Evaluating urban greening scenarios for urban heat mitigation: a spatially-explicit approach

Urban green infrastructure, especially trees, are widely regarded as one of the most effective ways to reducing urban temperatures in extreme heat events, and alleviate its adverse impacts on human health and well-being. Nevertheless, urban planners and decision-makers are still lacking methods and tools to spatially evaluate the cooling effects of urban green spaces and exploit them to assess greening strategies at the urban agglomeration scale. This article introduces a novel spatially-explicit approach to simulate urban greening scenarios by increasing the tree canopy cover in the existing urban fabric, and evaluating their heat mitigation potential. The latter is achieved by applying the InVEST urban cooling model to the synthetic land use/land cover maps generated for the greening scenarios. A case study in the urban agglomeration of Lausanne, Switzerland, illustrates the development of tree canopy scenarios following distinct spatial distribution strategies. The spatial pattern of the tree canopy strongly influences the human exposure to the highest temperatures, and small increases in the abundance of tree canopy cover with the appropriate spatial configuration can have major impacts on human health and well-being. The proposed approach supports urban planning and the design of nature-based solutions to enhance climate resilience.

Urbanization is a global phenomenon that increasingly concentrates the world's 2 population in urban areas, with the latter expected to grow in both the number of 3 dwellers and spatial extent over the next decades [1][2][3]. As a major force of landscape 4 change, urbanization is characterized by the conversion of natural to artificial surfaces, 5 which alters the energy and water exchanges as well as the movement of air. Such 6 changes often result in the urban heat island (UHI) effect, a phenomenon by which 7 urban temperatures are warmer than its rural surroundings [4][5][6][7][8][9]. The negative impacts 8 of UHI have been widely documented and include increased energy and water 9 consumption [10][11][12], reduced workplace productivity [13,14] and aggravation of health 10 risks [15][16][17]. As urban areas grow and global temperatures rise, the UHI effect is 22 fine-grained approaches to evaluate the cooling effects of the spatial pattern of the tree 23 canopy at the urban agglomeration scale. 24 With the aim of addressing the above shortcomings, the present work introduces a 25 novel spatially-explicit method to evaluate the heat mitigation potential of altering the 26 abundance and spatial configuration of the urban tree canopy cover in realistic settings. 27 The proposed method consists of two major parts. First, synthetic scenarios are 28 generated by increasing the tree canopy cover in candidate locations where the existing 29 urban fabric permits it. Then, the spatial distribution of air temperature of each 30 synthetic scenario is estimated with the InVEST urban cooling model, which simulates 31 urban heat mitigation based on three biophysical processes, namely shade, 32 evapotranspiration and albedo. Finally, the simulated temperature map is coupled with 33 a gridded population census in order to evaluate the human exposure to urban heat in 34 the scenario. By applying such a procedure in the urban agglomeration of Lausanne, 35 Switzerland, this study aims to map the heat mitigation potential that can be achieved 36 starting from the existing urban fabric. With the aim of quantifying the effects of the 37 abundance and spatial configuration of the tree canopy cover on urban heat mitigation, 38 a set of synthetic scenarios are generated by increasing different proportions of tree 39 canopy cover in distinct spatial configurations. 40

