Accurate regional near-surface temperature forecasting is fundamental to short-range weather services and downstream risk assessment. Existing deep learning-based regional forecasters commonly produce a fixed set of future frames on a prescribed grid, limiting their use when forecast products must be evaluated at query-dependent lead times or display resolutions. To overcome these fixed-output constraints, we formulate regional T2M forecasting as query-conditioned continuous spatiotemporal temperature field evaluation and propose the Continuous Spatiotemporal Temperature Forecaster (CSTF), a neural field that turns forecast lead time and output resolution into explicit queries when evaluating 2-m temperature (T2M). Specifically, CSTF first encodes multivariable ERA5 histories into latent meteorological states and then decodes T2M as a coordinate-based field. Accordingly, spatial location, forecast lead time, and output resolution are introduced as queries, enabling standard hourly forecasts, intermediate lead-time diagnostics, and resolution-controllable outputs within a unified field-evaluation framework. Furthermore, to maintain coherence across flexible field queries, we design spatial-gradient, temporal-difference, and scale-consistency objectives that regularize regional thermal structures, lead-wise evolution, and cross-resolution agreement. Experiments on the Southeast China 0-6 h ERA5-Land benchmark demonstrate that CSTF achieves the best aggregate deterministic skill, including a 17.0 percent reduction in Bias, with global-scope diagnostics further illustrating flexible lead-time and resolution-controllable inference.
Figures & tables
Fig. 1: Comparison between conventional fixed-output and proposed query-conditioned temperature forecasting paradigms. (a) Fixed multi-head models predict predefined lead times and output resolutions. (b) Autoregressive rollout extends forecasts step by step but can accumulate errors across forecast steps. (c) Post-hoc resizing changes the display grid without enforcing consistency across resolutions. (d) Our CSTF formulates regional T2M prediction as continuous field evaluation conditioned on spatial, lead-time, and resolution queries, enabling flexible forecast requests.
Fig. 2: Overall architecture of the proposed CSTF framework. The model encodes ERA5 multi-variable input history into a latent regional weather state within a unified query-conditioned forecasting pipeline. Conditioned on this latent state, a coordinate decoder predicts regional T2M fields from lead-time, spatial-coordinate, and output-resolution queries. Training combines deterministic, gradient, temporal-difference, and scale-consistency objectives, so the same trained model supports continuous lead-time queries, resolution-controllable outputs, and scale-consistency evaluation.
Model
t+6 h
t+5 h
t+4 h
t+3 h
t+2 h
t+1 h
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
Baseline
8.196
3.741
7.992
3.378
7.782
2.968
7.583
2.529
7.419
2.108
7.309
1.779
SMAAT-UNet [ 28 ]
1.989
1.425
1.785
1.270
1.581
1.141
1.347
1.000
1.106
0.840
1.000
0.765
EarthFormer [ 18 ]
2.784
1.254
2.668
1.154
2.600
1.064
2.590
0.990
2.600
0.939
2.615
0.902
ARROW [ 9 ]
1.625
1.131
1.444
1.015
1.246
0.892
1.051
0.768
0.870
0.655
0.739
0.567
Weather-RF [ 29 ]
1.609
1.112
1.446
1.008
1.284
0.908
1.126
0.812
0.970
0.720
0.853
0.647
TABLE I: Lead-wise quantitative comparison for short-range T2M forecasting over Southeast China. RMSE and MAE are reported in degrees Celsius at each supervised forecast lead, where lower values indicate better accuracy. The best result in each column is highlighted in green , and the second-best result is underlined.
Model
RMSE ↓
MAE ↓
Bias ↓
Corr. ↑
Baseline
7.720
2.751
1.247
0.576
SMAAT-UNet [ 28 ]
1.510
1.074
0.329
0.987
EarthFormer [ 18 ]
2.644
1.051
0.598
0.963
ARROW [ 9 ]
1.203
0.838
0.389
0.992
Weather-RF [ 29 ]
1.243
0.868
0.446
0.992
Ours
1.123
0.811
0.273
0.993
TABLE II: Overall quantitative comparison for integer-lead T2M forecasting over Southeast China. Lower is better for RMSE, MAE, and bias magnitude, while higher is better for correlation.
Fig. 3: Qualitative comparison of deterministic T2M forecasts over the Southeast China benchmark. Panels show the ERA5 input context, integer-lead CSTF predictions, ERA5-Land references, and absolute-error maps.
Fig. 4: Global-scope diagnostic visualization of T2M field prediction. Panels compare ERA5-Land references, CSTF predictions, and absolute-error maps under a Robinson projection with polar insets. Zoomed-in views are provided in Appendix.
