Repository logo
 

Development of Novel Image Analysis Approaches for Seaweed Discrimination – Species Level Study Using Field Spectroscopy and UAV Multispectral Remote Sensing

Files

Thesis(17.33 MB)

Date

2021

Authors

Supervisor

White, Lindsey
Case, Bradley

Item type

Thesis

Degree name

Doctor of Philosophy

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

Seaweeds play important roles in coastal ecosystems such as providing habitat, feeding grounds and improving water quality. It is crucial to map their distribution to quantify biodiversity and assess changes over time especially due to invasive species. The seaweed Undaria pinnatifida (Harvey) Suringar, native to north-western Asia, is one of the top 100 invasive species in the world and has become established across much of New Zealand (NZ), competing and co-existing with native seaweed species. Remote sensing is an efficient tool for mapping since current seaweed mapping practices in NZ such as snorkelling and SCUBA surveys can be time-consuming and do not cover large extents. Despite the invasive nature of U. pinnatifida, there was no spectral information available that would assist in remote sensing surveys in NZ. A hyperspectral library of common NZ native and invasive seaweed species was created to identify the key wavelengths that discriminated NZ seaweed species at both inter- and intra- taxonomic levels. The hyperspectral data of the native and invasive seaweed species collected from field survey were subjected to two supervised classification methods - Partial Least Square Discriminant Analysis (PLS-DA) for wavelength selection/classification and random forest for validating the wavelengths from PLS-DA. The seaweeds were separable at broad taxonomic level (red, green and brown seaweeds) with accuracies > 85% using PLS-DA. Some of the influential wavelengths identified were consistent with pigment absorption peaks unique to red and brown seaweeds. U. pinnatifida differed from native browns in the visible (574 nm) and near-infrared (716 – 721, 750 nm) region of the electromagnetic spectrum and the classification accuracies were 97.7% and 90.7% using random forest and PLS-DA, respectively. Variations in season or location may affect the spectral reflectance which in turn would affect the accuracy of mapping aquatic and terrestrial vegetation from remote sensing surveys. Such variations are widely studied in terrestrial plants compared to seaweeds. The hyperspectral data of the two commonly found New Zealand native seaweed species, Ecklonia radiata (C. Agardh) J. Agardh. and Carpophyllum maschalocarpum (Turner) Grev from four locations across four seasons were used to analyse spatial and seasonal effects on their spectral reflectance values using mixed-effects modelling. The modelling showed that season affects spectral reflectance of the seaweed species across the four locations, specifically in summer, which is likely due to the higher rates of photosynthesis. There are many studies on the effect of depth and turbidity on seaweeds at broad taxonomic level globally. However, a detailed study on the depth and turbidity effects on seaweeds in New Zealand at broad taxonomic and species level is lacking. The hyperspectral data of the two seaweed species, U. pinnatifida and E. radiata, at five depths and two turbidity levels were used in two different models for different purposes. Mixed-effects modelling that was used to understand the effect of depth and turbidity on the spectral reflectance values of the seaweed species showed depth significantly affected the spectral reflectance of the seaweed species compared to turbidity. Random forest model was used to assess the feasibility to discriminate the two seaweed species from each other across different depths within a turbidity level. Two sets of wavelengths were used as explanatory variables to assess the suitability of bands for discrimination – wavelengths that discriminated U. pinnatifida from rest of the brown seaweed species and wavelengths that matched the Micasense RedEdge-m sensor. This comparison would help understand if existing multispectral sensors would suffice or if it would be better to customise the sensor for the application. The former set of wavelengths (574, 716-718, 720-721, 750 nm) discriminated the two seaweed species from each other better with accuracy in the range of 57 – 87% and the accuracy increased with the depth. The overall accuracy of the discrimination of the two seaweed species from multispectral data was better at flight height of 30m (63%, kappa = 0.45) compared to that at flight height of 10m (60%, kappa = 0.38). This is the first study on the spectral variability of seaweeds due to season and location, globally. This research is a significant step towards mapping common habitat forming NZ native and invasive seaweed species using remote sensing. This study identified key wavelengths for discriminating emerged seaweeds (out of water) at species level using robust discriminatory models developed. A novel image pre-processing technique that reduced noise was implemented before image classification. The study is also the first to use 5-band multispectral UAV data to classify two submerged spectrally similar invasive and native brown seaweed species of New Zealand.

Description

Keywords

UAV, Remote sensing, Seaweed, Undaria, Season, Spectral signature, Location, Discrimination, Hyperspectral, Multispectral, Micasense rededge, Phantom 4 pro, Direct georeferencing, Random forest, Mixed effects model, PLS-DA

Source

DOI

Publisher's version

Rights statement

Permanent link

https://hdl.handle.net/10292/14260

Collections

Doctoral Theses