Diurnal evolution of urban tree temperature at a city scale

Despite the importance of urban trees’ surface temperature in assessing micro-climate interactions between trees and the surrounding environment, their diurnal evolution has been largely understudied at a city-wide scale due to a lack of effective thermal observations. By downscaling ECOSTRESS land surface temperature imaginary over New York City, we provide the first diurnal analysis of city-scale canopy temperature. Research reveals a remarkable spatial variation of the canopy temperature during daytime up to 5.6 K (standard deviation, STD), while the nighttime STD remains low at 1.7 K. Further, our analysis shows that the greenspace coverage and distance to bluespaces play an important role in cooling the local canopy during daytime, explaining 25.0–41.1% of daytime spatial variation of canopy temperatures while surrounding buildings modulate canopy temperature asymmetrically diurnally: reduced daytime warming and reduced nocturnal cooling. Built on space-borne observations and a flexible yet robust statistical method, our research design can be easily transferable to explore urban trees’ response to local climate across cities, highlighting the potentials of advancing the science and technologies for urban forest management.


Diurnal Evolution of Urban Tree Temperature at a City Scale
Thuy Trang Vo and Leiqiu Hu Table section   Table S1. Summary of ECOSTRESS LST imaginary

Date
Time ( To explore the statistical coexistence of vegetation types (e.g. tree canopies and grassland), we resampled the original grid size (70 m x 70 m) to larger grid size (490 x 490 m) to explore the spatial patterns of vegetation. A preliminary test was conducted to decide that the grid size of 490 x 490 m (equals to a neighbor square) is sufficient to investigate the interaction between tree canopies and grassland at a city-scale size.
The relationship between the fraction of canopy coverage and its own surface temperature is shown ( Figure S1). It is clear that at fraction of canopy coverage less than 15 % is dominated by grass coverage (mean coverage fraction is 19 % for the whole domain). Within this fraction range, tree temperature is decreased linearly at the degree of 2 K per 5 % of canopy coverage. On the other hand, at a fraction of canopy coverage larger than 15 %, tree temperature remains almost constant. Therefore, we found that it is difficult to distinguish them statistically as shown that grass coverage plays a statistically important role in this model. Figure S1. Coexistence effect of the tree and grassland fraction on the estimated tree LST using LST and Tree Canopy Temperature Estimation approach. Figure S1 (a) represents the linear relationship between estimated tree canopies LST and the fraction of tree canopies. The red box indicates the abnormal tree LST in the range of tree canopies fraction less than 15 %. Figure S1 (b) represents the distribution of other land cover components within this fraction range (tree canopies < 15 %). The light green box plot demonstrates the distribution of grassland, which is dominant (accounts for 20 % on average in total) compared to other land cover types.

Explanation of warming effects of urban bluespaces at night (3 LT observation)
On July 04 2019 (3 LT), there was an abnormal cooling Sea Surface Temperature (SST) pattern in the ocean near NYC. This pattern cools down the surface temperature of bluespaces surrounding this city and thus cools down the nearshore tree canopies accordingly. It could be observed on July 01 2019 (4 LT) as well as other nocturnal observations that the surface temperature of water bodies is much warmer at night compared to the patterns on July 04 2019 (3 LT). These warmer patterns warmed up the nearshore tree canopies.

Borough-based GAMs model
By statistics, Staten Island has the highest tree cover among boroughs with almost 30 %, followed by The Bronx (23 %), Manhattan (20 %), Queens (18 %), and Brooklyn (16 %). The distribution of building heights among these boroughs demonstrated by the Building Height dataset. A series of borough-level GAMs were conducted separately for each borough, as summarized in Figure S3 -S8. Figure S3. Estimated effects of urban greenspaces, distance to waterbodies, and building height on the spatial and diurnal patterns of ∆ for Manhattan. The left column is for daytime hours and the right column is for nighttime hours. Domain mean LST was used as a reference value for ∆ . ∆ equals zero is indicated as black dashed lines. The shaded regions illustrate the confidence interval 95% of GAMs fitting. Figure S4. Estimated effects of urban greenspaces, distance to waterbodies, and building height on the spatial and diurnal patterns of ∆ for The Bronx. The left column is for daytime hours and the right column is for nighttime hours. Domain mean LST was used as a reference value for ∆ . ∆ equals zero is indicated as black dashed lines. The shaded regions illustrate the confidence interval 95% of GAMs fitting. Figure S5. Estimated effects of urban greenspaces, distance to waterbodies, and building height on the spatial and diurnal patterns of ∆ for Brooklyn. The left column is for daytime hours and the right column is for nighttime hours. Domain mean LST was used as a reference value for ∆ . ∆ equals zero is indicated as black dashed lines. The shaded regions illustrate the confidence interval 95% of GAMs fitting. Figure S6. Estimated effects of urban greenspaces, distance to waterbodies, and building height on the spatial and diurnal patterns of ∆ for Queens. The left column is for daytime hours and the right column is for nighttime hours. Domain mean LST was used as a reference value for ∆ . ∆ equals zero is indicated as black dashed lines. The shaded regions illustrate the confidence interval 95% of GAMs fitting. Figure S7. Estimated effects of urban greenspaces, distance to waterbodies, and building height on the spatial and diurnal patterns of ∆ for Staten Island. The left column is for daytime hours and the right column is for nighttime hours. Domain mean LST was used as a reference value for ∆ . ∆ equals zero is indicated as black dashed lines. The shaded regions illustrate the confidence interval 95% of GAMs fitting.

Sensitivity test to decide an appropriate kernel size for End-member Temperature Estimation
In order to decide an appropriate kernel size for End-member Temperature Estimation, we conducted a sensitivity analysis for various combinations of moving windows. There are 3 sets of moving size were chosen to test the sensitivity of this approach on estimating canopy temperature. As concluded, we found that the distribution of canopy temperature using these moving window sets is not significantly different. Therefore, we chose set [10,5], which means moving window size of 10 x 10 pixels and moving steps of 5 pixels or in other words, kernel size of 5 x 5 pixels. Based on this decision, we assumed that the end-member temperatures are homogeneous within a neighbor scale of 5 x 5 pixels (or, 350 x 350 meters). Figure S9. The distribution of estimated canopy temperature is plotted for various window sizes and moving pixels. The x-axis represents the symbol for each test, for example, [10,5] means window size 10 x 10 pixels with a moving step of 5 pixels.

Model performance of End-member Temperature Estimation approach
In our analysis, we conducted an MLR method to statistically decompose the sub-pixel temperatures among different urban elements at each pixel using Equation 1 in the main document. To confirm the consistency of the results, we showed the distribution of the residuals extracted from the model output for two times of the day: at 13 Local Time (LT) and at 3 LT ( Figure S10 Besides, we also showed residual maps (the difference between modeled LST and ECOSTRESS LST) for those 2 days in Figure S11. It could be seen that overall, our model has acceptable performance, proven by the fact that the residual values follow Gaussian distribution ( Figure S10) and the residual maps show near homogenous residual values in the range of -2 to 2 (K). Despite some outliers (e.g., high absolute residual values), especially on the daytime residual map ( Figure  S11 a), our modeled LST values show reliable results compared to the ECOSTRESS LST, which confirms a good accuracy in our downscaled approach.