Variant
t+6 h
t+5 h
t+4 h
t+3 h
t+2 h
t+1 h
Mean
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
w/o aux. losses
1.555
1.085
1.403
0.984
1.248
0.890
1.096
0.800
0.960
0.720
0.871
0.668
1.213
0.858
w/o Lscale
1.575
1.096
1.422
0.996
1.263
0.898
1.100
0.802
0.949
0.712
0.852
0.654
1.220
0.860
w/o Lgrad
1.581
1.106
1.420
1.003
1.264
0.907
1.116
0.817
0.974
0.730
0.875
0.669
1.230
0.872
w/o Ltemp
1.476
1.043
1.337
0.948
1.203
0.861
1.059
0.773
0.921
0.689
0.831
0.636
1.160
0.825
Ours
1.414
1.014
1.284
0.925
1.162
0.845
1.036
0.765
0.909
0.685
0.823
0.631
1.123
0.811
TABLE III: Extended lead-wise ablation of CSTF objective terms on the Southeast China test split. RMSE and MAE are reported in degrees Celsius at each supervised hourly lead; lower is better. This table extends the main-text lead-wise ablation by separately removing the spatial-gradient and temporal-difference regularizers. The best score in each column is highlighted in green , and the second-best score is underlined.
Fig. 5: Continuous lead-time query. The same trained model is queried at 0.5 , 2 , 4.5 , and 6 hours. The 0.5 h and 4.5 h panels are non-integer forecast-only queries, while the 2 h and 6 h panels align with supervised hourly targets.
Fig. 6: Global continuous lead-time query on a Robinson projection with polar insets. The same checkpoint is queried at 0.5 , 2 , 4.5 , and 6 hours, using the same lead-time set as the regional query visualization.
Fig. 7: Resolution-controllable prediction. Fixed input and lead-time query are decoded at 96×96 , 128×128 , 160×160 , and 192×192 grids. Panels show direct resolution-conditioned decoding rather than post-hoc resizing.
Fig. 8: Global resolution-controllable prediction. Orthographic visualizations are generated for different output grid sizes from the same input state and lead-time query.
Variant
96→128
160→128
192→128
Mean
w/o aux. losses
1.396
1.165
1.233
1.265
w/o Lscale
1.449
1.171
1.266
1.295
w/o Lgrad
1.111
1.082
0.997
1.060
w/o Ltemp
1.099
1.131
0.988
1.072
Ours
1.098
1.077
0.982
1.052
TABLE IV: Scale-consistency ablation for resolution-controllable prediction. Values are normalized cross-resolution RMSE after resampling to 128×128 ; lower is better. Best results are highlighted in green , and second-best results are underlined.
Variant
RMSE ↓
MAE ↓
Bias ↓
Grad. err. ↓
Scale err. ↓
w/o aux. losses
1.213
0.858
0.339
0.468
1.265
w/o Lscale
1.220
0.860
0.350
0.382
1.295
w/o Lgrad
1.230
0.872
0.356
0.455
1.060
w/o Ltemp
1.160
0.825
0.281
0.376
1.038
Ours
1.123
0.811
0.273
0.361
1.052
TABLE V: Objective-term ablation on the Southeast China test split. All variants use the same nine ERA5 input variables and architecture. Scale error is the mean normalized cross-resolution RMSE after resampling non-base queries to 128×128 . Best results are highlighted in green , and second-best results are underlined.
Fig. 9: Additional deterministic forecast diagnostic for the global domain ( 180.0∘ W– 180.0∘ E, 90.0∘ S– 90.0∘ N). The case shows input context, ERA5-Land references, CSTF predictions, and absolute-error maps under the main-text global visualization protocol. It supplements the idx0 main-text case and assesses whether the qualitative global diagnostic remains consistent across forecast initializations.
Fig. 10: Lead-wise absolute-error maps for the Southeast China case study ( 100.0∘ E– 120.0∘ E, 20.0∘ N– 40.0∘ N). Errors are shown over the six supervised integer lead times, providing a spatial diagnostic of where forecast discrepancies grow with lead time. The overlaid number on each panel reports the per-frame spatial RMSE ( ∘ C) against the ERA5-Land T2M reference over valid land pixels.
Fig. 11: Continuous lead-time query examples over Southeast China ( 100.0∘ E– 120.0∘ E, 20.0∘ N– 40.0∘ N). The three rows are independent test initializations at 2025-09-05 05:00 UTC, 2025-09-13 17:00 UTC, and 2025-09-22 05:00 UTC. Within each row, the same atmospheric input is queried at multiple lead times, including supervised hourly leads and non-integer lead times. The figure is therefore used to inspect whether the learned field evaluator produces temporally smooth and spatially coherent T2M fields when the requested forecast time is varied continuously. Non-integer lead panels are forecast-only diagnostics because the current ERA5-Land target sequence is available at hourly intervals.
