Monitoring the quality and quantity of urban green space in Amsterdam with remote sensing and truth ground measurements - A case study on the mitigation of air pollution and the urban heat island effect by urban parks d

(1)

Monitoring the quality and quantity of urban green space in

Amsterdam with remote sensing and ground truth

measurements

A case study on the mitigation of air pollution and the urban heat island effect by urban

parks during a heatwave in Amsterdam

Credits: _{42 EC} Supervisor: _{dr. ir. Nynke Schulp (IVM, VU)}

Period: _{01/12/2018 - 6/12/2019} Examinater: _{dr. ir. Emiel van Loon (IBED, UvA)}

Name: BSc. Florence van der Hoven Co-assessor: dr. Kenneth Rijsdijk (IBED, UvA)

ID: 10008047

Programme: _{MSc Earth Sciences} In collaboration with Green City Watch:

Track: _{Future Planet Ecosystem Science} Nadine Galle, Chris van Diemen, Jim Groot

Research institutes: IBED, IVM Date: January 19, 2020

(2)

Abstract

The continued development of cities will increasingly shape human well-being, considering more than two thirds of the global population is expected to live in cities by 2050. The consequences of urban densification endanger the liveability of cities, because it is associated with an increase in environmental threats like the urban heat island (UHI) effect and air pollution. Fortunately, urban green space can be used as a solution to mitigate the UHI effect and air pollution. Therefore this study focuses on developing a method to accurately monitor the quantity of urban green space in Amsterdam and assesses the quality of urban green space by exploring the ability of urban parks to lower temperatures and improve air quality. The monitoring method was developed by using remote sensing, very high-resolution (VHR) satellite imagery and machine learning. Three classification models (Decision Tree, Random Forest and XGBoost) were used to classify urban land cover in Amsterdam. All the models achieved an overall accuracy of 90%, but the Random Forest classifier achieved the highest overall accuracy with 99%. Correctly distinguishing urban green space from other land cover classes was also achieved, because green space was classified correctly in all cases by the Random Forest classifier. The quality was assessed with a case study that measured the difference in temperature and air quality between parks and a built-up area in Amsterdam. The results show that urban parks can function as urban cool islands during a heatwave. While the definition of a heatwave applied to the built-up area, it did not apply to the parks. Temperatures were between 3.8 °C and 5.8 °C lower in the parks than in the built-up area during the heatwave. Air pollution was worse during the heatwave than during average weather conditions, but was not worse in parks than in the built-up area. The traffic intensity of nearby roads most likely was the source of air pollution. The parks were not able to function as islands of clean air at the scale of this study.

Acknowledgements

This study was part of a master thesis at the University of Amsterdam in order to fulfil the requirements of the Future Planet Ecosystem Science track of the Earth Science master. The study was carried out in collaboration with Green City Watch, the Environmental Geography Institute of the Vrije Universiteit Amsterdam (IVM) and the Institute for Biodiversity and Ecosystem Dynamics at the University of Amsterdam (IBED). I would like to thank my supervisor Nynke Schulp for her advice and feedback on the quality of urban green space and ecosystem services throughout the process. I would like to thank my examinator Emiel van Loon for his advice and feedback on the modeling part of my thesis. I would like to thank both Emiel and Nynke for their encouragement throughout the process. And last but not least, I would like to give thanks to Chris, Jim and Nadine for teaching me new skills, letting me collaborate as one of their equals in their new start up, for their weekly support and encouragement and for teaching me that anything is possible.

(3)

1. Introduction

The continued development of cities will increasingly shape human well-being, considering the ongoing increase in urban population size. The percentage of the global population living in cities is expected to be 68% in 2050. A huge increase given the fact that in 1950 only 30%, and in 2018 only 55% of the global population lived in cities (United Nations, 2018). This increase results in urban expansion or when there is often no space for cities to expand in urban densification (Angel et al., 2011)_{. The need for sustainable development is becoming more important as migration to cities} continues and is therefore reflected in the 11th Sustainable Development Goal: sustainable cities and communities _{(United Nations, 2018a). To achieve this goal, investment in green public spaces,} public transport and improved urban planning is needed. Creating liveable cities that simultaneously inhabit two thirds of the global population will become one of the major challenges of this century. The increase in urban densification endangers the liveability of cities by causing threats to the environment and the well-being of urban citizens. Urban densification changes the physical characteristics of cities, resulting in more built-up areas and a loss of urban green space (Haaland and van den Bosch, 2015). Additionally, urban densification increases the environmental footprint. The International Energy Agency (IEA) estimated that cities currently account for approximately 67 percent of energy-related global greenhouse gases and will account for 74 percent by 2030 (Global Energy Assessment, 2012). The increase in emissions and the changes in physical characteristics can cause an increase in urban heat islands (UHI) and air pollution (Keeler et al., 2019). Urban heat islands are urban areas that are significantly warmer than the surrounding rural areas due to human activity. The loss of urban green space amplifies these negative consequences, because green space is able to mitigate these threats (Keeler et al., 2019). Additionally, climate change increases the impact of environmental threats. Extreme temperatures and droughts are intensifying the UHI effect and air pollution (Leichenko, 2011). Indirectly and directly the loss of urban green space threatens the living conditions and therefore the health of urban citizens (Wolch et al., 2014). Maintaining a sufficient amount of green space in cities is therefore essential to cope with health threats, environmental threats and climate change.

Increasing the amount of urban green space as a solution to the threats of urban densification is not always possible, due to the lack of available space in cities. Therefore it is crucial that the current amount of urban green space in cities is monitored and maintained. Smart solutions that increase the amount of urban green space without taking up additional space can play a role in achieving this. Some smart solutions that have been successfully implemented include the planting of trees, the transformation parks and open spaces, green roofs and walls, pocket parks, ecopassages and more (Keeler et al., 2019; Municipality of Amsterdam, 2011) _{. Nonetheless, the quantity of urban green} space decreases in cities. In Amsterdam, the amount of urban green space decreased with 11% between 2003 and 2016 _{(Giezen et al., 2018). An unexpected result, since the municipality of} Amsterdam aimed to increase the amount of urban green space (Municipality of Amsterdam, 2011). Giezen et al., (2018) showed that remote sensing can be used to quantify the amount of urban green space in cities. This study tests similar methodologies that can be used to quantify the amount of urban green space, but adds ground measurements to assess the quality of urban green space.

(6)

A work-around to increase the amount of urban green space is to improve the quality of existing urban green space. In this study, the quality of urban green space is interpreted as the ability of urban green space to mitigate environmental threats, such as the UHI effect and air pollution. The quality is measured by quantifying ecosystem services of urban green space. Ecosystem services are the benefits that humans obtain from ecosystems. Many of the ecosystem services provided by urban green space mitigate threats that are associated with urban densification and climate change. Ecosystem services include for example climate regulation and air purification(Derkzen et al., 2015). Moreover, urban green space can enhance physical and mental health and promote social and cultural well-being (Tzoulas et al., 2007). These services are particularly valuable in cities, where urban green space is often the only way people can experience nature. Improving the quality of urban green space can therefore increase a city’s ability to mitigate environmental threats and promote human well-being.

