With a concentration of people, activities, and infrastructures, urban areas are particularly vulnerable to the negative effects of climate change. Among others, the intensification of the Urban Heat Island (UHI) effect is leading to an increased impact on citizen health and the urban ecosystem. In this context, this study aims to investigate the effect of urban morphology and land cover composition—which are established by exploiting the Local Climate Zone (LCZ) classification system—on two urban climate indicators, i.e., Land Surface Temperature (LST) and air temperature. The study area is the Metropolitan City of Milan (northern Italy). LCZ and LST maps are derived by leveraging satellite imagery and building height datasets. Both authoritative and crowdsourced in situ measurements are used for the analysis of air temperature. Several experiments are run to investigate the mutual relation between LCZ, LST, and air temperature by measuring LST and air temperature patterns in different LCZs and periods. Besides a strong temporal correlation between LST and air temperature, results point out vegetation and natural areas as major mitigating factors of both variables. On the other hand, higher buildings turn out to increase local air temperature while buffering LST values. A way lower influence of building density is measured, with compact building areas experiencing slightly higher air temperature yet no significant differences in terms of LST. These outcomes provide valuable tools to urban planners and stakeholders for implementing evidence-based UHI mitigation strategies.

Spatial discretization in hydrological models has a strong impact on computation times. This study investigates its effect on the performance of the Soil and Water Assessment Tool (SWAT) applied to a French Mediterranean watershed. It quantifies how spatial discretization (the number of sub-basins and hydrological response units (HRUs)) affects the SWAT model’s performance in simulating daily streamflow and whether this effect depends on the choice of soil and land use input datasets. Sixty-eight SWAT model configurations were created using various soil and land use datasets and 17 discretization setups, evaluated from 2001 to 2021 with the Kling–Gupta efficiency (KGE) metric. The key findings include (1) while the number of sub-basins does not impact model performance, increasing HRUs significantly degrades it (KGE loss of 0.13 to 0.26) regardless of the number of sub-basins or input datasets. (2) SWAT is found to be more sensitive to variations in soil datasets than in land use datasets, but the observed performance decline with more HRUs is attributed to the calibration process and the increased heterogeneity in soil types rather than input dataset spatial resolution. (3) Minimizing the number of HRUs may improve both the accuracy of streamflow simulations and the computational efficiency of the SWAT model.

The quantification of present and future water resources at the regional scale is critical for designing climate change (CC) adaptation strategies and assessing flood/drought risks. Process-oriented Hydrological Models have traditionally been used for hydrological projections under climate change. Recurrent neural networks of the Long Short-Term Memory (LSTM) type have recently emerged as a powerful substitute, showing strong performance in temporal induction (TI) task [1], (when historical data of predicted basins are used for model training). 10-year daily streamflow predictions over French catchments demonstrated excellent accuracy [2], highlighting the potential of LSTM for predictions over future periods. However, their ability to operate under non-stationary CC conditions remains largely unexplored. The Explore 2 project [3] provides a multi-scenario and multi-model ensemble of future streamflow projections over France (3 greenhouse gas emission scenario - 17 climate projections - 9 hydrological models). The selection of hydrological models was based on the state of the art in 2021, just before the emergence of LSTM in hydrological applications. Deep learning based models were therefore not included in the ensemble. In this study, we harness the Explore2’s spatially consistent datasets to evaluate LSTM performance against the process-oriented hydrological models. A subset of catchments with low anthropogenic influence and sufficiently long streamflow records is selected. These catchments are used for model training. Model evaluation is then performed under the RCP 8.5 emission scenario until 2100. A focus on streamflow elasticity curves [4] is proposed to assess the hydrological response to a change in climatic forcing. Catchments are clustered using their hydrological signatures to assess the ability of models to preserve hydrological dynamics.This work aims to clarify the role of LSTM in hydrological CC studies and opens new opportunities for reliable streamflow projections. Beyond this application, predictions outside the country could be investigated to extend the projections outside the national boundaries. [1] Kratzert, F.; Klotz, D.; Brenner, C.; Schulz, K.; Herrnegger, M. Rainfall–Runoff Modelling Using Long Short-Term Memory (LSTM) Networks. Hydrology and Earth System Sciences 2018, 22 (11), 6005–6022. https://doi.org/10.5194/hess-22-6005-2018. [2] Assessing Temporal and Spatial Generalization of Lstms for Streamflow Modeling in French Watersheds with and Without European Training Data. https://doi.org/10.2139/ssrn.5286855. [3] Sauquet, E. et al. A Large Transient Multi-Scenario Multi-Model Ensemble of Future Streamflow and Groundwater Projections in France. EGUsphere 2025, 1–41. https://doi.org/10.5194/egusphere-2025-1788. [4] Anderson, Bailey J., Manuela I. Brunner, Louise J. Slater, et Simon J. Dadson. « Elasticity Curves Describe Streamflow Sensitivity to Precipitation across the Entire Flow Distribution ». Hydrology and Earth System Sciences 28, no 7 (2024): 1567‑83. https://doi.org/10.5194/hess-28-1567-2024.

