Heat and Trees¶

Written by Julia Signell
Created: October 1, 2018
Last updated: July 30, 2021

Urban Heat Islands and Street Trees¶

In this notebook we'll be exploring the urban heat island effect by looking at the impact on surface temperature of roof color and street trees. We'll be replicating the process described here: http://urbanspatialanalysis.com/urban-heat-islands-street-trees-in-philadelphia/ but using Python tools rather than ESRI.

Extra packages: To run this notebook, you'll need the PyViz tools and a library of top of atmosphere calculations from rio-toa: pip install rio-toa

Data sources: This notebook uses Landsat data from Google Cloud Storage as well as some geographic data from OpenDataPhilly.

In [1]:
import intake
import xarray as xr
import pandas as pd
import numpy as np

import geopandas as gpd

import cartopy.crs as ccrs

import hvplot.xarray  # noqa
import hvplot.pandas  # noqa

from geoviews.tile_sources import EsriImagery

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
/tmp/ipykernel_6160/1571672325.py in <module>
----> 1 import intake
2 import xarray as xr
3 import pandas as pd
4 import numpy as np
5

ModuleNotFoundError: No module named 'intake'

Just some extra info about Landsat data:

In [ ]:
band_info = pd.DataFrame([
(1,  "Aerosol", " 0.43 - 0.45",    0.440,  "30",         "Coastal aerosol"),
(2,  "Blue",    " 0.45 - 0.51",    0.480,  "30",         "Blue"),
(3,  "Green",   " 0.53 - 0.59",    0.560,  "30",         "Green"),
(4,  "Red",     " 0.64 - 0.67",    0.655,  "30",         "Red"),
(5,  "NIR",     " 0.85 - 0.88",    0.865,  "30",         "Near Infrared (NIR)"),
(6,  "SWIR1",   " 1.57 - 1.65",    1.610,  "30",         "Shortwave Infrared (SWIR) 1"),
(7,  "SWIR2",   " 2.11 - 2.29",    2.200,  "30",         "Shortwave Infrared (SWIR) 2"),
(8,  "Panc",    " 0.50 - 0.68",    0.590,  "15",         "Panchromatic"),
(9,  "Cirrus",  " 1.36 - 1.38",    1.370,  "30",         "Cirrus"),
(10, "TIRS1",   "10.60 - 11.19",   10.895, "100 * (30)", "Thermal Infrared (TIRS) 1"),
(11, "TIRS2",   "11.50 - 12.51",   12.005, "100 * (30)", "Thermal Infrared (TIRS) 2")],
columns=['Band', 'Name', 'Wavelength Range (µm)', 'Nominal Wavelength (µm)', 'Resolution (m)', 'Description']).set_index(["Band"])
band_info


For this example, we will be using landsat data stored on Google Cloud Storage. Since these data are accessed via https, there is no guaranteed directory structure, so we will need to specify the url pointing to each file and then iterate over the files to create a concatenated dataset. We use jinja template notation in intake to pass parameters to the urlpath.

In [ ]:
cat = intake.open_catalog('catalog.yml')
list(cat)


Let's take a look at what the google_landsat_band looks like:

yml
description: Landsat bands from Google Cloud Storage
driver: rasterio
parameters:
path:
description: landsat path
type: int
row:
description: landsat row
type: int
product_id:
description: landsat file id
type: str
band:
description: band
type: int
args:
urlpath: https://storage.googleapis.com/gcp-public-data-landsat/LC08/01/{{ '%03d' % path }}/{{ '%03d' % row }}/{{ product_id }}/{{ product_id }}_B{{ band }}.TIF
chunks:
band: 1
x: 256
y: 256

The following might feel a bit arbitrary, but we have chosen the path and row corresponding to the area over Philadelphia using the earth explorer. We have also found the id of the particular date of interest using the same tool. With these values in hand, we can access parts of each file on Google Cloud Storage.

In [ ]:
path = 14
row = 32
product_id = 'LC08_L1TP_014032_20160727_20170222_01_T1'


The first step to using intake is to initialize the catalog entry with user parameters to create a data source.

In [ ]:
data_source = cat.google_landsat_band(path=path, row=row, product_id=product_id)


From this data source we can get a lazily loaded xarray object using dask. To make sure that we can inspect what dask is up to, it can be helpful to create a dask client.

