Sebastian Scher

Machine learning emulators provide a computationally efficient alternative to physically based regional climate and surface mass balance models, but their applicability depends on their ability to generalize beyond training conditions. In this study, we assess the transferability of a melt emulator trained on output from the regional climate model HIRHAM5 forced by ERA-Interim, CESM2, and EC-Earth under historical and future (SSP5-8.5) climates. Transferability is evaluated across forcing datasets and climate states without retraining. The emulator generalizes reasonably well within a given climate regime, producing biases below the inter-model spread, but exhibits limited transferability between historical and future climate conditions. Predictor distributions analyses, perturbation experiments, and ablation studies indicate that these errors arise from shifts in predictor distributions together with an insufficient representation of the surface state. Although perturbation experiments demonstrate physically plausible relationships between surface energy fluxes and melt generation, physical consistency alone does not ensure robust extrapolation to novel climate conditions. Increasing training data diversity does not overcome these transferability limitations, highlighting the importance of training on data that are representative of the target application. Future emulator development should focus on incorporating information across multiple temporal scales and systematically evaluate transfer performance to improve generalizability.

The unprecedented use of GPS jamming in the Russo-Ukrainian war severely disrupts GNSS Radio Occultation, a key remote sensing technique for climate monitoring and weather forecasting. Since 2018, a growing systematic gap in atmospheric observations has extended far beyond the direct conflict zone. This loss of data substantially impairs atmospheric temperature monitoring, introducing observation errors large enough to bias climate trend estimates.

High-resolution hourly 2 m air temperature fields are essential for weather, climate, and impact applications in complex terrain, but coarse atmospheric products cannot directly resolve the local thermal structure produced by Alpine topography. This study develops and evaluates a terrain-aware residual U-Net for kilometre-scale hourly 2 m temperature downscaling over Austria and the surrounding Alpine region. The model learns the correction between bilinearly interpolated ERA5 reanalysis and the Integrated Nowcasting through Comprehensive Analysis (INCA) high-resolution temperature analysis using dynamic atmospheric predictors, digital elevation model-derived terrain descriptors, and cyclical time features. Performance is evaluated over an independent 2023–2025 test period against INCA and compared with two reference approaches: 1) interpolated ERA5 and 2) a physically interpretable lapse rate baseline with monthly orographic and bias corrections. In addition to conventional deterministic skill metrics, the evaluation examines whether the downscaled fields reproduce physically meaningful Alpine temperature structure, including spatial error patterns, hourly lapse rate variability, elevation-dependent diurnal cycles, high-frequency spatial variability, ridge-valley temperature contrasts, and heat- and cold-event biases. The residual U-Net substantially improves predictive skill relative to both reference methods. It reduces root-mean-square error from 2.39 °C for interpolated ERA5 and 1.97 °C for the lapse rate baseline to 1.22 °C and reduces mean absolute error from 1.69 °C and 1.37 °C for ERA5 and the lapse rate baseline, respectively, to 0.89 °C. These improvements correspond to root-mean-square error reductions of approximately 49 % relative to ERA5 and 38 % relative to the baseline. The model also weakens terrain-locked error structures, produces smaller and less spatially coherent biases, and better reproduces observed lapse rate variability, high-elevation diurnal cycles, small-scale temperature variability, nocturnal ridge-valley contrasts, and event-scale warm and cold biases. A source-transfer experiment using the NASA MERRA-2 reanalysis predictors without retraining shows that the learned correction remains beneficial relative to raw MERRA-2, but with reduced skill compared to the ERA5-driven model, indicating sensitivity to the driving reanalysis distribution. Overall, the results demonstrate that residual deep-learning downscaling can generate physically realistic INCA-scale hourly temperature fields in Alpine terrain, while also highlighting the need to treat such products as high-resolution emulations of the reference analysis rather than direct observations.

High-resolution hourly 2 m air temperature fields are essential for weather, climate, and impact applications in complex terrain, but coarse atmospheric products cannot directly resolve the local thermal structure produced by Alpine topography. This study develops and evaluates a terrain-aware residual U-Net for kilometre-scale hourly 2 m temperature downscaling over Austria and the surrounding Alpine region. The model learns the correction between bilinearly interpolated ERA5 reanalysis and the Integrated Nowcasting through Comprehensive Analysis (INCA) high-resolution temperature analysis using dynamic atmospheric predictors, digital elevation model-derived terrain descriptors, and cyclical time features. Performance is evaluated over an independent 2023–2025 test period against INCA and compared with two reference approaches: 1) interpolated ERA5 and 2) a physically interpretable lapse rate baseline with monthly orographic and bias corrections. In addition to conventional deterministic skill metrics, the evaluation examines whether the downscaled fields reproduce physically meaningful Alpine temperature structure, including spatial error patterns, hourly lapse rate variability, elevation-dependent diurnal cycles, high-frequency spatial variability, ridge-valley temperature contrasts, and heat- and cold-event biases. The residual U-Net substantially improves predictive skill relative to both reference methods. It reduces root-mean-square error from 2.39 °C for interpolated ERA5 and 1.97 °C for the lapse rate baseline to 1.22 °C and reduces mean absolute error from 1.69 °C and 1.37 °C for ERA5 and the lapse rate baseline, respectively, to 0.89 °C. These improvements correspond to root-mean-square error reductions of approximately 49 % relative to ERA5 and 38 % relative to the baseline. The model also weakens terrain-locked error structures, produces smaller and less spatially coherent biases, and better reproduces observed lapse rate variability, high-elevation diurnal cycles, small-scale temperature variability, nocturnal ridge-valley contrasts, and event-scale warm and cold biases. A source-transfer experiment using the NASA MERRA-2 reanalysis predictors without retraining shows that the learned correction remains beneficial relative to raw MERRA-2, but with reduced skill compared to the ERA5-driven model, indicating sensitivity to the driving reanalysis distribution. Overall, the results demonstrate that residual deep-learning downscaling can generate physically realistic INCA-scale hourly temperature fields in Alpine terrain, while also highlighting the need to treat such products as high-resolution emulations of the reference analysis rather than direct observations.