41
Study area 42 Lausanne is the fourth largest Swiss urban agglomeration with 420757 inhabitants as of 43 January 2019 [28]. The agglomeration is located at the Swiss Plateau and on the shore 44 of the Lake Léman, and is characterized by a continental temperate climate with mean 45 annual temperatures of 10.9 • C and mean annual precipitation of 100 mm, with a 46 dominating vegetation of mixed broadleaf forest. The spatial extent of the study has 47 been selected following the recent application of the InVEST urban cooling model to 48 Lausanne by Bosch et al. [29], and covers an area of 112.46 km 2 . 49 In order to evaluate the human exposure to UHI, the population data for the study 50 area has been extracted from the population and households statistics (STATPOP) [30] 51 provided at a 100 m resolution by the Swiss Federal Statistical Office (SFSO) with the 52 Python library swisslandstats-geopy [31]. 53 Simulation with the InVEST urban cooling model 54 The spatial distribution of air temperatures is simulated with the InVEST urban 55 cooling model (version 3.8.0) [32], which is based on the heat mitigation provided by 56 shade, evapotranspiration and albedo. The main inputs are a land use/land cover 57 (LULC) raster map, a reference evapotranspiration raster and a biophysical table 58 2/20 containing model information of each LULC class of the map. The LULC maps have 59 been obtained by rasterizing the vector geometries of the official cadastral survey of the 60 Canton of Vaud [33] as of August 2019 to a 10 m resolution. Such a dataset 61 distinguishes 25 LULC classes which are relevant ot the urban, rural and wild 62 landscapes encountered in Switzerland. The reference evapotranspiration pixel values 63 are estimated with the Hargreaves equation [34] based on the daily minimum, average 64 and maximum air temperature values of the 1 km gridded inventory of by the Federal 65 Office of Meteorology and Climatology (MeteoSwiss) [35]. The biophysical table used in 66 this study is shown in Table S1. A more thorough description of the model and the data 67 inputs can be found in Bosch et al. [29]. 68 The parameters of the model are set based on its calibration to the same study area 69 in previous work [29]. Finally, the rural reference temperature (T ref ) and UHI 70 magnitude (U HI max ) values are derived from the air temperature of 11 monitoring 71 stations in the study area (see Figure S1). More precisely, T ref is set as the 9 p.m. air 72 temperature measurement -the moment of maximal UHI intensity in Switzerland [36]  Refining LULC classes based on tree cover and building density 80 A procedure to redefine the LULC classes from the cadastral survey has been designed 81 to distinguish the LULC classes depending on their proportional cover of both trees and 82 buildings. The reclassification is achieved by combining the 10 m raster LULC map 83 with two 1 m binary raster masks, one for the tree canopy raster and another for the 84 buildings. The 1 m binary tree canopy mask has been derived from the SWISSIMAGE 85 orthomosaic [37], by means of the Python library DetecTree [38], which implements the 86 methods proposed by Yang et al. [39]. The estimated classification accuracy of the tree 87 canopy classification is of 91.75%. On the other hand, the 1 m binary building mask has 88 been obtained by rasterizing the buildings of the vector cadastral survey [33].

89
The reclassification procedure consists of three steps. Firstly, each 10 m pixel is 90 coupled with the tree canopy and building masks in order to respectively compute its 91 proportion of tree and building cover. Secondly, the set of 10 m pixels of each LULC 92 class are grouped into a user-defined set of bins to form two histograms, one based on 93 their proportion of tree cover and the other analogously for the building cover. Lastly, 94 the two histograms are joined so that each LULC class is further refined into a set of 95 classes. For example, if two bins were used for both the tree and building cover, the 96 "sidewalk" LULC code might be further refined into "sidewalk with low tree/low 97 building cover", "sidewalk with low tree/high building cover", "sidewalk with high 98 tree/low building cover" and "sidewalk with high tree/high building cover".

99
In the present work, four equally spaced bins (i.e., distinguishing 0-25%, 25-50%, 100 50-75% and 75-100% intervals) have been used to reclassify each LULC class according 101 to both the tree and building cover. Following the advice given by the directorate of 102 resoures and natural heritage in the Canton of Vaud (DGE-DIRNA), the threshold over 103 which a pixel is considered to have a high tree canopy cover has been set to 75%, which 104 corresponds to placing trees of a spheric crown with a 5 m radius spaced 10 m from one 105 another so that they form a continuous canopy. Therefore, adjacent pixels with a tree 106 canopy cover over 75% can Finally, in order to adapt the biophysical table of the 107 InVEST urban cooling model to the reclassified LULC classes, the shade coefficients are 108 computed as the midpoint of the bin interval of each level of tree cover (i.e., 0.125, 109 3/20 0.375, 0.625 and 0.875), whereas the albedo coefficients have been linearly interpolated 110 based on the level of building cover (see Table S1).