In [ ]:
ds = data_source.to_dask()
ds.name = 'value'


In [ ]:
def load_google_landsat_metadata(path, row, product_id):
"""
def parse_type(x):
try:
return eval(x)
except:
return x

suffix = f'{path:03d}/{row:03d}/{product_id}/{product_id}_MTL.txt'

df = intake.open_csv(
urlpath=f'{baseurl}/{suffix}',
csv_kwargs={'sep': '=',
'names': ['index', 'value'],
'skiprows': 2,
'converters': {'index': (lambda x: x.strip()),

In [ ]:
metadata = load_google_landsat_metadata(path, row, product_id)


Sub-setting to area of interest¶

So far we haven't downloaded any band data. Since we know that we are interested in Philadelphia, we can just take a smaller square of data that covers the extents of the city. First we need to know the projection of the dataset:

In [ ]:
ds.crs


We'll convert that into something directly usable for later:

In [ ]:
crs = ccrs.epsg(32618)


Now if we were just looking for one particular point we could use that point, converted to the coordinate system of the data, and then select the data nearest to it:

In [ ]:
x_center, y_center = crs.transform_point(-75.1652, 39.9526, ccrs.PlateCarree())
nearest_to_center = ds.sel(x=x_center, y=y_center, method='nearest')
print(nearest_to_center.compute())

nearest_to_center.hvplot.line(x='band')


In this case, though, we are interested in a subset of data that covers that city of Philadelphia. So we need some geometry to specify the bounds of the city. We can get a GeoJSON of neighborhood data from OpenDataPhilly.

In [ ]:
url = 'https://github.com/azavea/geo-data/raw/master/Neighborhoods_Philadelphia/Neighborhoods_Philadelphia.geojson'


We can compute the bounds of each of these neighborhoods and then using min and max get a rectangle that encompasses all of Philly.

In [ ]:
bounds = geoms.geometry.bounds
lower_left_corner_lat_lon = bounds.minx.min(), bounds.miny.min()
upper_right_corner_lat_lon = bounds.maxx.max(), bounds.maxy.max()

print(lower_left_corner_lat_lon, upper_right_corner_lat_lon)


Using the crs defined above, we can transform these lat lons into map coordinates.

In [ ]:
ll_corner = crs.transform_point(*lower_left_corner_lat_lon, ccrs.PlateCarree())
ur_corner = crs.transform_point(*upper_right_corner_lat_lon, ccrs.PlateCarree())

print(ll_corner, ur_corner)


Then we can use those corners to slice the data. If the subset is empty along x or y, the ordering of the coordinates might not be what you anticipated. Try swapping the order of arguments in the slice.

In [ ]:
subset = ds.sel(x=slice(ll_corner[0], ur_corner[0]), y=slice(ur_corner[1], ll_corner[1]))


We can persist this slice of the dataset in memory for easy use later.

In [ ]:
subset = subset.persist()
subset


To check that we got the right area, we can do a simple plot of one of the bands and overlay the neighborhoods on top of it. We'll use hvplot to quickly create a holoviews object rendered in bokeh.

In [ ]:
band_plot = subset.mean('band').hvplot(x='x', y='y', datashade=True, project=True, crs=crs, cmap='gray')
hood_plot = geoms.hvplot(geo=True, alpha=.5, c='mapname', legend=False, frame_height=450)

band_plot * hood_plot


Calculate NDVI¶

We'll calculate NDVI but we won't yet do any computations -- our bands are actually dask arrays, which allow for lazy computation.

In [ ]:
subset = subset.where(subset > 0)
NDVI = (subset.sel(band=5) - subset.sel(band=4)) / (subset.sel(band=5) + subset.sel(band=4))
NDVI = NDVI.where(NDVI < np.inf)
NDVI


In order to visualize NDVI, the data will need to be loaded and the NDVI computed. We can expect this to take some non-trivial amount of time (on the order of 20 sec on my machine).

In [ ]:
NDVI.hvplot(x='x', y='y', crs=crs, datashade=True, project=True, cmap='viridis', frame_height=450)


Calculate land surface temperature¶

Given the NDVI calculated above, we can determine land surface temperature. For ease, we'll use some top of atmosphere calculations that have already been written up as Python functions as part of rasterio work in the rio_toa module. We'll also need to specify one more for transforming satellite temperature (brightness temp) to land surface temperature. For these calculations we'll use both Thermal Infrared bands - 10 and 11.

