1. Satellite imagery analysis for
Sustainable Development Goals:
requirements and spatial
approaches
Jacinta Holloway
Queensland University of Technology (QUT), Brisbane,
Australia
ARC Centre of Excellence for Mathematical and Statistical
Frontiers
2. Presentation summary
• Satellite images are a useful and low cost data source.
• 3 minimum requirements to measure SDGs using satellite
images.
• If validation data is expensive or unavailable, spatial data and
methods can help.
3. United Nations (UN) Global Working Group on
Big Data for Official Statistics
• Created in 2014.
• International collaboration to
investigate benefits and challenges of
Big Data sources.
• Focus on monitoring and reporting
on Sustainable Development Goals.
5. Why satellite imagery data?
• Range of freely
available sources.
• Relatively long time
series of high quality
imagery available.
5Source: USGS (2015).
6. Why satellite imagery data?
•Can be combined
with other data
for new insights.
• Useful to create
visualisations.
6
Source: Presentation by Ms Flora Kerblat, CSIRO. Workshop on
Satellite Imagery for Official Statistics
Australian Bureau of Statistics, Canberra,
30th January 2017.
7. Minimum requirements to measure SDGs
from satellite images
1. Indicator can be measured from satellite images.
2. Analysis-ready satellite images are available.
3. Access to validation data and/or spatial information.
8. 1. Indicator can be measured from satellite images.
Crop area estimation, crop yield, land cover classification, extreme weather event detection.
Indicator 2.4.1 Proportion of agricultural area under productive and sustainable agriculture.
Water quality monitoring.
Indicator 6.6.1 Change in the extent of water-related ecosystems over time.
Indicator 6.3.2 Proportion of bodies of water with good ambient water quality.
Forest cover monitoring, deforestation detection.
Indicator 15.1.1 Forest area as a proportion of total land area.
Indicator 15.3.1 Proportion of land that is degraded over total land area.
Remote sensing application and indicator
9. Water quality in Ghana
Bui Reservoir
Feb 2013 to Dec 2016
Maximum Total Suspended
Matter (TSM)
Source: Killough, 2017.
11. 2. Analysis-ready satellite images are available.
US Geological Survey Global Visualisation tool
(Source: USGS, 2018)
12. 3. Access to validation data and/or spatial
information.
•Validation data includes directly collected and survey
data.
•Expensive to collect and access.
•When validation data is limited or not available,
spatial data can improve statistical modelling.
13. Why does spatial autocorrelation matter?
In Beltrá’s photograph, there is forest,
agricultural land and cleared land.
Standard models treat all the pixels as
independent.
Spatial autocorrelation, and visual
inspection, tells us observations are
not independent.
Source: Beltrá (2017).
14. Spatial autocorrelation
Spatial autocorrelation can be expressed as
Γ𝑖𝑗 = 𝑖=1
𝑛
𝑗=1
𝑛
𝑊𝑖𝑗 𝑌𝑖𝑗
Spatial autocorrelation, Γ, between location 𝑖 and all other
sites 𝑗 is given by two matrices, W and Y (Getis, 2010).
15. Methods: decision trees
Source: Garni et al., (2014)
𝑥 ∈ 𝑅𝑗
𝑥 ∈ 𝑅𝑗
𝑥 ∈ 𝑅𝑗
𝑥 ∈ 𝑅𝑗 𝑥 ∈ 𝑅𝑗
𝑅𝑗
𝑅𝑗
𝑅𝑗 𝑅𝑗 𝑅𝑗 𝑅𝑗
The variables, 𝑅 include MSAVI,
Texture, Brightness Index, Slope
and Moisture index. A constant
set value of each variable, 𝛾𝑗 are
the split points that separate
observations into different
classes.
A decision tree is given by
𝑇 𝑥; Θ =
𝑗=1
𝐽
𝛾𝑗 𝐼 𝑥 ∈ 𝑅𝑗
Source: Hastie et al., 2008, p.356.
16. Spatial decision trees
• There are a small number of spatial decision tree models.
• These are implementations of random forest algorithm.
• Some methods can only handle 100 observations, but to
classify satellite images requires thousands or millions of
pixels.
17. Application example: filling in missing image data
Location of the Injune study area in Queensland.
Satellite image of Injune
with data missing due to
cloud cover
18. Application example: filling in missing image data
Fit spatial
model
Satellite image with
missing data due to
cloud cover
Satellite image
completed with spatial
model predictions
19. Summary
• Satellite images are a useful and low cost data source for
measuring SDGs.
• To measure SDGs using satellite images, there are 3 minimum
requirements:
1. Indicator can be measured from satellite images.
2. Analysis-ready satellite images are available.
3. Access to validation data and/or spatial information.
20. • Where validation data is
expensive or unavailable, spatial
data and methods can help.
• Can’t allow lack of validation
data to prevent measuring
SDGs.
• New spatial machine learning
methods continue to be
developed and can measure
SDGs.
Summary
Editor's Notes
Explore feasibility for producing official statistics based on Big Data sources.
Costs are a main barrier to developing countries measuring SDGs.
Imagine this is captured in a satellite image. Within the forest area, the pixels next to each other are likely to also be forest rather than grass, because they are close together. Equally, these forest pixels are unlikely to be bare earth, which is far away from them.
Standard models treat all the pixels as independent; a pixel in the forest area is just as likely to be bare earth or grass as it is forest.
MSAVI is a modified soil-adjusted vegetation index