High-resolution hourly 2 m air temperature fields are essential for weather, climate, and impact applications in complex terrain, but coarse atmospheric products cannot directly resolve the local thermal structure produced by Alpine topography. This study develops and evaluates a terrain-aware residual U-Net for kilometre-scale hourly 2 m temperature downscaling over Austria and the surrounding Alpine region. The model learns the correction between bilinearly interpolated ERA5 reanalysis and the Integrated Nowcasting through Comprehensive Analysis (INCA) high-resolution temperature analysis using dynamic atmospheric predictors, digital elevation model-derived terrain descriptors, and cyclical time features. Performance is evaluated over an independent 2023–2025 test period against INCA and compared with two reference approaches: 1) interpolated ERA5 and 2) a physically interpretable lapse rate baseline with monthly orographic and bias corrections. In addition to conventional deterministic skill metrics, the evaluation examines whether the downscaled fields reproduce physically meaningful Alpine temperature structure, including spatial error patterns, hourly lapse rate variability, elevation-dependent diurnal cycles, high-frequency spatial variability, ridge-valley temperature contrasts, and heat- and cold-event biases. The residual U-Net substantially improves predictive skill relative to both reference methods. It reduces root-mean-square error from 2.39 °C for interpolated ERA5 and 1.97 °C for the lapse rate baseline to 1.22 °C and reduces mean absolute error from 1.69 °C and 1.37 °C for ERA5 and the lapse rate baseline, respectively, to 0.89 °C. These improvements correspond to root-mean-square error reductions of approximately 49 % relative to ERA5 and 38 % relative to the baseline. The model also weakens terrain-locked error structures, produces smaller and less spatially coherent biases, and better reproduces observed lapse rate variability, high-elevation diurnal cycles, small-scale temperature variability, nocturnal ridge-valley contrasts, and event-scale warm and cold biases. A source-transfer experiment using the NASA MERRA-2 reanalysis predictors without retraining shows that the learned correction remains beneficial relative to raw MERRA-2, but with reduced skill compared to the ERA5-driven model, indicating sensitivity to the driving reanalysis distribution. Overall, the results demonstrate that residual deep-learning downscaling can generate physically realistic INCA-scale hourly temperature fields in Alpine terrain, while also highlighting the need to treat such products as high-resolution emulations of the reference analysis rather than direct observations.

High-resolution hourly 2 m air temperature fields are essential for weather, climate, and impact applications in complex terrain, but coarse atmospheric products cannot directly resolve the local thermal structure produced by Alpine topography. This study develops and evaluates a terrain-aware residual U-Net for kilometre-scale hourly 2 m temperature downscaling over Austria and the surrounding Alpine region. The model learns the correction between bilinearly interpolated ERA5 reanalysis and the Integrated Nowcasting through Comprehensive Analysis (INCA) high-resolution temperature analysis using dynamic atmospheric predictors, digital elevation model-derived terrain descriptors, and cyclical time features. Performance is evaluated over an independent 2023–2025 test period against INCA and compared with two reference approaches: 1) interpolated ERA5 and 2) a physically interpretable lapse rate baseline with monthly orographic and bias corrections. In addition to conventional deterministic skill metrics, the evaluation examines whether the downscaled fields reproduce physically meaningful Alpine temperature structure, including spatial error patterns, hourly lapse rate variability, elevation-dependent diurnal cycles, high-frequency spatial variability, ridge-valley temperature contrasts, and heat- and cold-event biases. The residual U-Net substantially improves predictive skill relative to both reference methods. It reduces root-mean-square error from 2.39 °C for interpolated ERA5 and 1.97 °C for the lapse rate baseline to 1.22 °C and reduces mean absolute error from 1.69 °C and 1.37 °C for ERA5 and the lapse rate baseline, respectively, to 0.89 °C. These improvements correspond to root-mean-square error reductions of approximately 49 % relative to ERA5 and 38 % relative to the baseline. The model also weakens terrain-locked error structures, produces smaller and less spatially coherent biases, and better reproduces observed lapse rate variability, high-elevation diurnal cycles, small-scale temperature variability, nocturnal ridge-valley contrasts, and event-scale warm and cold biases. A source-transfer experiment using the NASA MERRA-2 reanalysis predictors without retraining shows that the learned correction remains beneficial relative to raw MERRA-2, but with reduced skill compared to the ERA5-driven model, indicating sensitivity to the driving reanalysis distribution. Overall, the results demonstrate that residual deep-learning downscaling can generate physically realistic INCA-scale hourly temperature fields in Alpine terrain, while also highlighting the need to treat such products as high-resolution emulations of the reference analysis rather than direct observations.