Urban ecosystem services like temperature regulation and air purification have become important in the Netherlands, because air pollution and the UHI effect are occurring more frequently due to climate change and urban densification (Heusinkveld et al., 2014). Both the UHI effect and air pollution have negative consequences for human health and therefore threaten the liveability of cities (Luber and McGeehin, 2008). As air pollution and the UHI effect often occur simultaneously and reinforce each other, the risks for human health become even worse (Fallmann et al., 2016; Sarrat et al., 2005). Fortunately, urban green space is able to mitigate the effects of UHIs air pollution by cooling and purifying the air (Harlan and Ruddell, 2011). Therefore urban green space can be used as a solution to battle air pollution and the UHI effect and to maintain the quality of life in cities.

(7)

2. Research aim and questions

This study consisted of two parts. In the first part methods to assess the quantity of urban green space were tested. In the second part a pilot study was done to measure the quality of urban green space.

The aim of the first part was to find an accurate classification method to classify urban land cover and distinguish urban green space from other land cover classes in order to quantify the percentage of urban green space in Amsterdam. Three classification models were used to classify urban land cover in Amsterdam (The Netherlands) with remote sensing. The first research question therefore was:

1. Which image classification method can most accurately classify urban land cover in Amsterdam?

In the second part, the ability of urban green space to mitigate air pollution and the UHI effect was explored by measuring the difference in air quality and temperature between urban green space and a nearby built-up area in Amsterdam. The second research question therefore was:

2. Is there a difference in temperature and air quality between urban green space in Amsterdam and a nearby built-up area?

(8)

3. Theoretical framework

This chapter provides more insight into assessing the quantity and quality of urban green space. The first section focuses on methods that can be used to quantify the amount of urban green space and monitor changes in urban land cover. In the second section the quality of urban green space in terms of ecosystem services that mitigate the UHI effect and air pollution is explored.

3.1 The quantity of urban green space

As discussed briefly in the introduction, urban expansion can cause a decrease in the amount of urban green space, while at the same time urban green space can be used to maintain and improve living conditions in cities. In Amsterdam, remote sensing was used to quantify the amount of urban green space within the inner ring road(Giezen et al., 2018). Results pointed out that the amount of urban green space decreased with 11% between 2003 and 2016, while the population increased by 7% _{(Giezen et al., 2018). The loss of urban green space was most likely the direct result of the} increasing population, because most of the lost green spaces were designated development areas. These changes were inconsistent with the long-term policy of the municipality of Amsterdam, which aimed to increase the amount of urban green space (Municipality of Amsterdam, 2011). Rapid population growth and the associated urban densification is challenging the preservation of urban green space in Amsterdam (Giezen et al., 2018). Therefore, monitoring urban green space in near real-time can be used to prevent unplanned losses of urban green space.

3.1.1 Remote sensing

The approach of Giezen et al. showed that remote sensing can be a valuable tool to quantify the amount of urban green space. Remote sensing can be used to detect and monitor physical characteristics of the earth from a great distance, typically by satellites. Nowadays, satellites are able to collect images at a resolution of less than a meter. With airborne photogrammetric frame-based cameras, very high-resolution (VHR) satellites collect panchromatic and spectral bands (Nikolakopoulos and Oikonomidis, 2015). Panchromatic bands supply high spatial resolution information, while spectral bands supply rich spectral information(Pohi and Van Genderen, 1998). VHR satellites like Quickbird, Worldview-2, Worldview-3, IKONOS and GeoEYE are able to collect multiple spectral bands at the same time, creating satellite images that contain extensive spectral information. With multi-spectral information, differences in land cover can be detected, because every land cover type has different radiometric features (i.e. different values for each spectral band). Calculating ratios between bands can reveal even more detailed information that leads to more specific differentiation between land cover types. For example, the normalized difference vegetation index can be used to distinguish vegetation from other land cover types and the normalized water index can be used to distinguish water from other land cover types (Gao, 1996; Tucker, C. J., 1979). With the multispectral information from high-resolution satellite imagery and the indices that can be calculated with it, accurate land cover classification methods can be developed. Using these provides a way to distinguish urban green space from built-up areas.

3.1.2 Machine learning

Manually extracting data from high-resolution satellite images is however very time consuming, because they contain huge bulks of data. Therefore, machine learning is used to evaluate the data and translate it into valuable information. Machine learning is an application of artificial intelligence

(9)

that uses algorithms and statistical models to learn and improve from experience without explicit instructions (Lary et al., 2016). In remote sensing it is for example used to classify land cover or extract specific features(Srivastava et al., 2017). With machine learning it is possible to easily classify urban land cover with the use of (high-resolution) satellite images.

3.1.3 Pixel-based approach

The dataset that is used to train an urban land cover classification model can be created with an object-based or pixel-based approach (Myint et al., 2011). The object-based approach uses groups of pixels that have similar properties as the training data samples (Jia, 2006). Object data is often pulled from an existing dataset that is not necessarily coming from the satellite image one is trying to classify. Therefore the characteristics of the training data samples and the testing data samples are likely to differ. On the other hand, the pixel-based approach uses single pixels as training data samples (Jia, 2006). Because the pixels are pulled from the same image that is going to be classified, the characteristics are the same and misclassifications are less likely to occur (Myint et al., 2011). With the pixel-based approach, radiometric characteristics are used to differentiate between land cover classes. Therefore it is possible to distinguish between general land cover types like trees, grass, water and buildings, but it is not possible to distinguish features like parks and sports fields. By using a pixel-based approach, it is possible to use the radiometric characteristics of urban land cover types to differentiate between them.

3.1.4 Expectations

To prevent unplanned loss of urban green space a monitoring method can be used to quantify urban green space in near real-time and advise urban planners. With a pixel-based approach and spectral information provided by high-resolution satellite images, a distinction between urban green space and other land cover types based on radiometric characteristics can be made. By using machine learning to train the classification models urban land cover in Amsterdam can be accurately classified.

3.2 The quality of urban green space

Increase the quality of existing urban green space can be used as an alternative or addition to increasing the amount of urban green space. Ecosystem services can be used to quantify the quality of urban green space. Ecosystem services of urban green space include climate regulation, air and water purification, storm and wastewater management, carbon sequestration, noise reduction, food production and providing a habitat for wildlife _{(Derkzen et al., 2015). Especially climate regulation} and air purification are ecosystem services that have become more important recently, because the densification of cities has led to an increase in air pollution and the UHI effect, which endangers the health of urban citizens (Harlan and Ruddell, 2011). This section focuses on the causes and risks of air pollution and the UHI effect and on the role of urban green space in mitigation.

