Urban adaptation to climate change - UC1#

The aim of this use case is to provide comprehensive data on urban adaptation to climate change, offering a clear picture of the current situation.

A persistent challenge in developing effective adaptation strategies is the complexity of integrating and analyzing heterogeneous datasets with varying formats and quality. The UC1 approach addresses this issue through two key components: a process to harmonize diverse datasets into structured data cubes, and a comprehensive toolkit for their analysis and presentation. These tools support UC applications at multiple scales.

At the European level, they help identify cities with similar characteristics and analyze how different factors influence their adaptation capacity. At the local level, specialized "city cubes" serve specific goals, as demonstrated through collaborations with Luxembourg City on managing invasive plant species and support provided to Vienna city initiatives.

EU level - City features collection#

Research questions#

• What are commonalities of and differences between cities in terms of their situation about climate adaptation considering comparable European data sets from various sources (Copernicus, ESTAT)?  Can we identify groupings of cities with similar conditions with respect to different climate pressures (e.g. heat islands, air quality…)
• Can we identify the impacts of certain measures in cities? E.g., what does it mean to increase the urban tree cover as requested by European policies? 

Workflow#

Data and ingestion#

Dataset name	Description	Link to Metadata
GISCO Urban Audit	Provides harmonized statistical data on quality of life in European cities, including demographics, housing, health, and environment.	https://ec.europa.eu/eurostat/web/gisco/geodata/statistical-units/urban-audit
Copernicus DEM	A high-resolution Digital Elevation Model representing Earth's surface, including buildings and vegetation, available in 10m, 30m, and 90m resolutions.	https://doi.org/10.5270/ESA-c5d3d65
Environmental Zones	A classification of Europe into ecological zones based on environmental stratification, supporting biodiversity and land management.	https://sdi.eea.europa.eu/catalogue/idp/api/records/6ef007ab-1fcd-4c4f-bc96-14e8afbcb688
CLMS HRL Imperviousness	Maps the degree of soil sealing (e.g., roads, buildings) across Europe, supporting urban planning and environmental monitoring.	https://land.copernicus.eu/en/products/high-resolution-layer-imperviousness
CLMS HRL Tree Cover Density	Provides percentage-based tree cover data across Europe, useful for forest monitoring and carbon accounting.	https://land.copernicus.eu/en/products/high-resolution-layer-tree-cover-density?tab=main
Urban Atlas (CLMS)	Detailed land cover and land use data for Functional Urban Areas in Europe, supporting urban planning and policy.	https://doi.org/10.2909/fb4dffa1-6ceb-4cc0-8372-1ed354c285e6
Thermal Comfort Indices (ERA5)	Derived from ERA5 reanalysis, this dataset includes indices like UTCI and MRT to assess human thermal stress in outdoor environments.	https://doi.org/10.24381/cds.553b7518
Eurostat Urban Audit (Socio-economic)	Offers socio-economic indicators for European cities, including data on population, education, employment, and income.	https://ec.europa.eu/eurostat/cache/metadata/EN/urb_esms.htm

• Land and climate data are harmonised and ingested in a spatial data cube (spatial coordinates are latitude/longitude). Harmonisation consists for example in ensuring that the dimensions all share the same projection and geographical extent. 
• The spatial data cube is used to compute various city indicators. One indicator has one value per city/timestamp. 
• These land- and climate-based indicators are then fed into a city data cube available at https://github.com/FAIRiCUBE/uc1-urban-climate/tree/master/data/city_features_collection
• At the same time, socio-economic data is ingested. Socio-economic data do not have spatially explicit coordinates, but rather they are indexed by the city identifier. Therefore, indicators are directly fed into the city data cube, and derived indicators are computed. 

Processing steps and ML applications#

To make the most out of the Eurostat socio-economic data, we experimented with ML gap filling methods based on feature correlation or on time series (Long Short-Term Memory method). Unfortunately the gaps proved to be too wide to be filled sensibly without hampering the data quality.

Gap-filling with bidirectional LSTM#

A widely adopted machine learning technique for modeling time series is the Long Short-Term Memory (LSTM) network, a type of Recurrent Neural Network (RNN) designed to learn from sequential data and retain information over long time horizons. An extension of this model, the Bidirectional LSTM (BLSTM), processes data in both forward and backward directions, enhancing its ability to capture temporal dependencies.

In this study, we applied BLSTM models to predict the 'Total Population' indicator using Eurostat data spanning 31 years (1991–2022). Two cities were selected for evaluation:

• Helsinki (FI001C): Exhibits a simple, linear increasing trend.
• Bari (IT008C): Displays a more complex, non-linear pattern.

Each BLSTM model was trained using the first 70% of the time series (approximately 22 years), with a sequence length of 3 (e.g., predicting the value for 2010 using data from 2007–2009). The remaining years were reserved for testing. Training was conducted over 100 epochs.

