Climate resilience is a defining challenge of the 21st century, yet public health authorities continue to face difficulties in operationalising state-of-the-art geospatial and environmental science. In Portugal, as in Europe more broadly, extreme temperatures have already increased in frequency and severity, contributing to measurable studies reporting impacts on excess mortality and morbidity. Of relevance, these impacts are often amplified by the simultaneous degradation of air quality. However, evidence on the compounding effect of temperatures and atmospheric composition has largely been event-specific, fragmented across case studies of individual heatwaves, cold waves, or air-quality exceedance episodes, limiting our ability to implement dedicated compound events early-warning systems. The ESA-funded AIR4health project, developed under the Early Digital Twin Components initiative, addresses these gaps by designing innovative algorithms and a user-driven climate service focused on predicting human mortality and morbidity excesses in Portugal, during compound extreme temperature and pollutant exceedance events. The project has developed two Machine Learning (ML)–based AIR4health Downscaling Algorithms and the corresponding Incidence Risk Ratio (IRR) Models depicting, in a 1 km x 1 km spatial resolution, and the daily mortality and hospital admissions excesses per municipality that are attributable to (i) Heat & Ozone (O3) and (ii) Cold & Nitrogen Dioxide (NO2) compound events, using a long (2000-2018) healthcare database for mainland Portugal. These indicators integrate EO data, in-situ air-quality records from the EEA, and CAMS/C3S model outputs to improve the spatial resolution. Results achieved included diverse ML architectures (random forest, XGboost, Neural Network, using linear regression as the non-ML benchmarking model) which were validated against observational records on withheld data, proving the ML models' superiority in spatially depicting O3 and NO2 exceedances, in space and time, proving the added-value of such architectures for hazard mapping, with implications for the health impact assessment, with mean absolute error metrics inferior to 10µg/m3 in almost all cases, including during extreme events. Furthermore, impact models showed the relevance of the combined effect, particularly concerning heat and O3 compound events during which the IRR may increase by 50%, compared to considering heat alone. In complement, users have been involved in defining the requirements for a graphical interface that resonates with the currently existing seasonal surveillance system, with the goal of delivering an interactive dashboard that automatically conveys risk indices into useful geospatial information for the public health sector. With these actions, AIR4health advances beyond current country-level systems by implementing fully spatiotemporal exposure–response modelling. Its dynamic and continuous framework will deliver a prototype DTC capable of providing fine-scale early warning for combined climate and air-quality extremes. By benchmarking results against European-level datasets, AIR4health will support scalable pathways towards relevant practices in planetary health and climate-preparedness, while contributing to the broader European Digital Twin ecosystem.

With an ever-growing recognition of Land Surface Temperature (LST) as a key Essential Climate Variable (ECV), it becomes utmost important to have such a variable at the fine spatial and temporal scales of urban spaces and dynamics. Sentinel-3 provides coarse LST (1 km, daily) based on thermal imagery acquired by its Sea and Land Surface Temperature Radiometer (SLSTR) as well as fine Spectral Directional Reflectances (SDRs, 300 m, every two days) synergically inferred from both SLSTR and the Ocean and Land Colour Instrument (OLCI), which gives the opportunity for using the latter as a predictor in the downscaling of the former. Herein, two scale-invariance-based architectures were developed: a single-timestamp (STS) model, trained with coarse data of the timestamp whose fine target it infers; and a multi-timestamp (MTS) one, trained with multiple timestamps. Note that while several Machine Learning (ML) models besides Linear Regression (LR) were considered for the MTS architecture, only LR was used for the STS one due to the limited amount of available data which the former require for hyperparameter tuning. The models were developed over four Danish Functional Urban Areas (FUAs) using SRD-derived indices and seasonal and geospatial predictors and validated against Landsat data. While Gradient Boosting (GB) achieved the best coarse-scale performance it corresponded to the worst fine-scale performer together with Random Forest (RF), indicating scale invariance breakdown. Tree-based models performed poorly due to extrapolation limitations, whereas Neural Net (NN) and LR proved more robust. After residual correction, single-timestamp LR achieved the best fine-scale performance, making it the most reliable and recommended architecture for operations. The overall results showed that, although ML models may better predict the target at their training scale, their performance may not significantly generalise at others, therefore revealing scale specificity. Furthermore, the results suggested that usage of the more general multi-timestamp architecture instead of the single one may deteriorate performance.

