Remote sensing and public policy

Definitions

What is remote sensing?

Information obtained through long-range observation (e.g. from satellites, aircraft)

Collection of raw imagery from the surface of the Earth
1. passive sensors: collect information on emitted light/radiation
  - examples: photography, infrared
2. active sensors: emit energy, collect information on reflected light/radiation
  - examples: radar, LiDAR
Image processing
1. raw images are georeferenced to ground control points
2. emitted/reflected light matched to specific spectral signatures
  - examples: types of vegetation, land cover, CO\(_2\) emissions
3. processed data are stored as pixels in raster datasets

Advantages:

remote sensing is sometimes cheaper and safer than direct observation
(e.g. hard-to-reach areas, conflict zones)
measurement is consistent across regional, national borders

Remote sensing \(=\) raster data

not all raster data are derived from remote sensing imagery
but all remote sensing imagery originates as raster data

Raster data structure \(\neq\) vector data structure

vectors store information in “attribute tables” (\(N\) features \(\times\) \(K\) fields)
rasters store information in a grid of pixels (\(N_R\) rows \(\times\) \(N_C\) columns)
- pixels are of constant size, shape, area
- each pixel represents a unique location
- each pixel contains just one value
  (e.g. precipitation, land use)
- size of pixels determines resolution
  (e.g. 1 meter, 1 km, 1 degree)
rasters usually have larger file size than vectors, but not necessarily more precision

Raster data structure

Applications

Many variables of interest to public policy originate as remote sensing imagery

weather (precipitation, temperature)
climate model forecasts
flooding depth and risk
active fires
night light emissions
elevation, slope, line of sight
pollution and air quality
cloud cover
vegetation indices
soil quality, fertility
land use and land cover (LULC)
built-up areas
population density (derived from above)

Example use case

But raster data were not (originally) built for social science and public policy applications

original policy purpose: military reconnaissance, damage assessment
original scientific purpose: natural sciences (e.g. geology, ecology, biology)
no sensor systems, spectral measurements were designed for dedicated monitoring of social, economic processes
reliance on indirect/proxy measures

Divergent data structures, approaches

social science: “vector view” of world (e.g. organize data into discrete units)
natural science: “raster view” of world (e.g. organize data into regular lattice)
integrating raster and vector data requires interdisciplinary cooperation

Social science prefers vectors

Raster data analysis

Rasterization and Vectorization

In social science and public policy, raster data integration requires that we either

Rasterize the vector data
- convert discrete features into continuous field
- examples:
  1. frequency/density of features
  2. presence/absence of feature
  3. distance to features
  4. assignment to feature
Vectorize the raster data
- summarize values of continuous field at each feature
- examples:
  1. zonal statistics (e.g. mean, max cell values)
  2. image tracing (e.g. of georeferenced maps – we covered this earlier)

Point-to-raster: suppose points are locations of 100 events (e.g. wolf attacks)

Point geometries

Option 1: count number of features in each raster pixel/cell

Point counts per cell

Pixels values are local frequency (number of points) or point density (number/area)

Local point frequency

Line-to-raster: suppose this line is an infrastructural object (e.g. road, power line)

Line geometries

Option 2: presence/absence of features at each raster pixel/cell

Line presence/absence per cell

Pixels values are indicators of whether an object is locally present/accessible

Local line access

Option 3: distance from feature to each raster pixel/cell

Distance from line to cell

Pixels values represent proximity

Local distance

Polygon-to-raster: suppose polygons are 4 administrative areas (e.g. districts)

Polygon geometries

Option 4: assign pixels to overlapping features (e.g. by center of cell)

Polygon assignment by centroid

Pixel values are polygon labels or attributes (e.g. assumed constant)

Local polygon assignment

Rasterization overview

These operations can be done on all types of vector data

count/density of points/lines/polygons
presence/absence of points/lines/polygons
distance to points/lines/polygons
assignment to points/lines/polygons (with tie-breaking rule)

But problem: why do this?

pixels are not meaningful spatial units for public policy
policymakers don’t think of the world as a “continuous field”
policy is made in discrete geographic jurisdictions, with well-defined borders
more common approach to analysis: convert raster to vector

Raster-to-polygon: suppose raster represents a continuous variable (e.g. elevation)

Study region with 32 continuous raster grid cells and 4 polygons

Raster cell values

Option 1: calculate zonal statistics (e.g. average cell values) for each polygon

Zonal statistics: mean cell values

Average cell values are added to attribute table for polygons

Mean cell values for each polygon

Same operation could be used to obtain maximum cell values…

Maximum cell values for each polygon

…or minimum values (or any other summary statistic)

Minimum cell values for each polygon

But what if raster represents a categorical variable (e.g. land use)?

Study region with 32 categorical raster grid cells and 4 polygons

Raster cell values

Option 2: reclassify raster to binary (e.g. 1 if land use “A”, 0 otherwise)

Reclassified raster

Calculate zonal statistics: percent of each polygon with cell values of “A”

Zonal statistics: value “A” as percent of overlapping cells

Percentages are added to attribute table for polygons

Percent “A” per polygon

Scale-dependence

Scale-Pattern-Process

Scale of analysis (spatial, temporal) impacts which patterns are observable
- these observations shape inferences we draw about underlying social processes
Processes drive patterns whose observation is scale-dependent
- some research questions require high spatial resolution:
  1. urban/neighborhood policy
  2. bomb damage assessment
- some research questions require high temporal resolution:
  1. emergency response
  2. weather forecasting
- some questions can be answered at low resolution (e.g. long-term, large-scale)
  1. economic development
  2. deforestation, changes in land use

Trade-offs

lower resolution (large pixels) \(=\) more information loss
higher resolution (small pixels) \(=\) higher collection, storage, computation costs

How scale impacts rasterization and vectorization