Fig. 12: Resolution-conditioned query examples over Southeast China ( 100.0∘ E– 120.0∘ E, 20.0∘ N– 40.0∘ N). The three rows are independent test initializations at 2025-09-05 05:00 UTC, 2025-09-13 17:00 UTC, and 2025-09-22 05:00 UTC. Within each row, CSTF directly evaluates the same forecast state at different requested output grids rather than resizing a single fixed-grid prediction. The visual comparison checks whether regional thermal structures remain consistent as the output resolution changes, complementing the cross-resolution RMSE values reported in the main text.
Fig. 13: Continuous lead-time query diagnostics over the global domain ( 180.0∘ W– 180.0∘ E, 90.0∘ S– 90.0∘ N). The three rows are independent test initializations at 2025-09-05 05:00 UTC, 2025-09-13 17:00 UTC, and 2025-09-22 05:00 UTC, rendered with the same globe-style projection as the main-text global diagnostic. Within each row, the queried lead time is varied while the input state is fixed, allowing visual inspection of large-scale thermal evolution under the continuous-time interface. Non-integer leads are forecast-only diagnostics because the current target split provides hourly references.
Fig. 14: Resolution-conditioned query diagnostics over the global domain ( 180.0∘ W– 180.0∘ E, 90.0∘ S– 90.0∘ N). The three rows are independent test initializations at 2025-09-05 05:00 UTC, 2025-09-13 17:00 UTC, and 2025-09-22 05:00 UTC. Within each row, the model is queried at multiple output grid sizes for the same forecast case, so changes across columns reflect direct resolution-conditioned field evaluation. These visualizations are used as qualitative checks of large-scale coherence under resolution queries, whereas quantitative cross-resolution consistency is reported separately in the main tables.
Fig. 15: Additional training, benchmark, and ablation diagnostics supporting the main quantitative results. From left to right and top to bottom, the panels show optimization curves, aggregate deterministic skill, lead-wise RMSE, cross-resolution scale consistency, objective-term ablation metrics, and lead-wise ablation behavior. The first panel checks whether model training is stable, the next two panels summarize forecast accuracy on the supervised hourly benchmark, the scale-consistency panel evaluates the resolution-query interface, and the ablation panels isolate the effects of the auxiliary objective terms. These plots are diagnostic summaries; the standardized test metrics and primary comparisons are reported in the main-text tables.
Weather forecasting predicts future atmospheric conditions, from local storms to global patterns, using physical models, observations, and machine learning.
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
Regional climate prediction presents unique challenges for time series foundation models, which typically process temporal patterns through single-pass inference. Expert climatologists, in contrast, employ multi-scale temporal analysis and iterative refinement based on systematic error diagnosis. We present RGMR (Residual-Guided Multi-Resolution Refinement), an inference-time framework that adapts pre-trained foundation models to perform structured coarse-to-fine refinement for climate forecasting without updating backbone parameters. Applied to drought forecasting using the Standardized Precipitation Evapotranspiration Index (SPEI), RGMR is architecture-agnostic across the three TSFM backbones evaluated per site (TimesFM, TimeGPT, TabPFN) and consistently lowers test-set MSE on three South Australian sites and three additional regions outside South Australia. Applied to TimesFM, the wrapper reduces one-month-ahead SPEI MSE by up to 18.9% across the three South Australian sites (mean reduction ≈18.7%). Overall, RGMR provides a practical route for deploying frozen TSFMs in regional climate forecasting workflows.
Wentao Gao, Jiuyong Li, Lin Liu +4
Adelaide University, Adelaide, SA, Australia · CSIRO Technology, Canberra, Australia · CSIRO Environment, Canberra, Australia
Most deep learning weather models assign a fixed set of variables and pressure levels to predefined channels, limiting transfer across atmospheric field configurations. This dependence on a fixed field set limits the transferability of trained models across atmospheric field configurations. We propose FlexCast, a field-adaptive weather forecasting model that uses a single set of parameters to produce identity-aligned forecasts for variable-cardinality subsets drawn from a 69-field ERA5 registry. Specifically, a metadata-conditioned adapter the first encodes variable identity, pressure level, and field type and combines them with spatial features. Then, shared rank-16 projec?tions are modulated by metadata-dependent gates to produce field?specific features, while masked set fusion aggregates the available fields into a fixed-width representation. Subsequently, a multiscale U-Transformer processes the fused atmospheric features, while an identity-aware query decoder produces forecasts for the requested fields. Finally, FlexCast learns a standardized six-hour increment and applies it recursively to generate forecasts at longer lead times. Experiments on the 2020 ERA5 test set demonstrate that FlexCast operates across varying field configurations. Compatible cross-field context is associated with lower forecast errors, whereas mismatched context increases them.
Yuang Zhang, Chen Hui, Weisi Lin +4
Nanjing University of Information Science and Technology · Nanyang Technological University · Harbin Institute of Technology +1