Land use map of Istanbul

Land use data of Istanbul was obtained by developing an LCZ map. LCZ serves as a categorization system delineating areas based on uniform surface cover, structure, material composition, and human activity, extending across spatial scales ranging from hundreds of meters to several kilometers (Oke and Stewart, Reference Oke and Stewart2012). The LCZ classification system divides urban areas into distinct classes based on their land cover and morphology, with each class representing a unique combination of climate-related properties (Aslam and Rana, Reference Aslam and Rana2022). The classification system considers factors such as vegetation cover, impervious surfaces, building height, and sky view factors. These factors influence the energy balance and microclimate of urban areas, leading to variations in temperature, humidity, wind patterns, and other climatic parameters (Oke and Stewart, Reference Oke and Stewart2012).

LCZ maps are generated using the LCZ Generator website (Demuzere et al., Reference Demuzere, Kittner and Bechtel2021) which utilizes satellite images and user-created training areas (TAs). The method developed by Demuzere et al. (Reference Demuzere, Hankey, Mills, Zhang, Lu and Bechtel2020) involves drawing polygons on Google Earth Engine to represent different LCZ classes by each representative color. The polygons must meet certain criteria, according to Demuzere et al. (Reference Demuzere, Hankey, Mills, Zhang, Lu and Bechtel2020), such as being representative and homogeneous, with a minimum side length of 200 m and a simple shape. To prevent confusion, there should be a minimum distance of 100 m between polygons of different LCZ classes. Factors like persistence, seasonality, and land characteristics should be considered when selecting areas. Areas under construction or with specific land features should be appropriately classified. Finally, the polygons should be distributed across the region of interest or city.

The total created TAs were 165 which represented 15 out of 17 classes. LCZ classes 7 and 10, corresponding to lightweight low-rise and heavy industry, were excluded from the representation, as they were deemed non-existent in the Istanbul metropolitan city based on the examination of land use maps from each district in Istanbul, prepared by the Directorate of Urban Planning of Istanbul Metropolitan City (IMM, 2017). A couple of control methods were used to ensure the homogeneity of the polygons. The Istanbul map of the buildings’ number of floors provided by the Directorate of Geographic Information Systems of Istanbul Metropolitan Municipality (IMM, 2016) was used to control the heights of the buildings in the selected TA that correspond to the LCZ class. Furthermore, the Street View map from Google Earth was used to inspect the type as well as the heights of the buildings of the selected area to validate the selection of the LCZ class.

The file of the Google Earth Engine that contains the TAs was uploaded to the LCZ Generator website, where the personal information of the user, and the reference date of the satellite images along with other information are required. After submitting the TAs file, the subsequent back-end process involves LCZ classification and quality control. To address potential issues, TAs undergo a preliminary step, reducing the surface area of large polygons to approximately 350 m radius, preventing imbalances and computational inefficiencies in the classifier (Demuzere et al., Reference Demuzere, Kittner and Bechtel2021). The default World Urban Database Access Portal Tools (WUDAPT) workflow employs TAs and Landsat 8 data in a random forest classifier (Demuzere et al., Reference Demuzere, Hankey, Mills, Zhang, Lu and Bechtel2020). The LCZ Generator enhances this process by incorporating additional earth observations, including data from Sentinel-1, Sentinel-2, and terrain and forest canopy height data (Demuzere et al., Reference Demuzere, Kittner and Bechtel2021). The system is designed to easily integrate new datasets. Quality control is ensured through an automated cross-validation approach with 25 bootstraps (Bechtel et al., Reference Bechtel, Alexander, Beck, Böhner, Brousse, Ching, Demuzere, Fonte, Gál, Hidalgo, Hoffmann, Middel, Mills, Ren, See, Sismanidis, Verdonck, Xu and Xu2019; Demuzere et al., Reference Demuzere, Kittner and Bechtel2021). The LCZ map, generated from TAs and input features, was sent via e-mail to the user, and an enhanced version was obtainable through a morphological Gaussian filter. This method surpasses the traditional majority post-classification approach by considering factors such as distance from the kernel center and variations in patch sizes, resulting in a more accurate representation (Demuzere et al., Reference Demuzere, Kittner and Bechtel2021).

Local climate zone areas: GIS analysis

QGIS software program (QGIS, 2023) was used to calculate the areas of each class for the city in general and for each district as well, to analyze the land use of the districts, and to interpret its relationship with the environmental issues addressed in this study.

First, the LCZ map of Istanbul was prepared going through several steps. Foremost, the style layer of the LCZ Classes consisting of the corresponding colors and names was defined. Then the map was intersected with the map layers of the district’s boundaries in the country, by exporting the layer and then cropping it according to the LCZ map, which eliminated the districts outside the LCZ map’s boundaries. The boundary layers were retrieved from OpenStreetMap (Raifer, Reference Raifer2023) using the coding developed by Kargın (Reference Kargın2020). Consequently, the LCZ map was also cropped to fit the boundaries of the districts and eliminate the areas outside the city. This was achieved by executing the Clip raster by Mask Layer function from the raster menu in QGIS.