111
Generation of urban greening scenarios 112 Starting from the refined LULC map, a set of urban greening scenarios are generated by 113 altering the LULC classes of certain candidate pixels in a way that corresponds to 114 reasonable transformations that could occcur in urban areas. More precisely, pixels 115 whose base LULC class corresponds to "building", "road, path", "sidewalk", "traffic 116 island", "other impervious" and "garden" are changed to the LULC code that has the 117 same base class but with the highest tree cover, e.g., pixels of a post-refinement class 118 "sidewalk with low tree/low building cover" are be changed to "sidewalk with high 119 tree/low building cover". In order to ensure that such an increase of the tree canopy 120 cover is performed only where the existing urban fabric permits it, pixels might only be 121 transformed when two conditions are met. First, the proportion of building cover in the 122 candidate pixels must be under 25%, i.e., there is a 75% of the pixel area which could 123 be occupied by a tree crown. Secondly, pixels of the "road, path" class might only be 124 transformed when they are adjacent to a pixel of a different class, which prevents 125 increasing the tree canopy cover in pixels that are in the middle of a road (e.g., a 126 highway).

127
After mapping the candidate pixels where the tree canopy cover can be increased, 128 scenarios are generated based on two key attributes: the extent of tree canopy 129 conversion (expressed as a proportion of the total number of candidate pixels), and the 130 selection of pixels to be converted. A set of scenarios is generated by transforming a 131 12.5, 25, 37.5, 50, 62.5, 75 and 87.5% of the candidate pixels respectively. For each of 132 these canopy areas, three distinct selection approaches are used. The first consists in 133 randomly sampling from the candidate pixels until the desired proportion of changed 134 pixels is matched. In the second and third approaches, the candidate pixels are sampled 135 according to the number of pixels with high tree canopy cover (i.e., greater than 75%) 136 found in their Moore neighborhood (i.e., the 8 adjacent pixels). In the second approach, 137 pixels with higher number of high tree canopy cover neighbors are transformed first, 138 which intends to spatially cluster pixels of high tree canopy cover. The third approach 139 intends to spatially scatter pixels of high tree canopy cover by prioritizing pixels with 140 lower number of high tree canopy neighbors. Given that the three sampling approaches 141 are stochastic, for each scenario configuration, i.e., each pair of proportion of 142 transformed candidate pixels and sampling approach, the corresponding temperature 143 maps will be computed by averaging a number of simulation runs. After observing little 144 variability among the simulation results, the number of runs of a each configuration has 145 been set to 10. Lastly, the set of scenarios is completed with a configuration where a 146 100% of the candidate pixels are transformed, which is independent of the sampling  For each scenario, the spatial pattern of the tree canopy is quantified by means of a 152 set of spatial metrics from landscape ecology [40,41], which are computed for the pixels 153 whose post-refinement LULC class has a tree canopy cover over 75% 1 . Based on other 154 studies that explore the relationship between the spatial of tree canopy and UHIs, four 155 spatial metrics have been chosen to quantify both the composition and oconfiguration of 156 1 Following the advice given by the directorate of resoures and natural heritage in the Canton of Vaud (DGE-DIRNA), the threshold over which a pixel is considered to have a high tree canopy cover has been set to 75%, which corresponds to placing trees of a spheric crown with a 5 m radius spaced 10 m from one another so that they form a continuous canopy.