Results#

• Helsinki: The model achieved a Mean Absolute Percentage Error (MAPE) of 0.63%, closely matching the actual values due to the linear nature of the time series.
• Bari: Despite a lower MAPE of 0.53%, the predictions did not accurately capture the complex trend, indicating limitations in model generalization for non-linear patterns.

The figure below illustrates the predicted vs. actual values for Helsinki (FI001C) on the left and Bari (IT008C) on the right.

Limitations#

This approach has several constraints:

• Limited Data Points: With only 31 observations, training is insufficient for complex time series. A minimum of 50 data points is recommended for reliable performance.
• City/Indicator Specificity: Each model must be trained individually per city and indicator, which is computationally intensive.
• Zero Data Cities: Cities with no historical data cannot benefit from this method unless a proxy city with similar characteristics is identified—a non-trivial task.
• Applicability: The method is most effective for cities/indicators with 1–5 missing years and regular trends (e.g., linear growth).

Find the BLSTM gap-filling script here: Python script

Gap-filling with gradient boosting regressor#

Some indicators exhibit strong interdependencies. For example, the Proportion of population aged 75 years and over is inherently related to the Total Population. These relationships can be leveraged to estimate missing values using machine learning (ML) regression models.

Regression models learn a function that estimates the value of a target feature based on a set of input features. These models range from simple linear regression to more complex approaches such as ensemble methods and deep learning architectures. As a demonstration, we implemented a regression model to estimate the Total number of hours of sunshine per day (EN1002V) using the input features Average temperature of warmest month – degrees (EN1003V) and Average temperature of coldest month – degrees (EN1004V). To train the model, we selected 1673 data points from Eurostat where all three indicators were available. The trained model was then used to gap-fill 299 data points where the two input features were present, but the target feature was missing.

The figure above presents the correlation matrix computed from the 1673 complete data points. The matrix reveals a positive correlation between the target and input features, particularly between EN1002V and EN1003V, with a correlation coefficient of 0.74. This strong correlation suggests that a regression model could perform well in this context. Several regression models were trained using 70% of the available data (1173 data points). Among them, the Gradient Boosting Regressor—an ensemble method based on decision trees—yielded the best performance, with Mean Squared Error (MSE) of 0.48, equivalent to an average error of less than 30 minutes of sunshine per day, and F1 Score of 0.677, indicating promising classification-like performance in a regression context. The figure below shows the distribution of predicted vs. actual sunshine hours, demonstrating a close alignment between the two.

Limitations#

While regression models can be effective for gap filling, several limitations must be considered:

• Manual Feature Selection: Each indicator requires a custom model and manually defined input features based on domain knowledge or correlation analysis.
• Insufficient Training Data: Many indicators lack enough complete data points to train a reliable model.

Find the gradient boosting regressor gap-filling script here: Python script

Cluster analysis based on Urban Atlas land use classes#

To get a general overview of the cities that are similar with respect to land use distribution, clustering can be used. Clustering is a type of Unsupervised Machine Learning tool that aims to put a given data into different clusters, each containing data points that are similar with respect to a set of features. For the task of clustering cities with respect to land use distribution, we have experimented with three different clustering algorithms: k-means, Mean-Shift and agglomerative hierarchical clustering (AHC).

K-Means Clustering#

K-Means is a widely used clustering algorithm that partitions data into K clusters by iteratively assigning points to the nearest cluster center and updating the centers until convergence. The Elbow Method helps determine an optimal K by plotting the inertia (sum of squared distances to cluster centers) for various K values. The "elbow" point indicates diminishing returns in variance reduction. In this study, K = 4 was selected.

Mean-Shift Clustering#

Mean-Shift identifies clusters by shifting a window toward the mean of points within it, repeating until convergence. Unlike k-means clustering, there is no need to specify the number of clusters in advance. Instead, one has to set a window size (bandwidth), which significantly affects results. With a bandwidth of 0.25, four clusters were formed, but 93% of cities were grouped into a single cluster, making the method unsuitable for this dataset.

Agglomerative Hierarchical Clustering (AHC)#

AHC builds a hierarchy of clusters by successively merging the closest pairs of clusters. The user can then select the desired number of clusters from the hierarchy. We selected 4 clusters for consistency with K-Means.

The figure below shows pairwise combinations of the most relevant features (Artificial surfaces, Agricultural areas and Natural and semi-natural areas) for all three methods. We can see that Mean-Shift produced imbalanced clusters, while K-Means and AHC yielded similar, more meaningful groupings, though AHC showed slight overlap. In conclusion, K-Means provided the most interpretable and balanced clustering for this study.

Solutions#

• Interactive view of the city features collection: Streamlit app
• Interactive demo clustering analysis: Jupyter Notebook

Local level: invasive alien plant species#

At local level, the focus was on leveraging earth observation data cubes to monitor invasive plant species. These research questions were developed in collaboration with the Environment delegate of Luxembourg City administration.

Research questions#

• Can we identify correlations between plant neophytes’ presence and spatial and environmental factors in urban areas?
• Can we identify patterns of neophytes spread, and how they affect sensitive areas?

