COALESCENCE OF REMOTELY SENSED LAND CLASSIFICATION PRECURSORS AND IN-SITU TEMPERATURE MEASUREMENTS:


INFORMING URBAN/REGIONAL HEAT ISLAND PREDICTIONS


Jackson Voelkel | Portland State University | 3/13/2017

Introduction

Global Warming

  • Not all temperature increases are created equal
  • Pacific Northwest: up to 5.7°C temperature increase in 100 years (Mote & Salathé Jr.,2010)
  • With over 50% of the world’s population living in cities (United Nations, 2015), we have a problem…

The Urban Heat Island Effect

  • Intra-urban phenomena
  • Spatially and temporally variable
  • Cities are warmer than rural areas
    • Documented since early 1800’s (Howard, 1820)

Graphic from Wikimedia Commons. Is this an over-simplification?

Heat and Health

The City of Portland (among many others) considers UHI an important long-term planning point.

Satellite- vs. Traverse-based UHI

90m Grid (ASTER TIR spatial resolution) Example

Previous Work

Traverse-based UHI Modeling

  • Gell out of style with the introduction of free satellite imagery.
  • Models human-level temperature based on spatial configuration of land classifications.


Graphic from Saitoh et al., (1996)

City of Portland

  • LiDAR-derived land descriptors
  • Required massive resources (data, time, money, and computational power)

Previous modeling of UHI has been created with remote sensing-derived land cover data (Voelkel, Shandas, and Haggerty, 2016); the intent of this project is to both remove feature classification from the workflow, and leverage freely available satellite data

Research Questions

Question #1:

Can vehicle traverse-based in-situ temperature readings, coupled with non-Thermal Infrared satellite bands, accurately model the Urban Heat Island Effect?

Hypothesis #1:

By converting individual imagery bands into regional descriptions of their own internal variability/texture, it is possible to accurately predict urban heat. Using the spectral structure that informs classification should sidestep the need to classify land uses in the UHI model.

Question #2:

Can a UHI model created in Portland be transferred to another city?

Hypothesis #2:

A Portland UHI will have some relation to other nearby cities, however will not be as powerful of a predictor as in the area that in-situ measurements were collected.

Study Area

Portland Metropolitan Area, Oregon, USA

[Validating with Eugene Metro, Oregon, USA]

Methods

Project Workflow

Overview:

  • Create regional descriptions of raster variability in the form of:
    • textons
    • multi-distance means of textons
    • multi-distance means original bands
  • Use these newly-created data sets as independent variables in a series of modeling exercises to predict the dependent variable: temperature.
  • Test transferability of the UHI model to other cities (with Eugene, Oregon as a test case)

Data

Sentinel-2 Bands

Band Name Wavelength
(nm)
Spatial
Resolution (m)
B2 Blue 490 10
B3 Green 560 10
B4 Red 665 10
B5 Red Edge 1 705 20
B6 Red Edge 2 740 20
B7 Red Edge 3 783 20
B8 NIR 842 10
B8a Red Edge 4 865 20
B11 SWIR 1 1610 20
B12 SWIR 2 2190 20

Data for Portland Metro and Eugene, Oregon taken on July 27, 2016. Acquired through Google Earth Engine.

Sentinel-2: 10m bands

Color infrared (Red=B8, Green=B3, Blue=B2)

Sentinel-2: 20m bands

Composite of Red-edge and SWIR bands

Texton Filter Bank

Textons are simply the result in raster format of a moving window analysis wherein the kernel filter weights optimize the exaggeration of texture and variability. Here is a regular kernel window, which would calculate the mean within a 5 pixel window:

##            [,1]       [,2]       [,3]       [,4]       [,5]
## [1,] 0.00000000 0.00000000 0.07692308 0.00000000 0.00000000
## [2,] 0.00000000 0.07692308 0.07692308 0.07692308 0.00000000
## [3,] 0.07692308 0.07692308 0.07692308 0.07692308 0.07692308
## [4,] 0.00000000 0.07692308 0.07692308 0.07692308 0.00000000
## [5,] 0.00000000 0.00000000 0.07692308 0.00000000 0.00000000

sum of weights adds up to 1, therefore this will calculate the mean.