3.2.1 Increase in heatwaves leads to urban heat islands and urban air pollution

Worldwide climate change is causing a rise in temperatures, more frequent and extreme heat events and the UHI effect (IPCC, 2014). These changes in climate are also associated with an increase in air pollution (Harlan and Ruddell, 2011). In the Netherlands, scenarios point out that warm days and heatwaves will occur more frequently in the future (van den Hurk et al., 2007). Next to the changes in climate, the characteristics of cities are changing. The amount of built-up area is increasing, while the amount of urban green space is decreasing (Giezen et al., 2018). These changes make cities more vulnerable to the UHI effect and air pollution (Harlan and Ruddell, 2011; Oke, 1982). The UHI effect

(10)

and air pollution interact with each other in several ways. Increased temperatures can promote the dispersal of pollutants through the increased mixing of air leading to higher pollutant concentrations (Sarrat et al., 2005). In turn, increased air pollutant levels can trap solar radiation leading to increased temperatures(Jin et al., 2011). Climate change and urban densification are leading to an increase in heatwaves, enabling the UHI effect and air pollution to occur more frequently and enforce the interaction between those.

3.2.2 Health risks and mortality

Both air pollution and the UHI effect are associated with severe risks to human health and well-being (Haines et al., 2006; Harlan and Ruddell, 2011). The UHI effect can cause overheating and air pollution can cause severe health risks like cardiovascular and respiratory diseases (Kampa and Castanas, 2008; Sena et al., 2014). Especially infants and elderly people are more vulnerable to these risks. Moreover, mortality rates are directly linked to air pollution and the UHI effect. A striking example is the heatwave that occurred between the 1st and 20th of August in 2003 in France, which caused 15,000 excess deaths (Fouillet et al., 2006). Notable results were also found in the Netherlands, where heatwaves caused an excess mortality between 1400 and 2200 people in 2003 and 2006 respectively(Hoek et al., 2000). The increased mortality during the heatwave in 2003 was in turn directly associated with an increase in air pollution (Fischer et al., 2004). Compared to an average summer there were between 400 and 600 more air pollution induced deaths. Because the UHI effect and air pollution are occurring more frequently due to the urban densification in the Netherlands, there is an urgent need for solutions that mitigate these threats.

3.2.3 Air pollution

Air pollution has gotten attention in the Netherlands since the associated health risks became known. The most important source of air pollution is traffic, which in particular causes problems in cities where the most people live. There are different components that cause air pollution, such as carbon dioxide, sulfur dioxide, nitrogen dioxide and particulate matter. Particulate matter includes all particles smaller than 10 μm, which are usually divided in PM10 (<10 μm) and PM2.5 (<2.5 μm). PM2.5 particles are thought to be more harmful, because smaller particles can penetrate further into the respiratory pathway (World Health Organisation, 2013). Because particulate matter can cause severe health problems, the World Health Organisation (WHO) gives advice on limiting PM concentrations (World Health Organisation, 2013). In the Netherlands, PM10 concentrations are allowed to reach 40 μg/m3 on average on a yearly basis. Next to that, the daily average can exceed 50 μg/m3 on maximum 35 days per year (RIVM, 2018). The WHO advises to keep PM10 concentrations below 20 μg/m3 and limit the number of days that PM10 concentrations exceed 50 μg/m3 to 3 (World Health Organisation, 2013). In the Netherlands, PM2.5 concentrations need to stay below an average concentration of 25 μg/m3 each year (RIVM, 2018). No limit is set for the number of days PM2.5 concentrations can exceed dangerous limits. The WHO advises to keep PM2.5 concentrations below 10 μg/m3 and limit the number of days that PM2.5 concentrations exceed 25 μg/m3 to 3 days (World Health Organisation, 2013). On average, PM concentrations stay within the boundaries set by the Dutch government. However, PM concentrations are still exceeded locally, mainly next to busy roads and in cities (RIVM, 2018). In 2018, all the measurement stations in Amsterdam exceeded the limit set by the WHO for PM2.5 and PM10 concentrations (GGD Amsterdam, 2019). On average PM2.5 and PM10 concentrations have significantly decreased over a longer period, but the downward trend has weakened over the last couple of years. Therefore, further solutions to reduce PM concentrations are needed to meet the advice of the WHO.

(11)

3.2.4 Urban green space as mitigation strategy

Previous sections showed there is a need for solutions that mitigate the effects of UHIs and air pollution in cities. Because vegetation has multiple characteristics that have a cooling and purifying effect on the atmosphere, urban green space can be used to mitigate the negative consequences of UHIs and air pollution(Bowler et al., 2010; Derkzen et al., 2015). Vegetation is able to purify the air by taking up air pollutants and transform them to biomass and oxygen (Mudd, 2012). When pollutant concentrations are high, green space takes up more pollutants than when levels are low, which shows that increasing the amount of green space near the source of emissions in cities could therefore benefit air quality (Derkzen et al., 2015). A key characteristic in temperature regulation by plants is evapotranspiration; the process by which absorbed water is transferred to the atmosphere (Oke, 1982). Evapotranspiration has a cooling effect on the surrounding air. Built-up areas and impervious surfaces lack the ability to cool their surroundings. Additionally, the type of vegetation plays a role in air movement and heat exchange (Bonan, 1997). For example, warm air can be preserved under tree canopy cover, while open grassland has low resistance to airflow which increases cooling by convection. Therefore the type of vegetation also plays a role in temperature regulation. Vegetation and impervious surfaces like built-up areas have a different influence on temperature and air quality. Therefore, different land cover types should be taken into account when studying temperature regulation and air quality. This study focuses on differences in temperature and air quality between built-up areas and urban green space.

3.2.5 UHIs in the Netherlands

In the Netherlands, climate regulation in cities only recently became the focus of researchers, because the UHI effect did not occur before. The Netherlands has a mild mid latitude climate (Köppen classification: Cfb), with mild summers and mild winters. Now that the number of heatwaves and warm days is increasing, there is an emerging need for implementing solutions. For several researchers urban green space as a solution for mitigating the UHI effect became a point of interest. In Rotterdam, spatial variability of the UHI was found to be influenced by urban land cover (Heusinkveld et al., 2014). During midday, city parks were 4.0 °C cooler than built-up areas, while the surrounding rural area was 1.2 °C cooler than built-up areas in cities. In Amsterdam, shading, water and grass reduced the air temperature by 1 °C during the day (Klok et al., 2019). The physiological equivalent temperature (PET), i.e. the temperature as perceived by humans, on the other hand differed with 12 to 22 °C between shaded and sunlit areas. These findings show that temperatures were lower near vegetation, shade and in parks in cities in the Netherlands.

3.2.6 Expectations

Air pollution and UHIs are occurring more frequently in the Netherlands, while they are associated with increased health risks and mortality. Therefore there is a need for solutions to mitigate these threats and ensure the health of urban citizens. Multiple studies have shown that urban green space is able to cool and purify urban air and can therefore be used as a solution. This study assesses the ability of urban parks to mitigate the effect of air pollution and UHIs by comparing temperature and PM concentrations between urban parks and a built-up area. It is expected that temperature and PM concentrations are lower in urban parks, because vegetation contains multiple characteristics that are able to cool and purify the air.

