This document discusses using machine learning to classify Antarctic sea ice and count seal populations from very high-resolution satellite imagery. The goals are to 1) automatically classify ice into types, 2) develop a seal detection method, and 3) explore ecological questions about seal habitats. Challenges include the large study area and complexity of classifying sea ice. Next steps involve collaborating to refine ice and seal classification techniques to enable robust population estimates and identify high priority habitats for conservation amid climate change impacts.
2. Outline
• Introduction and background
• British Antarctic Survey
• Wildlife Monitoring from Space
• Antarctic seals
• The big data problem
• Data
• Study site
• Sea ice and seal counts
• Goals
• Future steps
• Questions
3. Seals or Space?
Guess whether these cosmic noises
originate from Antarctic seals or
outer space.
8. British Antarctic Survey (BAS)
• World leading Polar
Research Centre
• Interdisciplinary
research within
regions of major
scientific and societal
urgency
• Wildlife From Space
• Power of Poop -
Doubled Emperor
Penguin population
estimate!
9. We hope to use innovative data science to understand
Antarctic ice seals
• These seals live and breed in the
pack ice zone, a region that is
notoriously difficult to monitor
• Although these seals have some of
the largest populations of any large
mammal, habitat preferences, true
population size and population
trend estimates are very poorly
understood
• As the sea ice on which they live will
reduce in coming decades due to
climate change, an understanding of
their habitat relationship and a
baseline estimate of population is
critical
10. Ice Seals = Southern Ocean Ecosystem Health Indicators
• Keystone species of the Antarctic foodweb
(feeds seals, penguins, seabirds, whales,
squid, fish etc…)
• Krill is tied to the seasonal cycles of sea ice
and primary productivity
• Knowledge of krill distribution, abundance,
trends is important for monitoring the
status of the Southern Ocean ecosystem
• Monitoring krill is extremely difficult
• Threat from changing sea ice conditions and
growing krill fisheries
11. Past Seal Surveys: Costly, Difficult and Restricted
● Population trends are unknown:
● Wide range of population estimates…● Dynamic and inaccessible habitat
● Seals are widespread
● No standardised census techniques
● Costly, limited coverage and dangerous
● 2 - 70 million crabeater seals
● 200,000 - 1 million Weddell seals
13. The aim of the ML research is twofold:
• We have a large amount of
very high resolution satellite
imagery, but we need an
automated method of
classifying the sea ice habitat
and the associated seals
• Using machine learning, can
we:
1. Classify the sea ice in the
imagery into different ice
types
2. Develop a method that
automatically finds ice seals
in the imagery
seals
seals
Pressure
ridges
Pressure
ridges
Seals are visible as the black dots surrounding the ridges.
Pressure ridges are caused by sea ice floes pushing up against one another
14. Seals
A-B) Detect Seal Habitat Hotspots at a broad scale
using 10-15 m satellite imagery (Landsat / Sentinel-2)
A BA
C
D
C-D) Download VHR imagery (0.31-0.5 m) and run
seal detection
Habitat Hotspots: targeted surveys and dynamic conservation
16. Data Overview
• Very high-resolution (VHR)
satellite images at sub-meter
resolution from WorldView-3
and WorldView-2 satellites
• For the two survey sites we
have satellite imagery
depicting a variety of sea ice
conditions
• Antarctic Peninsula:
• Panchromatic data at
0.50 m resolution
• Signy Island:
• Panchromatic images at
0.3 m and multispectral
images at ~ 1 m
Signy Island
Antarctic Peninsula
17. Data Overview: Seal counts
• manual seal counts
• Seals are found along coastal headlands,
thin pack ice, polynyas and the thick fast
ice. What drives the differing distribution
of these seals?
• Crimson blood – can we use multispectral
data to locate seas?
Very High
Resolution (VHR)
imagery of
Weddell seals and
their pups at
30cm resolution
(Signy Island)
A
B
C
B
Blue stars = Weddell seals,
Yellow stars = crabeater seals
blood
seals
18. Data overview: sea ice features
• Sea ice in satellite imagery -
easy to spot, difficult to
classify.
• A complexity of forms, shapes
and sizes.
• Tested brightness threshold
• Multi-scale levels of
classification:
1. Broad sea ice types (open
water, thick sea ice, thin
brash ice)
2. Small-scale sea ice features
(ice edge, ice floe size, leads,
cracks etc)
1. Very High-
Resolution satellite
sea-ice image
2. Sea ice classified into
broad-scale type (open
water, thin brash ice,
thick sea ice)
3. Sea ice classified by
features (individual ice-
floes, leads, cracks,
polynyas)
19. UAV Surveys
UAV Fieldwork
• Ground-truth images
• Detection based on temperature,
size, and shape of thermal
signatures
CAMMLR Methodology
• Create standardised UAV survey
protocol for repeated surveys of
krill-dependent species
Equipment
• Prion-3 UAV
• PhaseOne 50 MP visible camera
• MicroCAM3 thermal camera
• STS spectrometer
21. Goals
1. Automatically classify sea ice into open water, thick sea
ice and thin sea ice
2. Classify sea into broad categories: Pack ice, fast ice and
open water
3. Classify sea into small scale features: ice floe size,
cracks, pressure ridges, polynyas
4. Build seal detector
5. Explore ecological questions – what is average seal
colony size, what is the distance between seals, do
seals like the ice edge or middle of the ice floe, what
size ice floe do seals prefer…?
6. First VHR sea ice data, first robust population counts,
NRT habitat hotspots for dynamic conservation, IUCN
red list…
A. Sea ice classification B. Seal detector
C. Habitat model
Good habitat
Bad habitat
22. Seal counting
Can pre-trained CNN developed for 30x30 cm seal
counts work on lower res 50x50 cm data?
Off the shelf CNN ill-equipped - Poor performance.
Future: finetuning of the parameters using lower-
resolution imagery to avoid building 50 cm CNN
from scratch.
23. Ice classification
• Satellite image masking
• Distance to open water (Do seals prefer
to be close or far from open water).
• Distance to Brash (Do seals like to be
close to regions of brash ice to access
water, or far away to limit predation)
• Brash Intensity (What brash conditions
do seals prefer? This is something that
may easily alter with climate change in
the future)
• Ice size (Do seals prefer 4x4 m or 10x10
m ice floes, for example? Ice size
classified as being <2 m or >2 m)
24. Ecological Modelling
• Identify which environmental features
influence the density of seals on the ice.
• Test if the clustering pattern of the seals
locations is caused by the geography or by
individuals preferences.
25. Issues and future steps
• Scale – sea ice extends 18 million km2 over Southern Ocean,
creating a dynamic surface roughly twice the size of Europe.
• Discriminating Weddells found within crabeater aggregations and
from foraging elephant seals – look into different classification
techniques for different species.
• Sea ice – easy to spot, difficult to classify.
• Determine level of sea-ice classification complexity required for
seal habitat modelling.
• Use momentum from Hackathon to keep the ball rolling!
• Collaborate with DAMPT and others.
• Look into potential for fixed-wing and rotary UAV and the
application of aerial spectrometers.
27. Collaborator call!
Looking to combine my domain expertise as an
ecologist with data scientists to exploit state-of-art
satellite imagery for high impact conservation!