4/20
the tree canopy, which are listed in Table 1. The proportion of landscape (PLAND) of 157 pixels with high tree canopy cover serves to quantify the composition aspects, while the 158 configuration is quantified by means of the mean patch size (MPS), edge density (ED) 159 and the mean shape index (MSI) of patches of high tree canopy cover. The four metrics 160 have been computed with the Python library PyLandStats [42]. Table 1. Selected landscape metrics. A more thorough description can be found in the documentation of the software FRAGSTATS v4 [41] Category

Metric name Description
Composition Percentage of landscape (PLAND) Percentage of landscape, in terms of area, occupied by pixels with high tree canopy cover i.e., increasing the tree canopy in all the possible pixels, 61.50% of the pixels correspond 169 to the "garden" LULC class, followed by "road, path", "building", "other impervious", 170 (18.01, 10.81 and 7.69%, respectively). Finally, the LULC classes of "sidewalk" and 171 "traffic island" constitute only 1.67 and 0.3% of the pixels where the tree canopy can be 172 increased. The differences when considering the sampling approaches separately are 173 small relative to the total number of transformed candidate pixels. The largest 174 differences between sampling approaches can be noted in the number of transformed 175 pixels that originally belong to the "garden" class. When transforming 25, 50 and 75% 176 of the candidate pixels, clustering respectively transforms (on average among the 177 simulation runs) a 0.90, 0.38 and 0.12% more garden pixels than random sampling, and 178 1.28, 0.76 and 0.43% more garden pixels than the scattering approach ( Figure 2).   transformed and the simulated distribution of air temperature can be approximated as a 189 linear relationship with a negative slope (see Figure S2 and Figure S3 for more details 190 about this relationship).    Regarding the configuration metrics, the values of the mean patch area (AREA MN) 204 show that the clustering and random sampling approaches lead to larger patches of high 205 tree canopy cover than the scattering approach. When transforming a 12.5 and 25% of 206 the candidate pixels, clustering them to other pixels of high tree canopy cover increases 207 7/20 AREA MN from 0.14 to 0.54 hectares respectively (on average over the simulation runs). 208 For 37.5% of transformed candidate pixels in the clustering approach, AREA MN shows 209 a sudden decline to 0.20 hectares, followed by a monotonic increase that reaches 1.52 210 hectares when all the candidate pixels are transformed. Such a discernable kink in the 211 computed AREA MN reveals characteristics of the existing urban fabric, and describes 212 the point after which all the candidate pixels that are adjacent to other pixels of high 213 tree canopy have been transformed and hence new pixels have to be allocated as part of 214 new (and smaller) patches. The same kink is even more notable for the mean shape  growing existing patches by clustering the new pixels to them accounts for less total 224 edge length than scattering the same amount of new pixels in a leapfrog manner. In the 225 three approaches, the ED increases monotonically at first until an apex is reached when 226 the proportion of transformed pixels is between 50% and 60%, and then declines 227 monotonically.

228
The average simulated temperature T is overall negatively correlated with 229 AREA MN, which suggests that for the same amount of high tree canopy pixels, large 230 patches provide lower heat mitigation. On the other hand, configurations with the same 231 proportion of high tree canopy pixels show lower T for larger values of ED, which 232 suggests that edge effects between artificial patches and patches of high tree canopy 233 contribute to greater heat mitigation. Nonetheless, as higher proportions of candidate 234 pixels are transformed and the locations of the remaining candidate pixels force the 235 overall ED to decrease, the simulated average temperatures continue to decline. This 236 highlights how the cooling effects of the abundance of tree canopy overshadow those of 237 the spatial configuration, which is consistent with many related research.