Workflow#

We began by collecting species occurrence data and environmental variables (like shadows, wetness, temperature, and land cover). We then fed these inputs into a species distribution model using MaxEnt, which allowed us to create suitability maps showing where species can thrive. We used these suitability maps to conduct impact assessment to evaluate effects on our specific areas of interest (parks, schools, kindergartens, and ruderal areas). We also performed explanatory model analysis using Shapley Additive exPlanations to understand which factors most influenced our results. We initially considered using a habitat potential model (Ellenberg) component, as shown with the dotted line and red X, but we found this approach was unsuccessful or decided not to use it.

Data and ingestion#

Species occurrences were downloaded from mdata, a database mantained by the National History Museum of Luxembourg. We selected five species from the list of invasive species in Article 12(1) of the EU Regulation 1143/2014 on Invasive Alien Species which can pose risks to people and at the same time had a sufficient number of observations within the city boundary. The species along with the number of occurrences (from 2010) are listed in Table 2.

Species	No. occurrences
Robinia Pseudoacacia L.	205
Reynoutria Japonica	69
Impatiens Glandulifera	69
Heracleum Mantegazzianum	34
Prunus laurocerasus	81
Solidago canadensis	18

Table 3 lists the environmental layers used for this application. All the datasets were pre-processed to match the species distribution model requirements. Shadow index and Topographic wetness index were derived from Luxembourg DEM; gridded land cover was obtained by rasterizing the land cover map, which is a vector dataset; Soil acidity, soil nitrogen and temperature were extracted (clipped) from EU wide datasets, reprojected and regridded to match the target CRS (LUREF) at 10m resolution. Most of the preparatory work has been carried out in EOXHub using Python, except for the generation of the Shadow index and Topographic wetness index maps, which was done in QGIS.

Dataset name	Description	FAIRiCUBE STAC Catalog link
Shadow index	The normalized shadow index quantifies the duration of shadow within 10x10m area.	catalog.eoxhub.fairicube.eu/collections/index/items/shadow_index_city_of_luxembourg
Topographic wetness index	The TWI models soil moisture from the digital terrain model based on the size of the upslope contributing area (how much water flows to a point) and the local slope (how likely water is to stay or flow away).	catalog.eoxhub.fairicube.eu/collections/index/items/wetness_index_city_of_luxembourg
Average monthly air temperature for 2017	Average monthly air temperature for 2017 extracted from the Copernicus Climate Data Store. The data were generated using the urban climate model UrbClim.	catalog.eoxhub.fairicube.eu/collections/index/items/air_temperature_city_of_luxembourg_2017
Soil acidity	Soil Chemical properties based on LUCAS 2009/2012 soil surveys.	catalog.eoxhub.fairicube.eu/collections/index/items/soil_ph_city_of_luxembourg
Soil Nitrogen	Soil Chemical properties based on LUCAS 2009/2012 soil surveys.	catalog.eoxhub.fairicube.eu/collections/index/items/soil_nitrogen_city_of_luxembourg
Land cover	Land Cover of the City of Luxembourg	catalog.eoxhub.fairicube.eu/collections/index/items/land_cover_city_of_luxembourg

Processing steps and ML applications#

Species distribution model#

MaxEnt is a presence-only species distribution modelling technique that applies the principle of maximum entropy to predict species' potential geographic distributions by finding the probability distribution that is most spread out, or closest to uniform, while still satisfying constraints derived from occurrence data and environmental variables. The model estimates the relationship between species occurrences and environmental variables by using presence-only data points (where species have been observed), contrasting these with randomly sampled background points from the study area, and then generating a probability distribution across the landscape that indicates habitat suitability. More about MaxEnt: In this project we used elapid (Anderson, 2021/2023), a Python library that implements MaxEnt.

Explanatory model analysis#

SHAP (SHapley Additive exPlanations) is a unified framework for interpreting predictions of machine learning models (Lundberg & Lee, 2017). It provides consistent and locally accurate feature attribution values, so-called SHAP values, based on game theory principles. Broadly speaking, in the context of machine learning, SHAP treats each input feature as a "player" in a cooperative game where the model prediction is the "payout," with SHAP values quantifying each feature's contribution to the difference between the actual prediction and the baseline average prediction. More about SHAP here

Impact assessment#

The suitability maps were also used to establish the impact on certain areas of interest, such as city parks and areas around schools and kindergardens. We extracted the polygons of the areas of interest from OpenStreetMap. For schools and kindergardens, we further buffered them and clustered the schools/kindergardens close to each other. Finally we computed the mean suitability index for each area of interest ans species. Tthe figure below shows the parks location and the mean suitabilty index table.

Solutions#

• Interactive scrollytelling map: uc1.fairicube.nilu.no
• Poster presented at the GDDS event: doi.org/10.5281/zenodo.16570538

Resources#

• Getting started with FAIRiCube Hub

Partners#

space4environment and Stiftelsen Norsk Institutt for Luftforskning, Norwegian Institute for Air Research.