorter temporal windows is gradual and class-dependent. Contracting the window from one year to one month costs 39% of the segmentation accuracy on PASTIS-R but only 5% on DynamicEarthNet. Single-day embeddings still classify land cover in LUCAS at 3.4 times the chance level. Our study shows that temporal coverage is therefore a tunable cost rather than a fixed prerequisite, opening regimes such as near-real-time mapping and faster land-use/land-cover refresh cycles.
Julia Guerrero-Viu, Alex López-Cifuentes, Ignacio Pérez-Villar +1
Land Surface Temperature (LST) is a key variable for various applications, such as urban climate and ecology studies. Yet, existing satellite-derived LST products provide either high spatial or high temporal resolution, resulting in a fundamental trade-off between the two. To address this trade-off, we combine observations from a geostationary and a polar orbiting satellite and provide LST fields at high spatial and high temporal resolution (1 km at 15-min intervals). We demonstrate their application for intraday forecasting of LSTs. To estimate LST fields at high spatiotemporal resolution, a U-Net model is trained to map LST fields from SEVIRI/MSG (3 km and 15 min resolution) to LST fields from Terra/Aqua MODIS (1 km, 4 overpasses per day) that are collocated in space and time. The presented model has been trained on LSTs across large European cities with a population exceeding 1 million inhabitants, and achieves an RMSE = 1.92°C and near-zero bias MBE = 0.01°C on the hold-out test set. As a second step, we present an LST nowcasting model based on ConvLSTM architecture, trained across downscaled LST fields with forecast lead times of 15 to 75 minutes. The nowcasting model outperforms a persistence and a Climatological Rolling Median benchmarks, with RMSEs of 0.57 to 1.15°C for the considered lead times and biases ranging from −0.1 to 0.14°C. An additional validation conducted against independent MODIS overpasses confirms robust performance. Our LST forecast model at high spatiotemporal resolution is directly applicable to operational satellite-based LST monitoring.
Solomiia Kurchaba, Angela Meyer
Department of Geoscience and Remote Sensing · School of Engineering and Computer Science, Bern University of Applied Sciences