(12)

4. Materials and methods

4.1 Urban land cover classification

The aim was to build a suitable classification model that can accurately classify the urban land cover of Amsterdam (The Netherlands) using high-resolution satellite data and machine learning. Urban land cover was classified in the following categories: trees, grass, water, streets and buildings. To predict urban land cover, Decision Tree, Random Forest and XGBoost were used as machine learning classification models. The aim was to achieve an overall accuracy of 90% with the models.

4.1.1 Study area

Amsterdam (The Netherlands) was selected as a study area to test three classification methods. Amsterdam is the largest and capital city of the Netherlands, has over 850 000 inhabitants (almost 2.5 million in the metropole region) and covers almost 220 km2 ( Table 1). The population density is 5150 residents/km2 and the housing density is 2615 houses/km2 (OIS, 2018). The area of interest can be found in Fig. 1.

Table 1 Demographic data of Amsterdam on 1 January 2018 (CBS, 2018) Municipality Population (#) Surface area

(km2) Population density (citizens/km2) Housing density (houses/km2) Amsterdam 854 047 219.49 5 160 2 615 4.1.2 Data sources

A Worldview-2 image was used as high-resolution satellite data input. Worldview sensors provide a high resolution panchromatic band and eight multispectral bands: coastal, blue, green, yellow, red, red edge, near-infrared 1, and near-infrared 2. This highly detailed information was used to predict urban land cover. Because Worldview sensors are able to collect very large areas of multispectral imagery in a single pass, it was possible to use one image to classify the central city of Amsterdam. The selected Worldview-2 image was taken on the 5th of August in 2017. The requirements for the image included that it was taken during summer and that the image contained 0% cloud cover. The image had to be taken during the summer to be able to distinguish green space and had to contain 0% cloud cover to reduce noise. Images from 2018 were not taken into account, because there was a severe drought during this summer and most of the green space was not green anymore, but brown. The characteristics of the image used in this study can be found in _{Table 2.}

Table 2 Image characteristics: source, date, percentage cloud cover, resolution in meters and angle, where a 0 degree angle would be looking straight down.

ID Source Date Cloud cover Resolution Angle

(13)

Figure 1 Satellite image that was used to classify urban land cover in Amsterdam. The green square represents the area of interest.

4.1.3 Training and testing dataset

The dataset was created with an interactive script that could be used to select an independent pixel. First the image was divided into 100 squares of the same size. Within each square 2 pixels per class were selected, creating a training dataset that contained 1000 pixels in total and 200 pixels per class. Duplicates and pixels that were assigned to the wrong class were removed. Trees that were assigned to the wrong class were removed by selecting the pixels within the “trees” class that had an NDVI below 5. After the removal of duplicates and pixels that were assigned to the wrong class, the dataset contained 904 pixels. The pixel-based dataset was divided into training (75%, i.e. 681 pixels) and testing data (25%, i.e. 223 pixels) by taking a random subset.

4.1.4 Classification models

Three classification models were tested to classify urban land cover in Amsterdam. Urban land cover was classified into five classes: trees, grass, water, streets and buildings. Spectral band data and vegetation and water indices were used to predict the urban land cover class. Three classification models were tested: Decision Tree (DT), Random Forest (RF) and XGBoost (XGB). Urban land cover class was used as the response variable. Spectral band data of the eight Worldview-2 bands (coastal, blue, green, yellow, red, red edge, NIR1, NIR2) and the normalized difference vegetation index

(14)

(NDVI) and the normalized difference water index (NDWI) were used as explanatory variables. The aim was to achieve 90% accuracy after cross validation. The model that achieved the highest accuraacy was used to predict the land cover class of the complete image.

The Decision Tree classifier subdivides the dataset by testing the dataset at each node or branch of the tree (Fig. 2). The tree consists of one root node (containing the entire dataset), internal nodes called splits and terminal nodes called leaves. Decision Trees have

several advantages over traditional classification models; they are nonparametric and do not require specific input distributions, they are able to handle categorical and numeric data, they can handle non-linear relations between features and classes and they can handle missing values _{(Friedl and Brodley, 1997).}

The Random Forest classifier takes the average of multiple decision trees. Individual decision trees can have a high variance based on the tests they are based on. Averaging them increases the accuracy. Each decision tree in a Random Forest considers a random set of features and an individual random set of training data points. The random

selection makes the model robust to overfitting. To conclude, a Random Forest classifier has all the advantages that a Decision Tree classifier has and even more.

XGBoost is an ensemble learning method, which means it is able to combine multiple models called base learners. Bagging and boosting are examples of ensemble learners. They are mostly used with decision trees. Bagging takes the average of multiple decision trees that have been fed with data sampled with replacement. Random Forest is a bagging method. Boosting does not build parallel trees but builds trees sequentially. Each tree will perform slightly better than the previous tree reducing the errors. In contrast to bagging, where trees are built to their maximum extent, boosting builds trees with fewer splits. Advantages of XGBoost are that it is fast. Disadvantages are that XGBoost can be sensitive to overfitting and hard to tune.

The NDVI and NDWI were used to distinguish vegetation (e.g. trees, grass and shrubs) from build up area and bare land(Baret and Guyot, 1991). The NDVI is calculated with the difference between near infrared light (which is reflected by vegetation) and red light (which is absorbed by vegetation) as shown in equation (1). The NDVI ranges from -1 to 1, where -1 represents water, 0 represents little or no vegetation and 1 represents dense vegetation (Tucker, 1979).

DV I

N = _{(NIR + Red)}(NIR − Red) (1)

Especially in cities, where different land cover types occur close together with relatively small surface areas, the NDVI can be useful to differentiate between the different green land cover types. NDWI can calculate water content more accurately than the NDVI. The NDWI measures the liquid water content of vegetation and is calculated with the difference between near infrared light (which is reflected by vegetation) and short-wave infrared light (which is absorbed by vegetation) as shown in equation (2). The NDWI ranges from -1 to 1, where values between -1 and 0 represent no vegetation or water content and values between 0 and 1 represent vegetation with increasing water content (Gao, 1996). Both indices where used to differentiate between different vegetation types.

DW I

(15)

4.1.5 Parameter tuning and cross validation

The accuracy that the models achieve depends on how the training and testing set is sampled. For example, when the Decision Tree model is run once it achieves an accuracy of 85%. When the Decision Tree model is run 1000 times the accuracy ranges from 80% to 100% accuracy. It is a problem that the model performs differently depending on the subset of data it is trained on. This phenomenon is known as overfitting: the model is learning to classify the training set so well that it does not generalize and perform well on data it has not seen before.

To validate the accuracy of the model and prevent overfitting k-fold cross-validation was used. The original data were split into k subsets, where one of the subsets was used as the testing set and the other subsets were used as training sets. This process was repeated until each subset was used as the testing set exactly once. Finally, the average was taken of the outcomes. In this study the process was repeated 10 times, i.e. 10-fold cross-validation was used.