238
Effects on human exposure 239 The relationship between human exposure to air temperatures higher than 21, 22, 23, 240 24, 25 and 26 • C and the proportion of pixels transformed to their respective high tree 241 canopy cover class is shown in Figure 5. The number of dwellers exposed to 242 temperatures higher than 21 • C does not show a significant decrease (even when 243 converting all the candidate pixels), whereas for temperatures higher than 22 • C, it 244 diminishes from 269254 to 268601, 267683, 266518 and 264125 when the proportion of 245 transformed pixels is of 25, 50, 75 and 100% respectively, which represents a relative 246 share of 97.25, 97.02, 96.69 96.27 and 95.41% of the population of the study area. Such 247 a decline progressively becomes more notable as temperatures increase, e.g., the share of 248 the population exposed to temperatures over 24 • C declines from an initial 78.4% to 249 72.39, 59.57, 37.53 and 11.52% when transforming a 25, 50, 75 and 100% of the 250 candidate pixels respectively. Finally, the share of dwellers exposed to temperatures 251 over 25 • C, which is initially of 47.91%, is diminished to a 24.98 and 5.74% when 252 transforming a 25 and 50% of the pixels respectively, and becomes 0 after that, whereas 253 the 2508 dwellers originally exposed to temperatures over 26 • C do no longer meet such 254 temperatures after transforming a 25% of the candidate pixels. 255 The way in which the transformed pixels are sampled has significant effects on the to reduce the human exposure to the highest temperatures, followed by random 259 sampling. When transforming a 25 and 50% of the candidate pixels with the scattering 260 approach, the number of dwellers exposed to temperatures over 25 • C decreases from 261 124073 to 65108 and 4498 respectively. Such a reduction is larger than its random 262 sampling counterpart by 3125 and 8223 dwellers respectively, and larger than its The scenarios simulated in this study map locations where the tree canopy cover in the 267 urban agglomeration of Lausanne can be increased, and suggests that such changes can 268 result in urban nighttime temperatures that are up to 2 • C lower. The results indicate 269 that given the same proportion of tree canopy cover, a scattered configuration might 270 lead to more effective urban heat mitigation than a clustered one, which is in line with 271 previous studies in humid climates [43][44][45][46][47][48]. Nevertheless, the results suggest that effect 272 of the spatial configuration (measured by the metrics AREA MN, SHAPE MN and ED) 273 is secondary when compared to the effect of the composition (measured by the PLAND 274 metric). Overall, the effect of the spatial configuration of trees on its urban heat 275 mitigation depends on how it affects the shading and evapotranspiration processes.

278
The spatial effects observed in the results are due to the InVEST model equations from the observed spatial patterns of air temperature [35,54]. Moreover, the cooling 291 effects of large green spaces have been found to be non-proportional to their area and 292 shape complexity [55][56][57]. Improving how these non-linear components are represented 293 in the InVEST urban cooling model could enhance not only its validity, but also its 294 value to urban planning by identifying thresholds and regime changes in the cooling 295 efficiency of additional tree planting.

296
Despite the limitations noted above, a major advantage of the proposed approach is 297 that it can be used to evaluate urban heat mitigation of synthetic scenarios. The 298 simulations presented in this article focus on spatially exploring the effects of an 299 increase of the tree canopy cover, yet there is room for much more experimentation of 300 this kind. On the one hand, the generic sampling approaches explored above can be 301 extended to consider ad-hoc characteristics such as the spatial distribution of the 302 population, and design optimization procedures with specific goals. For instance, the 303 candidate pixels can be selected with the aim of minimizing the exposure of the most 304 vulnerable populations to critical heat thresholds. More broadly, the approach can be 305 used as part of decision support system to explore the trade-offs between ecosystem 306 services provided by trees, perform weighted optimizations and map priority planting 307 locations [58]. On the other hand, in line with recent studies [59][60][61], the approach 308 could be applied to examine the impact of distinct urbanization scenarios such as 309 densification and urban sprawl on air temperature and human exposure to extreme heat, 310 under current conditions as well as future climate estimates, e.g., by changing the T ref 311 or U HI max parameters. Similarly, InVEST urban cooling model might be coupled with 312 models of LULC change such as cellular automata in order to assess not only which 313 scenarios are most desirable in terms of urban heat mitigation, but also which planning 314 strategies might lead to them [62][63][64]. progressive coalescensce of artificial surfaces in its inner ring [65]. Such an infilling trend 319 urges for careful evaluation of the beneficial ecosystem services provided by urban green 320 spaces, which should be balanced against the adverse consequences of urban 321 sprawl [26,27].

322
The approach proposed in this study maps locations in the current urban fabric 323 where the tree canopy cover can be increased. While part of this urban greening might 324 occur in impervious surfaces (e.g., in sidewalks, next to roads and in other impervious 325 surfaces), most of the candidate locations currently correspond to urban green space 326 (i.e., the "garden" LULC class). Therefore, the potential heat mitigation suggested by 327 the results study is not attainable in a scenario of severe infill development.