Predicting flows in ungauged basins is a key challenge for integrated water resource management and hydrological risk prevention. To fill these gaps, regionalization approaches have generally relied on conceptual or physically-based hydrological models whose parameters are calibrated on gauged rivers and then transferred to ungauged rivers. However, these methods often have significant limitations in terms of robustness and accuracy, particularly when applied in heterogeneous hydrological contexts [1].With the growing adoption of machine and deep learning by the hydrological community, new opportunities are emerging for operational hydrology. In particular, recurrent neural networks such as Long Short-Term Memory networks (LSTM) have proven to be effective in exploiting large databases of observed flows and climate forcings compared to traditional locally or regionally calibrated approaches [2]. Nevertheless, LSTM is still rarely used in France for prediction outside of a few research projects (e.g. [3,4]).The objective of this study is to compare an LSTM model with the GR5J model [5,6] in a regionalization exercise with a special focus in flood prediction. The evaluation is carried out on the French catchments of the Explore 2 project, which gather more than 600 catchments [7]. The GR5J model, which stands for Génie Rural Journalier à 5 Paramètres (5-parameter daily rural engineering model), is a widely used reference model in catchment hydrology modeling due to its simplicity and flexibility. GR5J parameters are regionalized with different algorithms, including traditional spatial proximity and catchment similarity methods as well as a machine learning method based on random forest regression [8]. A direct assessment of flood statistics is also performed with random forest regression as a benchmark for flood prediction.The models are evaluated according to global criteria as well as hydrological signatures representative of flood hazards. The hydrological characteristics of the catchments are analyzed to identify favorable and unfavorable conditions for regionalization.This study discusses the prospects offered by deep learning for hydrological regionalization and its future integration into operational applications such as hydrological projection. [1] Guo et al. (2020). https://doi.org/10.1002/wat2.1487[2] Kratzert et al. (2019). https://doi.org/10.5194/hess-23-5089-2019[3] Hashemi et al. (2022). https://doi.org/10.5194/hess-26-5793-2022[4] Puche et al. (2026, in review). http://dx.doi.org/10.2139/ssrn.5286855[5] Perrin et al. (2003). https://doi.org/10.1016/S0022-1694(03)00225-7[6] Le Moine (2008). https://hal.inrae.fr/tel-02591478v1[7] Sauquet et al. (2025). https://doi.org/10.5194/egusphere-2025-1788[8] Saadi et al. (2019). https://doi.org/10.3390/w1108154

Assessing the benefits of increasing the discretization level of semi-distributed hydrological models is of great importance for hydrological applications. The impact of spatial discretization on model performance is investigated with the use of the Soil and Water Assessment Tool (SWAT) model when applied on a Mediterranean watershed (Argens, France). This study aims to explore how the spatial discretization (number of sub-basins and of hydrological response units (HRUs)) affects the model’s performance at simulating daily streamflows, and if the choice of soil and land use input datasets modifies model accuracy. Low and moderate resolution soil (5 km and 250 m) and land use (400 m and 100 m) maps are considered. Four SWAT input sets are created, each corresponding to a different combination of land use and soil datasets. Each input set is used to build 17 configurations with an increasing number of sub-basins (4, 12, and 18) and HRUs (from 4 to 320). The 68 models (4 input sets x 17 configurations) are evaluated on the 2001-2021 period using the Kling-Gupta efficiency (KGE) metric. Results indicate no influence of the number of sub-basins on SWAT performance. However, increasing the number of HRUs leads to a significant performance decrease (from 0.13 to 0.26 of KGE loss), regardless of the number of sub-basins and input datasets. The SWAT model is found to be more sensitive to soil dataset than to land use dataset. Despite significant differences in hydrological soil groups between the two soil maps, no clear impact on the derived hydrological properties is observed, such as the curve number. The observed decline in SWAT performance with an increasing number of HRUs is attributed to the calibration process rather than the soil and land use input datasets. This study suggests that, when the calibration of the semi-distributed SWAT model is not performed at the finer spatial HRU level, an increase in the spatial discretization does not lead to an improvement of the overall model accuracy. Therefore, minimizing the number of HRUs during the watershed subdivision is recommended for getting optimal simulations of streamflow while dealing with the computational efficiency of SWAT.