However, texton filters are slightly more complex. Here they are represented as rasters (for visualization purposes). All 12 filters are used in the analysis.

Filters mimic those found in the The Leung-Malik (LM) Filter Bank, which are commonly used in robotic vision. These are used to create new versions of imagery that represent a ‘vocabulary’ of structural information.

Leung, T. & Malik, J. International Journal of Computer Vision (2001) 43: 29. doi:10.1023/A:1011126920638

In-Situ Temperature

Temperature Collection

  • 4 to 9 vehicles
  • Measurements every 1s
  • GPS synced with temperature measurement time
  • 3 time periods: 6am - 7am, 3pm - 4pm, 7pm - 8pm
  • Nighttime cooling is of greatest concern for human well being.
    • For now, only nighttime measurements will be used in modeling.

Processing

Buffering

Tobler’s First Law of Geography:

“Everything is related to everything else, but near things are more related than distant things.”

In addition to the textons, a moving window mean function was run on each band and texton raster (130 total rasters) to account for distance-based decay in correlation at 50m, 100m, 200m, 400m, and 800m.

Buffered Texton Rasters


Band 11, Filter #1

Value Extraction

For all 650 rasters, cell values at each temperature measurement point are collected. This results in a table with 651 columns (including observed temperature) and a row for every temperature observation.

Modeling Temperatures

Modeling Workflow

  • Cross-validation done with a 70/30 holdout
  • Random Forest: top 10 most important variables selected
  • Multiple Linear Regression: Random Forest variables, with VIF and p-value filtering
  • Support Vector Machine: Iterative Model selection (variations in gamma and cost)

Results

Random Forest


RMSE: 0.287°C

Top 10 Variables:

No texton rasters were highly significant!

Rank Band Buffer
(meters)
Name Wavelength (nm)
1 B2 800 Blue 490
2 B8A 400 Red Edge 4 865
3 B8A 800 Red Edge 4 865
4 B2 400 Blue 490
5 B2 200 Blue 490
6 B12 400 SWIR 2 2190
7 B8 800 NIR 842
8 B8 400 NIR 842
9 B7 400 Red Edge 3 783
10 B7 800 Red Edge 3 783


Band 2: Pavement    |    Band 8/8A/7: Vegetation    |    Band 12: Vegetation

Source: European Space Agency

Multiple Linear Regression


RMSE: 0.968°C

SVM

… still running!

  • Started on 3/8/2017 - this slide will update if/when it ever finishes.
  • Unlikely to be much better than Random Forest
    • Processing time alone makes it less feasible.

Transferability of RF


RMSE: 1.264°C

Discussion

Hypothesis 1

Question #1:

Can vehicle traverse-based in-situ temperature readings, coupled with non-Thermal Infrared satellite bands, accurately model the Urban Heat Island Effect?

Hypothesis #1:

By converting individual imagery bands into regional descriptions of their own internal variability/texture, it is possible to accurately predict urban heat.

Yes! However, the complex texton rasters had little-to-nothing to do with the models. Simple spectral texture descriptions did well in predicting urban heat.

Hypothesis 2

Question #2:

Can a UHI model created in Portland be transferred to another city?

Hypothesis #2:

A Portland UHI will have some relation to other nearby cities, however will not be as powerful of a predictor as in the area that in-situ measurements were collected.

Yes, but not enough to create a satifactory UHI prediction map. The Portland UHI model did explain over 25% of the variation in heat in Eugene, however further analysis could yield a more transferable model.

Exposure

Within the Portland Metro region, we see temperatures vary between 78.7°F and 93.8°F! Further research should analyze demographic issues of disproportionate exposure to excessive heat.

Conclusion

Textons

Unnecessary! Literature suggests potential for classification (though is silent on regression).

Transferability

Currently, the model is not transferable; however, as more data is collect, I expect to see the intra-urban predictive power increase.

Future Work

  • Test other time periods
  • Collect more field data

Questions?

Unlinked References

- Howard, L. (1820). The Climate of London: Deduced from Meteorological Observations Made at Different Places in the Neighbourhood of the Metropolis. In Two Volumes. W.Phillips.