Second, the new cropped map was converted from raster to vector to be able to extract the LCZ data from the map. The class names were added to their corresponding numbers in the Attribute Table of the map layer using a case statement. As the map contains thousands of polygons with different LCZ classes, the data was dissolved by the class number using the Dissolve function under the vector menu, in order to aggregate the LCZ data for the whole city. This step was repeated to obtain the LCZ data in each district using the districts-intersected-LCZ map, but before dissolving the data, the vector LCZ map was intersected with the district’s boundaries layer using the Intersection function under the vector menu.

Finally, the areas of the districts as well as the classes were calculated using the area function under the Attribute Table of the map, subsequently, the percentage of each class was also calculated, and these data were extracted as an Excel file.

NbS characterization

For this research, 12 blue-green infrastructure solutions identified by Petsinaris et al. (Reference Petsinaris, Baroni and Georgi2020) were chosen to address the impacts of climate change in Istanbul. These solutions specifically target stormwater management and the urban heat island (UHI) effect. The selected solutions are detention basins, green roofs, green walls, infiltration basins, infiltration trenches, pervious surfaces, rain gardens, rainwater harvesting, retention ponds, roadside green infrastructure, swales, and urban gardens and parks. The effectiveness, cost, and maintenance values (CIRIA, 2023; Petsinaris et al., Reference Petsinaris, Baroni and Georgi2020) are based on general ranges and may vary depending on specific implementation factors and local conditions. The stormwater management, temperature mitigation values, and suitable LCZ classes are based on research studies and expert opinions. Different environmental issues in different districts of a city can be differentiated through a variety of methods. As an example, the characteristics of green roofs are presented in Table 1. These criteria were used subsequently to evaluate the solutions, assigning scores based on their effectiveness and suitability in addressing specific climate change challenges. These scores then were integrated into the Bayesian network model.

Table 1. Green roofs characteristics

Influence diagram

The study utilized GeNIe software (Druzdzel and Sowinski, Reference Druzdzel and Sowinski2023) to construct a Bayesian network model, enabling the assessment of variable probabilities within the study’s scope. GeNIe, developed by Druzdzel and Sowinski in 2015, is a user-friendly tool for creating and analyzing Bayesian networks, emphasizing inference using evidence to update variable probabilities based on observed data. The initial step involved building an influence diagram comprising variables such as temperature, humidity, precipitation, blue-green infrastructure, cost, maintenance, LCZ, and applicability. These variables were interconnected to depict their relationships, with meteorological parameters influencing BGI selection and cost impacting BGI applicability. LCZ directly influenced BGI suitability within the model. This approach facilitated the assessment of climate change-related challenges and the effectiveness of 12 selected BGI.

Figure 1 illustrates that meteorological parameters (temperature, humidity, and precipitation) serve as the direct parent variables for the BGI variable. This implies that the selection of any BGI is directly influenced by these meteorological parameters. Additionally, temperature has a significant influence on decision-making, acting as the upper parent variable, as it affects humidity, which, in turn, influences precipitation. Furthermore, the BGI, cost and maintenance, and LCZ nodes act as parents for the applicability node, which is categorized as a child node in this context.

Figure 1. Blue–green infrastructure applicability influence diagram.

Conditional probability tables (CPT)

The CPTs used in the BBN model were developed by the authors of this paper, who acted as experts in interpreting the obtained data. The CPTs were constructed based on extensive review and synthesis of relevant literature, expert knowledge, and data analysis. The authors applied their expertise to define the relationships between environmental factors and BGI applicability. The temperature, humidity, and precipitation nodes were trained with obtained data from the 24 districts. The LCZ node CPT was calculated as explained previously. The cost and maintenance nodes were trained based on literature information. The CPTs of the BGI and applicability were calculated, details are summarized in SI-A. Validation of the CPTs involved scenario testing and comparison with case studies to ensure their accuracy and reliability. This process included iterative consultations among the authors to refine the CPTs and align them with real-world conditions.

The model validation

In developing the BBN model for assessing the suitability of BGI solutions in Istanbul, we acknowledge the importance of validation to ensure its predictive accuracy and reliability. Given the potential use of this model in urban environmental decision-making, we have undertaken the following steps to validate the model.

We sourced meteorological and land use data from governmental and up-to-date databases. The model was trained using historical data from the Meteorological Data Information Presentation and Sales System associated with the Meteorological General Directorate of Türkiye and verified against recent environmental events specific to Istanbul. This approach helped in ascertaining the representativeness and accuracy of the data fed into the model.