In [ ]:
from rio_toa import brightness_temp, toa_utils

In [ ]:
def land_surface_temp(BT, w, NDVI):
"""Calculate land surface temperature of Landsat 8

temp = BT/1 + w * (BT /p) * ln(e)

BT = At Satellite temperature (brightness)
w = wavelength of emitted radiance (μm)

where p = h * c / s (1.439e-2 mK)

h = Planck's constant (Js)
s = Boltzmann constant (J/K)
c = velocity of light (m/s)
"""
h = 6.626e-34
s = 1.38e-23
c = 2.998e8

p = (h * c / s) * 1e6

Pv = (NDVI - NDVI.min() / NDVI.max() - NDVI.min())**2
e = 0.004 * Pv + 0.986

return BT / 1 + w * (BT / p) * np.log(e)


Now we'll set up a helper function to retrieve all the parameters from the metadata and general Landsat info table, and calculate the land surface temperature for bands 10 and 11.

In [ ]:
def land_surface_temp_for_band(band, data, units='F'):
# params from general Landsat info
w = band_info.loc[band]['Nominal Wavelength (µm)']

# params from specific Landsat data text file for these images

at_satellite_temp = brightness_temp.brightness_temp(data.sel(band=band).values, ML, AL, K1, K2)

temp = land_surface_temp(at_satellite_temp, w, NDVI).compute()

In [ ]:
temp_10_f = land_surface_temp_for_band(10, subset)
temp_11_f = land_surface_temp_for_band(11, subset)

temp_f = xr.concat([temp_10_f, temp_11_f],
dim=xr.DataArray([10,11], name='band', dims=['band']))
temp_f


Compare the results from the two different bands, noticing that the colorbars are different.

In [ ]:
temp_f.hvplot(x='x', y='y', groupby='band', cmap='fire_r',
crs=crs, rasterize=True, project=True, frame_height=350).layout()


We'll take the mean of the calculated land surface temperature for each of the bands and mimic the colormap used in the project that we are duplicating.

In [ ]:
mean_temp_f = temp_f.mean(dim='band')
mean_temp_f.hvplot(x='x', y='y', title='Mean Surface Temp (F)', crs=crs, tiles='EsriImagery',
frame_height=450, project=True, cmap='rainbow', alpha=.5, legend=False)


Notice how the hot spots are located over warehouse roofs and parking lots. This becomes even more visible when just the temperatures greater than 90F are displayed. To show this, we'll make a special colormap that just includes high intensity reds that are found at the top of the fire_r colormap.

In [ ]:
import colorcet as cc

special_cmap = cc.fire[::-1][90:]

In [ ]:
thresholded_temp_p = (mean_temp_f.where(mean_temp_f > 90)
.hvplot(x='x', y='y', title='Mean Temp (F) > 90',
cmap=special_cmap, crs=crs, frame_width=400,
frame_height=450,
colorbar=False, legend=False)
.redim(value='Temperature (F)'))

thresholded_temp_p + thresholded_temp_p.opts(alpha=.3, data_aspect=None) * EsriImagery


Adding in the Street Tree data¶

OpenDataPhilly released an inventory of all the street trees in the city. Street trees are trees that are planted along streets, not those in parks and private property. The original analysis considered these 100,000 points too many to plot, but that's nothing to datashader, which is happy with billions of points even on a laptop.

It is hypothesized that where there are more trees the land surface temperature will be less extreme. To explore this, we will overlay street trees with the thresholded land surface temperature:

In [ ]:
url = 'http://data.phl.opendata.arcgis.com/datasets/957f032f9c874327a1ad800abd887d17_0.geojson'


The trees plot can be generated straight from the geopandas dataframe, but that is rather slow. By inspecting the dataframe, we can see that each tree is represented as a point. We can create a simpler pandas dataframe with lat as one column and lon as the other.

In [ ]:
trees_df = pd.DataFrame({'Longitude': trees_gdf.geometry.x, 'Latitude': trees_gdf.geometry.y})

tree_p = trees_df.hvplot.points('Longitude', 'Latitude', title='Street Tree Density',