328
Additionally, densification strategies should consider that newly created urban green 329 space might result in less provision of ecosystem services than remnant natural 330 patches [25,66,67]. Finally, infilling might exacerbate the unevenness of the accessibility 331 to green areas by depriving dwellers of the most dense parts in city core from their few 332 remaining urban green spaces. Spatial heterogeneity of this kind, which are encountered 333 in many socioeconomic and environmental aspects of contemporary cities, are often 334 hard to represent with aggregate indicators and highlight the importance of spatially 335 explicit models to urban planning and decision making.

336
The explicit representation of space is also crucial when considering the impacts of 337 urban green space on human exposure to extreme heat. Although the simulated 338 scenarios suggest that the impact of the spatial pattern of tree canopy on the air 339 temperature is practically linear, the implications on human exposure to critical 340 temperatures exhibit important thresholds. For example, by increasing the tree canopy 341 cover of a 25% of the candidate pixels, the number of dwellers exposed to nighttime 342 temperatures over 25 • C can be reduced from 124073 to 74466, which respectively 343 represents a 45.08 and 27.06% of the total population in the study area. Furthermore, 344 the results suggest by selecting such pixels to prioritize a spatial scattering of the tree 345 canopy cover, such a population can be reduced by an additional 3125 or 6234 dwellers 346 when respectively compared to random sampling such pixels or clustering them to the 347 existing tree canopy cover. In Switzerland, the excess mortality associated to the heat 348 wave of 2003 occurred over-proportionally to urban and sub-urban residents of its 349 largest urban agglomerations [68]. Furthermore, the association between temperature suggest that the spatial configuration in which the tree canopy is increased influences its 361 heat mitigation effects. The configuration effects become more significant when 362 considering the impacts on the urban population, and small increases in the tree canopy 363 can result in important reductions in the number of dwellers exposed to the highest 364 temperatures. Overall, the presented approach provides a novel way to explore how the 365 urban tree canopy of can be exploited to reduce heat stress. Future studies can extend 366 the analyses by assessing the provision of other ecosystem services in the various tree 367 canopy strategies presented here. 368

11/20
Supporting Information 369 S1 Data 370 S1.1 Biophysical table 371 The biophysical table for the LULC codes (before the reclassification) is shown in 372   Table S1. The crop and water coefficients are based on Allen et al. [70], while rock, soil 373 and urban coefficients are derived from the results of Grimmond and Oke [71] in the city 374 of Chicago. Given that the evapotranspiration of the vegetation and crops is subject to 375 seasonal changes in temperate zones such as Switzerland [70], the values that correspond 376 to the mid-season estimation (June to August) in [72]. The albedo values are based on 377 the work of Steward et al. [73]. The shade column, which represents the proportion of 378 tree cover of each LULC class, is computed after the reclassification procedure described 379 in section "Refining LULC classes based on tree cover and building density". 380 Table S1. Biophysical table (before the reclassification). The source comma-separated value (CSV) file used in the computational workflow is available at https://github.com/martibosch/ lausanne-heat-islands/blob/master/data/raw/biophysical-   [74]. 386 The source CSV file with the operator, location and elevation in meters above sea level 387 of the monitoring stations used in the computational workflow is available at 388 https://github.com/martibosch/lausanne-greening-scenarios/blob/master/ 389 data/raw/tair-stations/station-locations.csv. The code to produce Figure S1 390 is available as a Jupyter Notebook (IPYNB) at https://github.com/martibosch/ 391 lausanne-greening-scenarios/blob/master/notebooks/stations.ipynb.

T [°C]
Approach random cluster scatter Figure S2. Relationship between the proportion of candidate pixels transformed and the average simulated temperature T for each scenario sample. The translucent bands around the regression line represent the 95% confidence intervals estimated using a bootstrap.