Every Machine Learning model comes with a variety of parameters to tune, and these parameters can be vitally important to the performance of the classification model. For example, the accuracy of the Decision Tree model drops tremendously when the depth is severely limited. To find the best parameter settings for the classification models Grid Search was used. With Grid Search the best-performing parameter combination was selected from a range of parameters ( Table 3). The parameter settings for the models were derived from (scikit-learn developers, 2019a, 2019b; XGBoost developers, 2019).

Table 3 Parameters that were used for parameter tuning of each classification model. Between brackets you find the possible settings.

Decision Tree Random Forest XGBoost

Criterion (gini, entropy) n Estimators (10, 25, 50, 100) n Estimators (10, 25, 50, 100) Splitter (best, random) Criterion (gini, entropy) Criterion (gini, entropy) Maximum Depth (1, 2, 3, 4, 5) Maximum Features (1, 2, 3, 4) Maximum Features (1, 2, 3, 4) Maximum Features (1, 2, 3, 4)

The criterion determines the quality of a split. Possible criteria are gini for the Gini impurity and entropy for information gain. The Gini Impurity is the probability of incorrectly classifying a randomly chosen element in the dataset if it was randomly labeled according to the class distribution in the dataset. The splitter represents the strategy that is used to choose the split at each node. There are two options where best chooses the best split and random chooses the best random split. The maximum depth represents the depth of the tree. If this parameter is set to _none the nodes are expanded until all leaves are pure. The maximum features parameter determines the number of features that is considered when looking for the best split. The number of estimators stands for the number of trees. The search for the best range of parameters was repeated 10 times. The best set of parameters discovered by then was chosen to be used in the final classification model.

(16)

4.1.6 Confusion matrix and accuracy

Even with the best set of parameters it is unlikely that the model achieves 100% accuracy. In other words, misclassification will occur. To find out where misclassifications occurred exactly, the confusion matrix was calculated. In a confusion matrix the predicted values are compared to the actual values. The result is an overview of all the true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN) (Table 4). The diagonal in the matrix contains all the correctly classified values (TP/TN). The confusion matrix is also used to find out which classes are confused for each other. With this information it is possible to finetune the choice of predictors.

Table 4 The confusion matrix with true positives (TP), true negatives (TN), false negatives (FN) and false positives (FP). Columns represent the reference values and rows represent the predicted values.

Reference values

Building Grass Street Tree Water Row total

Predicted values Building TP/TN FN FN FN FN Grass FP TP/TN FN FN FN Street FP FP TP/TN FN FN Tree FP FP FP TP/TN FN Water FP FP FP FP TP/TN Column total TOTAL

The overall accuracy is calculated by dividing the sum of the true positives by the total number of pixels in the confusion matrix ( Equation 3). Additionally the accuracy of individual categories can be calculated; the producer’s and user’s accuracy. The producer’s accuracy is calculated by dividing the true positives of one column by the column total ( Equation 4). The producer accuracy indicates the probability of a reference pixel being correctly classified (Congalton, 1991). The user’s accuracy is calculated by dividing the true positives of one row by the row total ( Equation 5). The user accuracy indicates the probability that a pixel actually represents the predicted land cover class (Congalton, 1991).

verall accuracy

O = T rue positives total_{T otal} (3)

roducer s accuracy

P ′ = Column true positives_{Column total} (4)

ser s accuracy

U ′ = Row true positives_{Row total} (5)

4.1.7 Software

The analysis was carried out in Python 3 in cloud-based Jupyter notebooks on the online GBDX platform. The platform contained Python 3 packages that were specialized in performing calculations on Worldview images. The packages that were used included the GBDX toolbox, sklearn, pandas, numpy, seaborn, xgboost. The machine learning notebook byOlson (2019) was used to create the classification script.

(17)

4.2 The influence of urban parks on temperature and air quality

The role of urban green space in regulating temperature and air quality in cities in the Netherlands is becoming more important. The effects of climate change are noticeable, since warm days and heatwaves are occurring more frequently (Heusinkveld et al., 2014). The Netherlands is characterized by a high population density, dense cities, extensive infrastructure and limited nature, which results more easily in UHIs and air pollution (Steeneveld et al., 2011). Therefore, a pilot study was conducted to explore the influence of urban green space on temperature and air quality by measuring the difference in temperature and air quality between urban parks and a nearby area in Amsterdam-Noord.

4.2.1 Study locations

Measurements were taken on three locations in Amsterdam-Noord ( Fig. 3). The locations were chosen based on the surrounding land cover as seen on satellite images and based on traffic intensity. The locations were different from each other, but still representative for Amsterdam. One location was in a built-up environment next to a road and two locations were in the middle of a park. The built-up location was next to the Johan van Hasseltweg, which is a busy road where approximately 23000 cars pass by each day (Fig. A4.1) (Gemeente Amsterdam, 2016). The other two measurement locations were located in the middle of Noorderpark and Vliegenbos. Vliegenbos mainly consists of trees and water, while Noorderpark has a varied landscape with a mixture of trees, grass and water (Fig. 5).

Figure 3 On the left the area of interest in Amsterdam is shown. On the right, the locations of the sensor units in Amsterdam-Noord are shown. From left to right: Johan van Hasseltweg, Noorderpark and Vliegenbos.

(18)

4.2.2 Measurements

The measurement equipment was manufactured by the Electronics Lab of the Vrije Universiteit in Amsterdam. Sensors were built into a plastic box to protect them from the outside environment. The specifics of the sensors can be found in Table 5. Measurements were sent every 10 minutes via the KPN LoRa (Long Range) network to the server. Data were also stored on a SD card to make sure the data would not be lost in case of network failure. Noise was measured once per minute for 10 seconds. This resulted in an average value and a peak value measured during 10 minutes. The PM sensor switched on after the sound measurements were completed. It absorbed air for 30 seconds by means of a fan and then the measurement was taken. The 30 second suction was necessary to get the air through the sensor and be able to get a reliable measurement.

Table 5 Characteristics of the sensors

Brand Type Variable Accuracy

PLANTOWER PMS5003 PM1, PM2.5, PM10 1 μg/m3

Microchip MCP9700AT Temperature ±2°C (max.)

Sensors were placed in the shade surrounded by bushes and trees to minimize the effect of sun and wind. Sensor units were attached to trees with tierips, a chain and a lock ( Fig. 4). The units were placed between 1 and 2 meters above ground level, with the sensor outlets facing down to protect them from possible rainfall. The units were left at the study locations from the 10th of July until the 17th of July and from the 21st of August until the 28th of August. After each measurement period, the data were downloaded from the SD-cards.

Figure 4 Placement of sensor units. From left to right: Johan van Hasseltweg, Noorderpark and Vliegenbos.

4.2.3 Analysis

First, the raw data was extracted from the SD cards of the sensor units. Then, time series of all the temperature and PM concentrations were visualized using the raw data. After that, averages and standard errors were calculated per part of the day. Days were divided into four parts: night (12-6AM), morning (6AM-12PM), afternoon (12-6PM) and evening (6PM-12AM). Then, the maximum daily temperature and PM concentrations were calculated to check whether the limits set by the Dutch government and the WHO were exceeded. Finally, the data was compared to weather and air quality data of the KNMI and RIVM. All calculations and visualizations were done in Microsoft Excel.