With an ever-growing recognition of Land Surface Temperature (LST) as a key Essential Climate Variable (ECV), it becomes utmost important to have such a variable at both the fine spatial and temporal scales of urban spaces and dynamics. Sentinel-3 provides coarse LST (1 km, daily) based on thermal imagery acquired by its Sea and Land Surface Temperature Radiometer (SLSTR) as well as fine Spectral Directional Reflectances (SDR, 300 m, every two days) synergically inferred from both SLSTR and the optical bands acquired through the Ocean and Land Colour Instrument (OLCI), which gives opportunity for using the latter as predictor in the downscaling of the former. Herein, two scale-invariance-based architectures were developed: a single-timestamp model, trained with the coarse data of the timestamp whose fine target it tries to infer; and a multi-timestamp one, trained with several timestamps and that can infer for any other. While for the case of the multi-timestamp architecture, Machine Learning (ML) models besides Linear Regression (LR) were trained, solely LR was considered for the single-timestamp architecture due to the smaller amount of data available, making it less suitable for hyperparameter tuning. The models were developed over four Danish Functional Urban Areas (FUAs) between 2020 and 2023 using SRD-derived indices, seasonal and geospatial predictors. From 112 Sentinel-3 scenes, 105 were used for training and 7 for validation against Landsat data. While Gradient Boosting (GB) achieved the best coarse-scale performance (test set Root Mean Square Error, RMSE, of 1.56 K), fine-scale predictions showed degraded performance, indicating scale-invariance breakdown. Tree-based models performed poorly due to extrapolation limitations, whereas Neural Net (NN) and LR proved more robust. After residual correction, single-timestamp LR achieved the best fine-scale performance (test set RMSE of 1.40 K), making it the most reliable and operationally recommended architecture.

Intra-urban air temperature is highly heterogeneous, and coarse-resolution products cannot resolve the neighbourhood-scale variability that urban heat management requires. The objective of this study is to assess whether the Urban Thermal Signal Downscaling (UTS-D) methodology, previously developed and tested over Danish Functional Urban Areas (FUAs) with the DAnish ReAnalysis (DANRA), can be re-implemented with a pan-European reanalysis and retrained in cities with contrasting climates, and to quantify the incremental value of a vegetation-seasonality predictor. The approach is applied to four European FUAs spanning contrasting climate zones, i.e., Copenhagen (Denmark), Lisbon (Portugal), Seville (Spain) and Athens (Greece). Two modifications were introduced: the regional weather component was replaced with the pan-European Copernicus Regional Reanalysis (CERRA), enabling application beyond Denmark, and a monthly climatological Normalized Difference Vegetation Index (NDVI) was added as a non-fixed predictor to capture the seasonality of vegetation cooling. A Random Forest model was trained to predict the residual bias between CERRA 2 m air temperature and quality-controlled observations from citizen-owned weather stations, combined with static landscape and urban predictors. Using a 2019–2022 training/cross-validation strategy and an independent 2023 test year, the downscaled model substantially outperformed raw CERRA in all four cities, reducing the test-year mean absolute error from 1.02–1.99 °C to 0.69–1.06 °C (a 32–55% reduction) and reaching R² of 0.94–0.98. Absolute error reductions were particularly relevant on hot days (HD) (0.52–1.21 °C); however, such periods are also the ones during which CERRA errors tend to be largest, indicating that the hot-day added value mainly reflects the correction of a larger reanalysis bias. NDVI provided only a marginal additional improvement (≤0.02 °C; ≤2%), statistically significant but negligible in practical terms. Errors were lowest in Copenhagen, which has the densest citizen-station network, and highest in Seville, which has the sparsest; however, a network-thinning experiment showed that station density explains only part of these differences, pointing to the combined role of station density, the magnitude of the reanalysis error, climate and urban morphology. These results show that UTS-D is methodologically transferable, i.e., it can be retrained successfully with CERRA wherever sufficiently dense citizen-station observations exist, and provide a basis for future heat-health and urban adaptation applications.