The model was subjected to a series of tests involving hypothetical scenarios reflective of Istanbul’s diverse urban and climatic conditions. These tests were designed to evaluate the model’s ability to make accurate predictions under varying environmental parameters.

A sensitivity analysis was performed to identify the influence of each variable on the model’s outcomes. This analysis helped in understanding the robustness of the model against variations in input data and in identifying key drivers in the decision-making process.

Rapid urbanization and changes in land use patterns can outpace data collection efforts. Istanbul, a city with dynamic growth, often encounters this issue. Accurate, up-to-date information on land use, urban density, and existing green spaces is crucial for identifying areas where BGI can be most effective. Recognizing the challenges posed by data availability, we explored alternative approaches for data collection, including the use of GIS to supplement gaps in land use data in Istanbul. This proactive approach allowed us to enhance the model’s robustness and adaptability to diverse data scenarios. Due to a limited number of BGI applications in Istanbul, we could not gather any implementation data to compare. However, the developed model can be trained with more implementation data in the future.

In summary, these validation steps were critical in enhancing the model’s fidelity to real-world data, thereby improving its applicability and credibility in the urban planning context of Istanbul.

Validation

Typically, model validation is conducted using GeNIe’s internal validation tool. However, due to the CPTs relying on assumptions and expert input rather than actual data, manual validation involves creating diverse scenarios to observe their impact on the model. For this purpose, two scenarios with different conditions were examined. The first scenario had the parameters set to represent the Fatih district, while the second scenario represented the Büyükçekmece district. The conditions can be seen in Table 2, and high applicability was selected to determine the most suitable BGI under the presented conditions.

Table 2. The validation scenario summary

In the initial scenario, results (Figure 3), green roofs emerge as the most fitting BGI with a 19% probability, closely followed by urban gardens at 16%. These outcomes align with their effectiveness in lowering temperatures and elevating humidity levels, which are the desired effects in a densely urban district facing specific meteorological conditions. In the second scenario, urban gardens and parks exhibit the highest suitability among BGI options, standing at a 19% probability, followed by rain gardens at 16%. These findings correlate with their effectiveness in reducing temperature and moderating humidity levels, specifically tailored for districts abundant in open green spaces and experiencing similar meteorological conditions.

Figure 3. Blue–green infrastructure applicability Bayesian belief network (BGIA-BBN) model a) scenario 1, b) scenario 2.

Sensitivity analysis

To assess model sensitivity, one approach involves altering the bin boundaries representing the data of parent nodes. The chosen method is distributing data entries equally into three bins. This process is applied to each parent node; average temperature, humidity, and maximum precipitation. Despite changes in the parameter state values, no impact is observed on the BGI and applicability nodes, as illustrated in the original model (SI-B). Consequently, this implies that the climate parameters do not have a great impact on the applicability of the solution directly, rather they directly affect the selection of the BGI solution, however, as the parameter values in the original and the modified models are very close, they did not affect the BGI node.

Another method for sensitivity analysis is stress testing, which involves the examination of model changes under extreme conditions or boundaries. The nodes for temperature, humidity, precipitation, and cost and maintenance were initially set to low and then high conditions, and the applicability node’s outcome was observed.

In the low condition, the applicability results indicate 54% for low, 16% for medium, and 30% for high states. Conversely, in the high condition, the results show 54%, 14%, and 32% respectively. This suggests that BGI solutions have a higher probability of being applicable in extremely high levels of various meteorological parameters compared to low levels. This implies that BGIs are more effective in extreme weather conditions, even when cost and maintenance are considered to be in the expensive range.

GeNIe incorporates an internal sensitivity analysis using the algorithm as described by Druzdzel and Sowinski (Reference Druzdzel and Sowinski2023). This algorithm efficiently calculates derivatives of posterior probability distributions for a set of target nodes in a Bayesian network. The aim is to unveil the sensitivity of the network’s numerical parameters concerning the accuracy of the target node probabilities. A high derivative for a parameter indicates that a slight change in that parameter significantly influences the target probabilities, while a low derivative suggests that even substantial parameter alterations have minimal impact on the target probabilities.

In this analysis, the ‘applicability’ node was selected as the target node. Subsequently, the model’s coloring shifts to various shades of red, where brighter colors signify a higher impact on the target node, and vice versa. The results reveal that the LCZ node exerts the most significant effect on applicability, justified by the fact that LCZ is not dependent on any other nodes and lacks prior conditional probabilities. The following figure (Figure 4) depicts tornado diagrams illustrating the target node for each of its states. The bars display the range of variation in the target node as the parameters are altered. Parameters are arranged from the most sensitive to the least sensitive. In these diagrams, the red color signifies a negative change in the target node, whereas green indicates a positive change.

Figure 4. a) Tornado diagram for applicability node (state low), b) Tornado diagram for applicability node (state medium), c) Tornado diagram for applicability node (state high).