mcneills_igarss2011_penguins.pdf

Semi-automated penguin countingfrom digital aerial photographs

S.J McNeill K Barton P Lyver D Pairman

Landcare Research New Zealand

Motivation

Understanding changes in penguin population is important,as these can be used as indicators of anthropogenic andfoodweb e�ects

Aerial photography is used in the Ross Sea (Antarctica) tocapture a reliable count of Adélie nesting penguins

From 1981, the Ross Sea area (158o�175o E) has beensurveyed annually

There are many di�culties in achieving this census count:

Timing is critical,Ground counting is di�cult or impossible,Counting using prints is di�cult to control and validate.

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Objectives

Determine if it is possibleto reliably detect Adéliebreeding penguins inimages

Generate software to(semi-)automate thecensus process.

Test, using an �expert�,and optimise interactivity.

Adult Adélie penguin (Pygoscelis adeliae)

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Adélie penguins

The most abundant and widespread Antarctic penguin

10 million Adélie make up 80% of the Southern Ocean birdbiomass

38% of all Adélie penguins are found in the Ross Sea

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Image capture

Images captured using a hand-held camera through the opendoors of a helicopter and/or C-130 Hercules

Hasselblad H1D with a Phase One digital camera back

Image size 5440× 4080, 3-bands natural colour, TIFFEXIF data provides date/time and aperture informationTypical ground resolution better than 0.5 m

Ten representative images were selected for analysis

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Full-scene example

5440× 4080 full-scene

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Sub-scene example

870× 510 sub-scene

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Analysis

Human detection of breeding Adélie not straightforward

There are many similar-looking objects in the imagesProposed revised approach:

Detect the distinctive area of the colonyOnly count penguins within colony areaProvide software features to easily add/delete penguins

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Colony/background discrimination

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Colony & penguin detection

Background is largely monochromatic

Colony area covered in guano and has a red excess overgreen or blue, with higher saturation

Use linear discriminant analysis to separate colony frombackground, based on:

Natural colour counts (RGB) converted to hue, saturation,lightness (HSL) space values,Two-way interactions of HSL space values,Aperture setting.

Classi�cation followed by morphological opening and closingde�ne the colony area

Penguins detected as dark local minima within colony area

Penguin �objects� pruned to upper threshold of circularityP2/ (4πA) to remove long thin objects

Adopt the centroid of the surviving objects as penguins

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Original image (350× 250)

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Detected colony

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Cleaned colony

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Candidate penguin locations

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Overlaid penguins

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Editing facilties

Detection procedure does not count all �real� penguins

False penguins counted

Non-breeding penguins within colonyPenguin shadows or spurious dark objects

True penguins missed

Breeding penguins outside colonyPenguins indistinct compared to surroundings

Editing facilities required:

Overlap between photographs requires group deletionsAdd or delete individual penguinsCheck that penguins are not double-countedRecord of editing steps maintainedNumber of editing steps requires �single-click� operation

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Implementation

Software written in Matlab 2010b, deployed with Matlabcompiler

Census results stored for each captured image in a small �le

Deployed for testing phase to a penguin ecologist

Second development phase to �x faults and improveinteractive response:

Reduce memory overhead for each counted penguinReduce keystroke e�ort for additions/deletionsAdd ability to count penguins within non-guano stained area

No problems reported after second phase deployment

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Editing software

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Bootstrap colony classi�cation rates

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Colony classi�cation rates

Accurate colony delineation is very important

Requirement is for high true positive, low false negative rates

About 5% of images give poor results:

Due to very poor colony/background distinctionNo clear reason for this poor result

CF001669 CF001720

IGARSS-2011, 25-29 July 2011, Vancouver, Canada

Conclusions

Semi-automated penguin counting is a pragmatic approach

Laborious counting automated; �ne editing left for an expert

Software allows editing, maintains counts, stores results

Emphasis is interactive productivity

Acknowledgements

Ministry for Science and Innovation (funding).Antarctica New Zealand (funding and logistics).

Helicopters New Zealand (�ying).Squadron 40, Royal New Zealand Air Force (�ying).

IGARSS-2011, 25-29 July 2011, Vancouver, Canada