(19)

5. Results

5.1 Urban land cover classification

Three classification models were tested to classify urban land cover in Amsterdam (The Netherlands) using machine learning and high-resolution satellite data. The aim was to find an image classification model that achieves an accuracy of at least 90%. Land cover was divided into five classes: trees, grass, water, streets and buildings. Three different classification models were tested: Decision Tree (DT), Random Forest (RF) and XGBoost (XGB). Land cover was used as the response variable. Spectral band data and indices were used as explanatory variables. The results of the three classification models are discussed in the following section.

5.1.1 Decision Tree

The Decision Tree model performed worse than the other two models but still achieved an overall accuracy of 96% after cross-validation and parameter tuning. Parameter tuning led to the highest accuracy whencriterion was set to entropy, splitter was set to best, maximum depth was set to 5 and maximum features was set to 4. The confusion matrix revealed that 9 out of 223 values were misclassified. In 7 cases street was mistaken for building, in one case grass was mistaken for tree and in one case building was mistaken for water (_{Table 6). The producer’s and user’s accuracies are given} below.

Producer’s accuracy: User’s accuracy:

Building = 39 / 41 = 95% Building = 39 / 46 = 85%

Grass = 48 / 49 = 98% Grass = 48 / 48 = 100%

Street = 47 / 54 = 87% Street = 47 / 47 = 100%

Tree = 48 / 48 = 100% Tree = 48 / 49 = 98%

Water = 32 / 32 = 100% Water = 32 / 33 = 97%

Table 6 Confusion matrix of Decision Tree classifier. Rows represent the predicted values and columns represent the actual values.

Building 39 0 7 0 0 46 Grass 0 48 0 0 0 48 Street 0 0 47 0 0 47 Tree 0 1 0 48 0 49 Water 1 0 0 0 32 33 Column total 41 49 54 48 32 223

(20)

5.1.2 XGBoost

The XGBoost classifier achieved an overall accuracy of 92% after cross-validation and parameter tuning. The highest accuracy was achieved after parameter tuning when criterion was set to gini, maximum depth was set to 3, maximum features were set to 1 and the number of estimators was set to 100. Misclassification occurred in 17 of the 223 cases. Building was mistaken for street in three cases and for water in one case. Street was mistaken for building in 4 cases and for grass in two cases. Tree was mistaken for grass in two cases. Water was mistaken for building in three cases, for grass in one case and for street in one case. Grass was classified correctly in all cases ( Table 7). The producer’s and user’s accuracies are given below.

Grass = 45 / 45 = 100% Grass = 45 / 50 = 90%

Street = 51 / 57 = 89% Street = 51 / 55 = 93%

Tree = 36 / 38 = 95% Tree = 36 / 36 = 100%

Water = 40 / 45 = 89% Water = 40 / 41 = 98%

Table 7 Confusion matrix of XGBoost classifier. Rows represent the predicted values and columns represent the actual values.

Building 34 0 4 0 3 41 Grass 0 45 2 2 1 50 Street 3 0 51 0 1 55 Tree 0 0 0 36 0 36 Water 1 0 0 0 40 41 Column total 38 45 57 38 45 223 5.1.3 Random Forest

Of the three classification models tested, the Random Forest classifier performed best with an overall accuracy of 99%. Parameter tuning led to the best score when the _{number of estimators was} set to 50, the criterion was set to gini and the maximum features was set to 4. According to the confusion matrix misclassification occurred in 3 of the 223 cases. Once building was mistaken for street, once street was mistaken for building and once grass was mistaken for street. The classes tree and water were classified correctly in all cases ( _{Table 8). The producer’s and user’s accuracies} are given below.

Grass = 47 / 48 = 98% Grass = 47 / 47 = 100%

Street = 45/ 46 = 98% Street = 45/ 47 = 96%

Tree = 38 / 38 = 100% Tree = 38 / 38 = 100%

(21)

Table 8 Confusion matrix of Random Forest classifier. Rows represent the predicted values and columns represent the actual values.

Building 59 0 1 0 0 60 Grass 0 47 0 0 0 47 Street 1 1 45 0 0 47 Tree 0 0 0 38 0 38 Water 0 0 0 0 31 31 Column total 60 48 46 38 31 223

The Random Forest classifier was used to classify the entire satellite image of Amsterdam ( Fig. 5). The original image can be found in Figure 6. Buildings are shown in red, streets in beige, water in blue, trees in dark green and grass in light green. Parks can be clearly distinguished from the surrounding city. Some unexpected features and classifications can be identified. Railways are classified as buildings. The red features in the water are boats, not buildings. The red features on roads are cars. Athletic courts and bicycle paths are classified as buildings, because they are red like the roofs of houses. Not all the waterways are visible, because of overhanging trees. The shadow of buildings is sometimes classified as water. Above Westerpark (visible in the top left corner of Figure

5) a large area has been classified as streets, while it actually consists mostly of trees. This misclassification probably occurred because of cloud cover ( Fig. 6). The classes were represented with 18.11% trees, 19.83% grass, 20.21% water, 23.40% building and 18.44% street.

Figure 5 Result of image classification using a Random Forest classifier. Trees are shown in dark green, grass is shown in light green, water is shown in blue, buildings are shown in red and streets are shown in beige.

(22)

(23)

5.2 The influence of urban parks on temperature and air quality

Differences in temperature and air quality between parks in Amsterdam and a nearby built-up area were explored with a pilot study. Sensor units that measured temperature (°C) and PM1, PM2.5 and PM10 (μg/m3) were placed next to a road (Johan van Hasseltweg) and in the middle of two parks in Amsterdam (Noorderpark and Vliegenbos) (_{Fig. 3). The units were placed in the shadow surrounded} by trees and bushes to minimize the effect of sun and wind. Measurements were taken between the 10th and the 18th of July and between the 21st and the 29th of August. In the following sections the results of the measurements are interpreted and compared to KNMI data from Schiphol and air quality data from Luchtmeetnet.

5.2.1 Temperature

During both measurement periods there was a clear trend showing the difference in temperature between night and day temperatures (Fig. 7 and Fig. 9). The results of the measurements in July were very different from the results of the measurements in August, because there was a heatwave in August. A heatwave in the Netherlands is described as five consecutive summer days (maximum temperature of 25.0 °C or higher), of which minimal three days are tropical (maximum temperature is 30 °C or higher) (KNMI, 2019). According to this definition, a heatwave occurred at Johan van Hasseltweg and at Schiphol, but not in Noorderpark and Vliegenbos. At Johan van Hasseltweg six consecutive summer days with four tropical days occurred, at Schiphol six consecutive summer days with three tropical days occurred and in Noorderpark and Vliegenbos four consecutive summer days without tropical days occurred (Table 9). During the heatwave temperatures were between 3.8 °C and 5.8 °C lower in the parks than at Johan van Hasseltweg (Figure 7).

