Automated recognition of harp seal pups in aerial photographs

Woodrow, Jennifer Renee (2008) Automated recognition of harp seal pups in aerial photographs. Masters thesis, Memorial University of Newfoundland.

[img] [English] PDF - Accepted Version
Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission.

Download (5Mb)

Abstract

Harp seals, Pagophilus groenlandicus, are the most abundant pinniped in the Northwest Atlantic. The Canadian and Greenland hunt for the harp seal is the largest marine mammal harvest in the world. To ensure a sustainable yield, it is important to monitor abundance and population trends on a regular basis. In addition to species management, assessing harp seal population is important in estimating the consumption of prey by the species. To estimate the total population of harp seals, the Canadian Department of Fisheries and Oceans (DFO) uses a population model that combines pup production estimates, pregnancy rates, and age-structured removals. -- Currently, the number of harp seal pups are estimated by conducting visual and photographic aerial surveys over whelping concentrations. A fixed-wing aircraft, equipped with a large format metric mapping camera with motion compensation, is used to take black-and-white photographs of whelping areas. To count seal pups, manual analysis of aerial photographs is performed by trained scientific personnel with extensive knowledge of harp seals and their environment. This process can take many months and involve several people. While extensive measures are taken to ensure the most accurate pup count, manual identification of seal pups is not always conclusive. This thesis attempts to address these issues by developing image processing and pattern recognition tools that automatically detect and classify harp seal pups in digitized aerial photographs. Automating this process will reduce the amount of time required to compute population estimates and potentially improve the accuracy of pup counts. -- The first step in the pattern recognition algorithm is to divide the large digitized aerial images into several sub-images for further analysis by the image processing and classification tools. The objective of the image processing algorithm is to segment and isolate potential harp seal objects to be used in pattern classification. The rigorous image processing component uses a combination of techniques including contrast stretching, adaptive thresholding using between-class variance, and a "cleaning" algorithm that employs edge detection, line dissection, and removal of objects based on size constraints. In addition, this thesis proposes a unique segmentation procedure called Isolate Connected Components that separates connected objects with minimal distortion to object shape. -- The image processing routine calculates nineteen features for each segmented object. Features are optimized using three different methods: scaling the data, Principal Component Analysis, and kernel whitening. One-class classification methods use these features to identify an object as 'seal pup' or 'not seal pup'. Two one-class methods are considered in this research: Parzen density estimation and Support Vector Data Description (SVDD). Optimal classifier parameters are determined by maximizing the Area Under the Receiver Operating Characteristic Curve (AUC). It is shown that the Parzen method performs better than the SVDD with an 82% success rate on test data.

Item Type: Thesis (Masters)
URI: http://research.library.mun.ca/id/eprint/9965
Item ID: 9965
Additional Information: Includes bibliographical references (leaves 122-132).
Department(s): ?? ComptSci ??
Date: 2008
Date Type: Submission
Library of Congress Subject Heading: Aerial surveys in wildlife management--Automation; Image processing; Pattern recognition systems; Seal populations--Estimates--Automation.

Actions (login required)

View Item View Item

Downloads

Downloads per month over the past year

View more statistics