High-resolution hourly 2 m air temperature fields are essential for weather, climate, and impact applications in complex terrain, but coarse atmospheric products cannot directly resolve the local thermal structure produced by Alpine topography. This study develops and evaluates a terrain-aware residual U-Net for kilometre-scale hourly 2 m temperature downscaling over Austria and the surrounding Alpine region. The model learns the correction between bilinearly interpolated ERA5 reanalysis and the Integrated Nowcasting through Comprehensive Analysis (INCA) high-resolution temperature analysis using dynamic atmospheric predictors, digital elevation model-derived terrain descriptors, and cyclical time features. Performance is evaluated over an independent 2023–2025 test period against INCA and compared with two reference approaches: 1) interpolated ERA5 and 2) a physically interpretable lapse rate baseline with monthly orographic and bias corrections. In addition to conventional deterministic skill metrics, the evaluation examines whether the downscaled fields reproduce physically meaningful Alpine temperature structure, including spatial error patterns, hourly lapse rate variability, elevation-dependent diurnal cycles, high-frequency spatial variability, ridge-valley temperature contrasts, and heat- and cold-event biases. The residual U-Net substantially improves predictive skill relative to both reference methods. It reduces root-mean-square error from 2.39 °C for interpolated ERA5 and 1.97 °C for the lapse rate baseline to 1.22 °C and reduces mean absolute error from 1.69 °C and 1.37 °C for ERA5 and the lapse rate baseline, respectively, to 0.89 °C. These improvements correspond to root-mean-square error reductions of approximately 49 % relative to ERA5 and 38 % relative to the baseline. The model also weakens terrain-locked error structures, produces smaller and less spatially coherent biases, and better reproduces observed lapse rate variability, high-elevation diurnal cycles, small-scale temperature variability, nocturnal ridge-valley contrasts, and event-scale warm and cold biases. A source-transfer experiment using the NASA MERRA-2 reanalysis predictors without retraining shows that the learned correction remains beneficial relative to raw MERRA-2, but with reduced skill compared to the ERA5-driven model, indicating sensitivity to the driving reanalysis distribution. Overall, the results demonstrate that residual deep-learning downscaling can generate physically realistic INCA-scale hourly temperature fields in Alpine terrain, while also highlighting the need to treat such products as high-resolution emulations of the reference analysis rather than direct observations.

High-resolution hourly 2 m air temperature fields are essential for weather, climate, and impact applications in complex terrain, but coarse atmospheric products cannot directly resolve the local thermal structure produced by Alpine topography. This study develops and evaluates a terrain-aware residual U-Net for kilometre-scale hourly 2 m temperature downscaling over Austria and the surrounding Alpine region. The model learns the correction between bilinearly interpolated ERA5 reanalysis and the Integrated Nowcasting through Comprehensive Analysis (INCA) high-resolution temperature analysis using dynamic atmospheric predictors, digital elevation model-derived terrain descriptors, and cyclical time features. Performance is evaluated over an independent 2023–2025 test period against INCA and compared with two reference approaches: 1) interpolated ERA5 and 2) a physically interpretable lapse rate baseline with monthly orographic and bias corrections. In addition to conventional deterministic skill metrics, the evaluation examines whether the downscaled fields reproduce physically meaningful Alpine temperature structure, including spatial error patterns, hourly lapse rate variability, elevation-dependent diurnal cycles, high-frequency spatial variability, ridge-valley temperature contrasts, and heat- and cold-event biases. The residual U-Net substantially improves predictive skill relative to both reference methods. It reduces root-mean-square error from 2.39 °C for interpolated ERA5 and 1.97 °C for the lapse rate baseline to 1.22 °C and reduces mean absolute error from 1.69 °C and 1.37 °C for ERA5 and the lapse rate baseline, respectively, to 0.89 °C. These improvements correspond to root-mean-square error reductions of approximately 49 % relative to ERA5 and 38 % relative to the baseline. The model also weakens terrain-locked error structures, produces smaller and less spatially coherent biases, and better reproduces observed lapse rate variability, high-elevation diurnal cycles, small-scale temperature variability, nocturnal ridge-valley contrasts, and event-scale warm and cold biases. A source-transfer experiment using the NASA MERRA-2 reanalysis predictors without retraining shows that the learned correction remains beneficial relative to raw MERRA-2, but with reduced skill compared to the ERA5-driven model, indicating sensitivity to the driving reanalysis distribution. Overall, the results demonstrate that residual deep-learning downscaling can generate physically realistic INCA-scale hourly temperature fields in Alpine terrain, while also highlighting the need to treat such products as high-resolution emulations of the reference analysis rather than direct observations.