The document discusses the application of drones and artificial intelligence in earth science, emphasizing the use of UAVs for high-resolution data collection and analysis across various environmental and geological applications. It highlights technological advancements such as photogrammetry, immersive visualization, and neural networks for data interpretation, as well as the ethical considerations surrounding data management and accessibility. The ambition is to position Australia as a leader in drone-sensed data for industry and research, enhancing conventional survey methods with flexible and rapid data acquisition.

1. Drones and A.I. in Earth Science
S. Micklethwaite, S. Thiele, J. Sissins, A.
Spek, S. Vollgger, T. Drummond
Thanks to G. Dering
300x300 m,
25 mm pixel resolution

2. UAV
Platform
Training:
Pilots licencing,
Data processing,
Data analytics
Equipment:
UAVs (multirotor, fixed wing, VTOL,
heavy lift, long-duration etc),
Sensors (high-end digital cameras,
LiDAR, mag, infrared, spectrometers,
Temp-gas sensors, grav, pXRF etc),
Service:
Qualified pilots,
Data processing,
Data analytics
eResearch
(MeRC)
Spatial processing:
Nectar Cloud,
MASSIVE3,
Distributed servers
Interactive Spatial
Visualisation:
MIVP (CAVE2, headsets etc)
3D
Printing
Driving
Developments
Smart analytics (e.g. CoE
Robotic Vision, Sensilab)
UAV developments (e.g. Swarm
robotics lab, Monash UAS Team)
Strategic collaborations (e.g.
UTas sensor calibration, USGS,
GA etc)

3. Disruptive Technology:
• High resolution data, large areas
• Integration of data across multiple
scales, ground-UAV-satellite
• Flexible & rapid – e.g. time series
analyses
• Interoperable – multiple data types for
different applications in a single survey
• Remote locations
Dangerous Spaces
Inglewood Mine (1890),
rehabilitation.
3. VTOL platforms

4. ARC LIEF 2017 “UAV Sensing and Data
Discovery for a Changing Planet”

5. Photogrammetry: The workhorse
1. Feature matching across photos
(stable under viewpoint and lighting
variations and generates a descriptor
for each point based on its local
neighborhood; similar to SIFT).
2. Solving for camera position,
orientation, and distortion
parameters + sparse point cloud
(greedy algorithm refined later using a
bundle-adjustment algorithm).
3. Dense surface reconstruction
(Exact, Smooth and Height-field
methods are based on pair-wise depth
map computation, while Fast method
utilizes a multi-view approach).
4. Texture mapping (parametrizes a
surface possibly cutting it in smaller
pieces, and then blends source photos
to form a texture atlas).
Bemis et al., 2014, JSG; Vollgger & Cruden 2016, JSG;
Smith et al., 2016, PPG

6. • Impressive efficiencies but computationally expensive with
increasing surface area or GSD
• E.g. 2x images, 4x processing time
• A 8 km2 region at 5 cm pixel resolution, 3-5 days
• How to maximise analysis of this data in timely
manner?
Type of
lens (mm)
Number of
photographs
Quality Number of
Polygons
(Approx.)
Total
processing
time
(hours)
28 150-250 High 90000 5-8
Medium 40000 4-6
50 500-700 High 300000 30-35
Medium 160000 20-25
low 50000 10-15
105 600-700 Medium 250000 30-35

7. VISION / AMBITION
 Australia the first drone-sensed nation (cm-scale)
 Pre-competitive data release for industry, environmental
management, education & research
 Conventional survey & remote sensing techniques at ultra-
high resolution and flexibility (time-series, rapid response etc)
 Next gen “UNDERCOVER” techniques (minerals and water
resources)

8. + Inter-operability: e.g. lightweight
T, CO2, humidity, pH etc
SENSOR DEVELOPMENTS:
e.g. thermal: street
scene
e.g. TMI - fluxgate
e.g. LiDAR for biomass
& landscape
e.g. L-band radar

9. Middlemiss 2016, Nature 531, 614-617 MEMS gravimeter:
• Microelectromechanical
system (MEMS) device with
a sensitivity of 40 microgal
per hertz1/2.
• Few cubic centimetres in
size.
• Can measure the Earth
tides, revealing the long-
term stability of the
instrument.
• High resolution gravity
surveys, possibly over large
areas. Initially, ground-
hopping.

10. ANALYTICAL & VISUALISATION DEVELOPMENTS:
Immersive viz.
3D feature maps & measurement
Semantic vision & rapid
outcrop interp of big data
dolerite
tonalite
gabbro

11. Thiele et al., in prep, 2017
Rapid fault, fracture,
lithology contact
mapping
Least-cost path solvers.
Geologist remains
involved in interpretation.
Future – NN automated

12. Thiele et al., in prep, 2017
Now implemented in QGIS

13. Process:
• Training a neural network to classify lithological classes
• Semantic segmentation problem
• Utilise methods to reduce overfitting and increase accuracy
• Data augmentation
• Colour space transforms
Progress:
• Tested a number of methods
• Currently achieving 70 - 75 % pixel
accuracy
• < 3 % of manual interpretation time

14. Vegetation
Dolerite
Tonalite
Background
Water
Aplite
Dacite Gabbro

15. DISCUSSION POINTS: PRACTICALITIES & ETHICS
Data storage, management coupled to virtual analytical pipelines (processing
and analyzing large to ‘big data’ on researchers & industry desktop/laptops).
Our objectives at Monash are not storage but data discovery.
• Begun to seed a virtual drone analysis environment under the
UNDERWORLD Virtual Lab
Nation-wide open data, or only partially open?
• Exploration leases with existing in-house high res magnetic data
suddenly supplanted and pre-competitive
• Cropping, where the data contains signals that can contain potentially
sensitive information on crop health
• How long is confidential information held, when collected on public
money?
Need to accelerate development of an ethical and legal framework
• National Parks and native title - Drones offer a potential solution to issues
of access without ground disturbance
• Accelerate the impact and reach of the platforms