How Many Hippos (HOMHIP): Algorithm for automatic counts of animals with infrared thermal imagery from UAV
Communication presented at GeoUAV Worshop in the frame of the ISPRS Geospatial Week 2015
La Grande Motte, 2nd October 2015
Ir Simon Lhoest
SIMON LHOEST, JULIE LINCHANT, SAMUEL QUEVAUVILLERS, CÉDRIC VERMEULEN, PHILIPPE LEJEUNE Forest Resources Management Axis, Biosystems engineering (BIOSE) University of Liège – Gembloux Agro-Bio Tech
2
Introduction Methodology Results Conclusion Discussion & Perspectives
Hippopotamus amphibius L.
• Vulnerable species • Groups : 10 to 200 individuals • In water during the day • Drawbacks in classic monitoring methods
Objective
Construction of an algorithm for the automatic count of hippos groups from thermal infrared images acquired by UAV
Integration in the open source QGIS software
3
Introduction Methodology Results Conclusion Discussion & Perspectives
Why doing monitoring ?
1. Quantify the impact of poaching Natural growth = 6%/year Reinforcement/moving of monitoring troops
2. Complete demographic analyses More precise previsions of probable populations evolution
3. Evaluate the impact of hippos on their environment Erosion? Soils compaction? Food disponibility? Competition between herbivorous species?
4
Introduction Results Conclusion Discussion & Perspectives
• Detection + counting • Mainly for birds colonies
+ Save time & efforts + Easy to use + Generally reliable results - More difficult than human interpretation
Automatic procedures
© Chabot, 2009
Canada geese Snow geese
Criteria : Aggregation of individuals (but too close together) High contrast (animal – background) Sufficient image quality
Methodology
© Julie Linchant
5
Introduction Results Conclusion Discussion & Perspectives
Methodology
Falcon Unmanned© UAV
Camera Tamarisk 640 (λ = 8-14 µm)
Garamba National Park (DRC)
Flight heights : 38 – 155 meters (Pixel GSD = 3.8 – 15.5 cm)
14 flights, 2 study sites
Manual extraction of images
Raster values = temperature
© Basile Luse Belanganayi
6
Introduction Results Conclusion Discussion & Perspectives
Methodology
• Reference data = visual and manual digitization of hippos 3 categories of hippos : o Completely emerged animals (CEA)
o Pairs of Polygons of Single Animals (PPSA)
o Nearly Immerged Animals (NIA)
• 37 extracted images • 2126 digitized polygons = 1856 hippos
7
Introduction Results Conclusion Discussion & Perspectives
Methodology
• Python script for all the geoprocessing • QGIS application • Graphical User Interface
8
Introduction Results Conclusion Discussion & Perspectives
Methodology
Algorithm steps
Relative coordinate system (pixel unit)
Local maxima
detection
Radius
Threshold
Minimum distance
Local maxima
(Loc_max)
Clipped image
(Cl_im)
9
Introduction Results Conclusion Discussion & Perspectives
Methodology
Algorithm steps Clipped image
(Cl_im)
Spatial
join
Local maxima
(Loc_max)
Local maxima
detection
Radius
Threshold
Minimum distance
Isolines
interval
Isolines
creation and
polygonization
Isoline polygons
(Iso_polyg)
10
Introduction Results Conclusion Discussion & Perspectives
Methodology
Algorithm steps Clipped image
(Cl_im)
Spatial
join
Local maxima
(Loc_max)
Local maxima
detection
Radius
Threshold
Minimum distance Isolines
interval Isolines creation
and polygonization
Isoline polygons
(Iso_polyg)
Polygons
selection
Minimum area
Maximum area
Minimum perimeter
Maximum perimeter
Emerged parts
(Em_parts)
y = 636631x-1.834
R² = 0.8423
0
100
200
300
400
500
600
700
800