Table 9 Daily average and maximum temperatures in °C measured at Johan van Hasseltweg and in Noorderpark and Vliegenbos between the 22nd and 28th of August. Tropical days (maximum temperature of 30 °C or higher) are shown in orange and summer days (maximum temperature of 25 °C or higher) are shown in yellow. *Maximum temperatures measured between 12 AM and 6 PM. Schiphol data is derived from KNMI (Figure A1.1).

Johan van Hasseltweg Noorderpark Vliegenbos Schiphol

Average Maximum Average Maximum Average Maximum Average Maximum

22-Aug 19.0 24.1 17.4 21.5 17.8 21.8 17 22.6 23-Aug 20.5 26.6 18.6 22.8 18.5 22.5 18.8 25.4 24-Aug 22.4 30.8 20.5 25.7 20.5 25.0 21.6 28.9 25-Aug 23.9 31.8 22.0 27.9 21.8 26.6 22.9 31.1 26-Aug 24.6 31.5 22.9 27.6 22.8 27.0 24.2 30 27-Aug 26.1 33.4 24.2 29.5 23.8 28.3 25.1 32.3 28-Aug 25.7* 24.1* 24.4* 21.7 25.4

(24)

Figure 7 Daily difference in maximum temperature between Johan van Hasseltweg and Noorderpark and between Johan van Hasseltweg and Vliegenbos between the 22nd and 28th of August.

In August, temperatures differed between the locations. During the day, the temperature at the Johan van Hasseltweg was higher than in Noorderpark and Vliegenbos ( Fig. 8). The temperature in Noorderpark seems higher during the day than in Vliegenbos, but temperatures did not differ between Noorderpark and Vliegenbos (Fig. 8 and Fig. 9). The average temperature per part of the day (night: 12AM - 6AM, morning: 6AM - 12PM, afternoon: 12PM - 6PM, evening: 6PM - 12 PM) differed between the parks and the built-up area during the afternoon and evening (Fig. 9). At the Johan van Hasseltweg temperatures were higher than in Noorderpark and Vliegenbos. During the night and morning, temperatures between the locations did not differ, except on the 22nd and 28th of August (Fig. 9). During the morning of the 22nd, temperatures were already higher at the Johan van Hasseltweg than in the parks. During the night and morning of the 28th, the same results were found.

Figure 8 Temperature [°C] at the Johan van Hasseltweg (red), Noorderpark (light green) and Vliegenbos (dark green) measured between the 21st of August and the 29th of August.

(25)

Figure 9 Average temperature in Vliegenbos, Noorderpark and Johan van Hasseltweg during night (12AM - 6AM), morning (6AM - 12PM), afternoon (12PM - 6PM) and evening (6PM - 12AM) between the 21st and the 29th of August. In Johan van Hasseltweg and Noorderpark 35 measurements were taken during each part of the day. In Vliegenbos, 20 measurements were taken during each part of the day. Error bars represent ±1SE.

In July, the differences in temperature between the locations were smaller than in August ( Fig. 10). Still, the average temperature per part of the day during was often higher at Johan van Hasseltweg than in Noorderpark and Vliegenbos (Fig. 11). On the 17th of July the difference between the minimum and maximum temperature was very high at all locations during the afternoon. At the Johan van Hasseltweg the temperature rose from 17.6 to 24.4 °C, in Noorderpark from 16.0 to 21.2 °C and in Vliegenbos from 16.3 to 22.1 °C. Because the temperature changed a lot during a short period of time, the standard error is much larger than on the other days ( Fig. 11). The total minimum and maximum temperature of the 17th of July were 13.1 and 24.7 °C at Johan van Hasseltweg, 12.8 and 22.1 °C in Noorderpark, 13.1 and 22.1 °C in Vliegenbos and 10.4 and 24.6 °C at Schiphol (Fig. S1).

Figure 10 Temperature [°C] at the Johan van Hasseltweg (red), Noorderpark (light green) and Vliegenbos (dark green) measured between the 10th and the 18th of July.

(26)

Figure 11 Average temperature in Vliegenbos (light green), Noorderpark (dark green) and Johan van Hasseltweg (red) during night (12AM - 6AM), morning (6AM - 12PM), afternoon (12PM - 6PM) and evening (6PM - 12AM) between the 10th and the 18th of July. In Johan van Hasseltweg and Noorderpark 35 measurements were taken during each part of the day. In Vliegenbos, 20 measurements were taken during each part of the day. Error bars represent ±1SE.

5.2.2 Air quality

Air quality was measured in PM1, PM2.5 and PM10 concentrations between Wednesday the 10th and Thursday the 18th of July and Wednesday the 21st and Thursday the 29th of August in Noorderpark, Vliegenbos and at the Johan van Hasseltweg ( Fig. 3). The 13th and 14th of July and the 24th and 25th of August were weekend days, which usually have a lower traffic intensity and therefore better air quality. PM1, PM2.5 and PM10 concentrations followed the same trends, i.e. maxima and minima followed the same trend for PM1, PM2.5 and PM10 (_{Fig. 12 and Fig. 14). PM10} concentrations are highest, followed by PM2.5 and PM1 concentrations. The closest measurement station of Luchtmeetnet that measures PM2.5 and PM10 is located at the Van Diemenstraat ( _Fig.

S2).

In August, In July, the average daily PM10 concentrations exceeded the limit of 50 μg/m3 on the 27th of August at all the measurement locations (Table 10). On the 26th, the PM10 limit was exceeded at the Johan van Hasseltweg and in Noorderpark. PM2.5 concentrations exceeded the limit of 25 μg/m3 on the 25th, 26th and 27th of August at all measurement locations ( Table 10). At the Van Diemenstraat the PM2.5 limit was exceeded on the 26th and 27th of August. Daily fluctuations in PM concentrations were more distinct than in July ( Fig. 12). Maxima are mostly found during mornings, while minima are mostly found during the evening (Fig. 12). The highest PM concentrations were measured in Noorderpark (Table A3.1). The maximum PM concentrations measured at the Van Diemenstraat did not differ much from the maximum PM concentrations measured at the study locations (Table A3.1). The average PM2.5 concentrations did exceed the allowed yearly average limit of 25 μg/m3 at all study locations. PM10 concentrations did not exceed the limit of 40 25 μg/m3.

(27)

There were some differences between the different locations. On the morning of the 24th of August, PM1 and PM2.5 concentrations were higher in Noorderpark then in Vliegenbos and at the Johan van Hasseltweg (Fig. 13). In the night and during the morning of the 25th of August, there were differences between all the locations for al PM sizes ( Fig. 13). The highest concentrations were found in Noorderpark and the lowest concentrations were found in Vliegenbos, where PM concentrations were already lower during the night before. During the night and the morning of the 26th of August, all PM concentrations were lower in Vliegenbos than in Noorderpark and at the Johan van Hasseltweg. The same results were found during the morning of the 27th of August.

PM concentrations reached levels below 6 μg/m3 in <1% of the measurements. PM10 concentrations exceeded 50 μg/m3 on the 25th, 26th, 27th and 28th of August at all locations. On the 24th, 50 μg/m3 exceeded only at the Johan van Hasseltweg and on the 22nd in Noorderpark and Vliegenbos.

Table 10Daily average PM1, PM2.5 and PM10 concentrations (μg/m3) at Johan van Hasseltweg, Noorderpark, Vliegenbos and Van Diemenstraat measured on the 22nd - 27th of August. PM1 concentrations are missing at the Van Diemenstraat, because they are not measured by RIVM. The 10th and the 17th of July are lacking, because measurements were only taken during part of the day. Concentrations that exceed the daily limit of 50 μg/m3 for PM10 concentrations and 25 μg/m3 for PM2.5 concentrations are marked in orange.

Johan van Hasseltweg Noorderpark Vliegenbos Van Diemenstraat

Date PM1 PM2.5 PM10 PM1 PM2.5 PM10 PM1 PM2.5 PM10 PM2.5 PM10 22-Aug 9 14 16 9 15 17 8 12 14 10 22 23-Aug 7 10 13 6 10 12 6 9 11 16 26 24-Aug 16 24 27 17 25 27 15 21 24 19 27 25-Aug 19 29 34 20 31 35 18 26 31 23 32 26-Aug 26 40 50 27 42 52 24 37 46 28 39 27-Aug 28 44 56 28 45 55 26 40 51 34 46

In July, the average daily PM10 concentrations exceeded the limit of 50 μg/m3 on the 11th of July at all the measurement locations (Table 11). PM2.5 concentrations exceeded the limit of 25 μg/m3 on the 11th of July at all locations and on the 12th of July in Noorderpark ( Table 11). PM concentrations peaked between the 10th and 13th of July and slightly on the 17th of July ( Fig. 14). Maximum concentrations for all PM sizes were measured on the 11th of July ( Fig. 14). In Noorderpark the maximum PM concentrations were highest, reaching 47, 79 and 104 μg/m3 for PM1, PM2.5 and PM10 respectively (Table A4.1). The maximum PM2.5 concentration at the Van Diemenstraat was 41 μg/m3 between the 10th and 18th of July, i.e. 23-38 μg/m3 lower than at the study locations. The maximum PM10 concentration at the Van Diemenstraat was 97 μg/m3 between the 10th and 18th of July, i.e 5-9 μg/m3 higher than at the Johan van Hasseltweg and in Vliegenbos and 7 μg/m3 lower than in Noorderpark. The average PM2.5 and PM10 concentrations did not exceed the allowed yearly average limit of 25 and 40 μg/m3 (Table A4.2).

(28)

Table 11Daily average PM1, PM2.5 and PM10 concentrations (μg/m3) at Johan van Hasseltweg, Noorderpark, Vliegenbos and Van Diemenstraat measured on the 11th - 16th of July. PM1 concentrations are missing at the Van Diemenstraat, because they are not measured by RIVM. The 10th and the 17th of July are lacking, because measurements were only taken during part of the day. Concentrations that exceed the daily limit of 50 μg/m3 for PM10 concentrations and 25 μg/m3 for PM2.5 concentrations are marked in orange.

Johan van Hasseltweg Noorderpark Vliegenbos Van Diemenstraat

Date PM1 PM2.5 PM10 PM1 PM2.5 PM10 PM1 PM2.5 PM10 PM2.5 PM10 11-Jul 24 37 47 28 44 56 26 40 54 22 35 12-Jul 15 24 30 16 26 30 12 20 27 13 24 13-Jul 5 10 12 6 11 13 5 9 12 5 12 14-Jul 1 2 2 1 2 3 1 2 2 3 9 15-Jul 1 2 3 1 3 4 1 2 3 5 12 16-Jul 3 6 7 4 7 8 2 5 6 7 13

The average PM concentrations were calculated per part of the day (night: 12AM - 6AM, morning: 6AM - 12PM, afternoon: 12PM - 6PM, evening: 6PM - 12 PM). Differences between the study locations were found on the 10th, 11th, 13th, 16th and 17th of July ( Fig. 15). No differences between the locations were found during the rest of the week. During the evening of the 10th and the night of the 11th, PM concentrations were highest in Noorderpark, followed by Vliegenbos and Johan van Hasseltweg. During the morning of the 11th, PM concentrations were higher in Noorderpark and Vliegenbos than at the Johan van Hasseltweg. During the night of the 13th, PM concentrations were higher in Noorderpark than in Vliegenbos and at the Johan van Hasseltweg. On the evening of the 16th, PM concentrations were lower in Vliegenbos than in Noorderpark and at the Johan van Hasseltweg. During the night of the 17th, PM concentrations stayed lowest in Vliegenbos. During the morning PM concentrations were higher in Noorderpark than in Vliegenbos and at the Johan van Hasseltweg. During the afternoon of the 17th of July, PM concentrations reached a high peak in Noorderpark, causing a large standard error (Fig. 15).

Between the 13th and the 17th of July, PM concentrations were very low. At the Johan van Hasseltweg, 51% of the measured PM1-, 39% of the PM2.5- and 33% of the PM10-concentrations were lower than 6 μg/m3. In Noorderpark, 50% of the measured PM1, 39% of the PM2.5 and 32% of the PM10 concentrations were lower than 6. In Vliegenbos, 60% of the measured PM1, 44% of the PM2.5 and 38% of the PM10 concentrations were lower than 6. At the van Diemenstraat 36% of the PM2.5 measurements was lower than 6 and only 0.03% of the PM10 measurements was lower than 6. PM10 concentrations exceeded the limit of 50 μg/m3 on the 10th, 11th and 12th of July at all locations.

(29)

Figure 12Air quality measured in PM1, PM2.5 and PM10 concentrations between the 21st and the 29th of August next to the Johan van Hasseltweg, in Noorderpark and in Vliegenbos.

(30)

Figure 13 Average PM1, PM2.5 and PM10 in Vliegenbos, Noorderpark and Johan van Hasseltweg during night (12AM - 6AM), morning (6AM - 12PM), afternoon (12PM - 6PM) and evening (6PM - 12AM) between the 21st and the 29th of August. In Johan van Hasseltweg and Noorderpark 35 measurements were taken during each part of the day. In Vliegenbos, 20 measurements were taken during each part of the day. Error bars represent ±1SE.

(31)

Figure 14Air quality in PM1, PM2.5 and PM10 concentrations measured between the 10th and the 18th of July next to the Johan van Hasseltweg, in Noorderpark and in Vliegenbos.

(32)

Figure 15Average PM1, PM2.5 and PM10 in Vliegenbos, Noorderpark and Johan van Hasseltweg during night (12AM - 6AM), morning (6AM - 12PM), afternoon (12PM - 6PM) and evening (6PM - 12AM) between the 10th and the 18th of July. In Johan van Hasseltweg and Noorderpark 35 measurements were taken during each part of the day. In Vliegenbos, 20 measurements were taken during each part of the day. Error bars represent ±1SE.

Monitoring the quality and quantity of urban green space in Amsterdam with remote sensing and truth ground measurements - A case study on the mitigation of air pollution and the urban heat island effect by urban parks d