0 50 100 150 200 Max
imal
surf
ace
(pix
el u
nit
²)
Flight height (m)
y = 195.87x-0.898
R² = 0.4713
0
2
4
6
8
10
12
14
16
0 50 100 150 200
Min
imal
surf
ace
(pix
el u
nit
²)
Flight height (m)
y = -0.9051x + 152.76
R² = 0.8805
0
20
40
60
80
100
120
140
0 50 100 150 200 Max
imal
per
imet
er (
pix
el u
nit
)
Flight height (m)
y = -0.0402x + 11.46
R² = 0.4551 0
2
4
6
8
10
12
14
16
0 50 100 150 200 Min
imal
per
imet
er (
pix
el u
nit
)
Flight height (m)
Each one contains at least 1 LM
11
Introduction Results Conclusion Discussion & Perspectives
Methodology
Algorithm steps Clipped image
(Cl_im)
Spatial
join
Local maxima
(Loc_max)
Local maxima
detection
Radius
Threshold
Minimum distance
Isolines
interval Isolines creation
and polygonization
Isoline polygons
(Iso_polyg)
Polygons
selection
Minimum area
Maximum area
Minimum perimeter
Maximum perimeter
Emerged parts
(Em_parts)
Polygons
classification
Completely
emerged
animals
(CEA)
Pairs of
Polygons of
Single
Animals
(PPSA)
Nearly
Immerged
Animals
(NIA)
Area limit between CEA and NIA
Maximum distance between PPSA polygons
Maximum |Δϑ| angle
Maximum |α-ϑ0| angle
y = 20536x-1.421
R² = 0.7828
0
20
40
60
80
100
120
140
160
0 50 100 150 200
Surf
ace
(pix
el u
nit
²)
Flight height (m)
12
Introduction Results Conclusion Discussion & Perspectives
Methodology
Algorithm steps
Criteria for Pairs of Polygons corresponding to a Single Animal :
1. Polygons size Area < CEA polygons
2. Proximity Maximal distance between polygons = fct (flight height)
3. Alignment - Minimum Bounding Boxes orientation characteristics - 2 angles computed : |Δϑ| &|α-ϑ0| maxima values
y = -0.1957x + 32.068
R² = 0.7626
0
5
10
15
20
25
30
35
0 50 100 150 200
Max
imal
dis
tance
bet
wee
n
centr
oid
s (
pix
el u
nit
)
Flight height (m)
13
Introduction Results Conclusion Discussion & Perspectives
Methodology
30% of images (rainy season)
Mean total error +3.9%
Mean NIA error +28.5%
Mean PPSA error -55.4%
Mean CEA error +20.2%
Minimal total error -4.9%
Maximal total error +12.9%
NIA (Nearly
Immerged Animals)
PPSA (Pairs of Polygons
corresponding to Single Animals)
CEA (Completely
Emerged Animals) Total
Automatic –
manual correlation 0.47 (p = 0.149) 0.14 (p = 0.674) 0.42 (p = 0.197) 0.98 (p < 0.001)
14
Introduction Results Conclusion Discussion & Perspectives
Methodology
Image processing Influence of input parameters! Difficult to deal with hippos very close together Hippos categories tend to compensate in counts, but … necessary improvement! Automation of the masking process & selection of images?
Practical recommendations of flight Rainy season (April to November) Not in the end of the afternoon Avoid fog
Exploitation of results Some hippos (under water) are not detected Necessity of a correction factor: 1.25, according to Delvingt (1978) & Lhoest (2015)
Sensors and UAV improvement Combination of several sensors? Use of high resolution infrared photos? Multicopter platform?
Rainy season
Dry season
15
Introduction Results Conclusion Discussion & Perspectives
Methodology
• Such algorithms are really important to deal with the huge amount of UAV data:
Great perspectives
Save time
Easy to use (open source software)
Standardized & reproducible procedure avoid operator effect in counts!
Adaptable parameters
• Improvement still necessary: Automation of pre-process
Determination of optimal input parameters values
+
-
Thanks for your sustained attention!
© Photos : Simon Lhoest, except other mentions
Simon Lhoest [email protected]