Wetland ecotones: testing remote sensing techniques to map ecotones in a Fynbos embedded wetland

Date
2021-12
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Various researchers starting as early as 1903, have developed many definitions of an ecotone (Clements 1905; Livingston 1903; Odum & Barrett 1971). The definition by Holland (1988) described ecotones as zones of transition between adjacent ecological systems, having a set of unique characteristics defined by space and time scales, and by the strength of interactions between adjacent ecological systems (Holland 1988). This definition paves the way for research that may exemplify various aspects of landscape ecology and spatial heterogeneity. Although a niche of high scientific interest, ecotonal research is very understudied, especially research on using Remote Sensing to identify and map fine-scale wetland ecotones. A bibliometric analysis and literature review showed that limited research has been conducted on wetland ecotones in southern Africa, however with sufficient literature covered on wetland delineation, classification, and mapping. Wetlands which are highly dynamic and considered moving entities in a landscape due to their varying hydroperiods, are especially challenging to map. Two main experiments were carried out both of which used Machine Learning (ML) algorithms namely Random Forest (RF) and the naïve Bayes classifier. The aim of the first experiment was to review and test remote sensing techniques to accurately identify and map distinct vegetation communities within the Du Toits River wetland, Western Cape South Africa. The second experiment was then to use probabilistic classification measures to map and characterize the ecotones prevailing in a fynbos embedded wetland ecosystem. The study used freely available satellite imagery namely Landsat 8 Surface Reflectance Tier 1, and Sentinel-2 MSI: MultiSpectral Instrument, Level-2A, obtained from the United States Geological Survey (USGS) through open-source resources such as Google Earth Engine (GEE). This research suggests that Random Forest (RF) classifier showed great potential in accurately mapping landcover, specifically four distinct and dominant vegetation types within the wetland namely Prionium serratum, Psoralea pinnata (referred to as palmiet wetland vegetation), a condensed group of Pteridium aquilinum, Restio paniculatus and Merxmuellera cincta (referred to as Sclerophyllous Wetland Vegetation), and Temporary Wetland Fynbos. RF results showed little spectral confusion between classes and produced moderate to high overall accuracies for classifications run through both the winter and summer seasons. The efficacy of using the fuzzy logic i.e. supervised probabilistic measures to identify and map ecotones in a spatially heterogenous landscape was showcased. Probabilistic mapping and fuzzy graphs showed complex and diverse ecotones within the wetland. It was evident that clear ecotones in the form of rapid and sharp high probabilities of one vegetation type intersected and replaced another. These ecotones may provide useful information about wetland ecosystem functioning and how vegetation zones may contribute to wetland ecosystem services (e.g. flood attenuation and carbon storage). Using a per-pixel based approach to map ecotones is highly useful as ecotones are more complex in reality and mapping them as single vector lines is not useful nor accurate. Although this study aimed to identify and map fine-scale wetland ecotones, further research using even finer scale data and in-depth field analysis that specifically focuses on the identified and mapped ecotonal areas will be significant.
AFRIKAANSE OPSOMMING: Verskeie navorsers wat so vroeg as 1903 begin het, het baie definisies van 'n ekotoon ontwikkel (Clements 1905; Livingston 1903; Odum & Barrett 1971). Die definisie deur Holland (1988) het ekotone beskryf as sones van oorgang tussen aangrensende ekologiese sisteme, met 'n stel unieke eienskappe wat gedefinieer word deur ruimte en tydskale, en deur die sterkte van interaksies tussen aangrensende ekologiese sisteme (Holland 1988). Hierdie definisie baan die weg vir navorsing wat verskeie aspekte van landskapekologie en ruimtelike heterogeniteit kan illustreer. Alhoewel 'n nis van hoë wetenskaplike belang is, word ekotonale navorsing baie onderbestudeer, veral navorsing oor die gebruik van Afstandswaarneming om fynskaalse vleiland-ekotone te identifiseer en te karteer. 'n Bibliometriese analise en literatuuroorsig het getoon dat beperkte navorsing oor vleiland-ekotone in Suider-Afrika gedoen is, maar met voldoende literatuur gedek oor vleilandafbakening, klassifikasie en kartering. Vleilande wat hoogs dinamies is en beskou word as bewegende entiteite in 'n landskap as gevolg van hul wisselende hidroperiodes, is veral uitdagend om te karteer. Twee hoofeksperimente is uitgevoer wat albei Masjienleer (ML) algoritmes gebruik het, naamlik Random Forest (RF) en die naïewe Bayes klassifiseerder. Die doel van die eerste eksperiment was om afstandswaarnemingstegnieke te hersien en te toets om afsonderlike plantegroeigemeenskappe in die Du Toitsrivier-vleiland, Wes-Kaap Suid-Afrika akkuraat te identifiseer en te karteer. Die tweede eksperiment was dan om waarskynlikheidsklassifikasiemaatreëls te gebruik om die ekotone wat in 'n fynbos ingebedde vleiland-ekosisteem heers te karteer en te karakteriseer. Die studie het vrylik beskikbare satellietbeelde gebruik, naamlik Landsat 8 Surface Reflectance Tier 1, en Sentinel-2 MSI: MultiSpectral Instrument, Level-2A, verkry van die Verenigde State se Geologiese Opname deur middel van oopbronbronne soos Google Earth Engine (GEE). Hierdie navorsing dui daarop dat Random Forest (RF) klassifiseerder groot potensiaal getoon het in die akkurate kartering van landbedekking, spesifiek vier duidelike en dominante plantegroeitipes binne die vleiland, naamlik Prionium serratum, Psoralea pinnata (na verwys as palmiet-vleilandplantegroei), 'n gekondenseerde groep Pteridium aquilinum, Restio paniculatus en Merxmuellera cincta (na verwys as sklerofilagtige vleilandplantegroei), en Tydelike Vleilandfynbos. RF resultate het min spektrale verwarring tussen klasse getoon en matige tot hoë algehele akkuraatheid getoon vir klassifikasies wat deur beide die winter en somerseisoene loop. Die doeltreffendheid van die gebruik van die fuzzy logika d.w.s. toesighoudende waarskynlikheidsmaatreëls om ekotone in 'n ruimtelik heterogene landskap te identifiseer en te karteer, is ten toon gestel. Probabilistiese kartering en fuzzy grafieke het komplekse en diverse ekotone binne die vleiland getoon. Dit was duidelik dat duidelike ekotone in die vorm van vinnige en skerp hoë waarskynlikhede van een plantegroeitipe gekruis en 'n ander vervang het. Hierdie ekotone kan nuttige inligting verskaf oor vleiland-ekosisteemfunksionering en hoe plantegroeisones kan bydra tot vleiland-ekosisteemdienste (bv. vloeddemping en koolstofberging). Die gebruik van 'n per-pixel-gebaseerde benadering om ekotone te karteer is baie nuttig aangesien ekotone in werklikheid meer kompleks is en om dit as enkelvektorlyne te karteer is nie nuttig of akkuraat nie. Alhoewel hierdie studie daarop gemik was om fynskaalse vleiland-ekotone te identifiseer en te karteer, sal verdere navorsing deur gebruik te maak van selfs fyner skaaldata en meer in-diepte veldanalise wat spesifiek op hierdie geïdentifiseerde en gekarteerde ekotonale gebiede fokus, betekenisvol wees.
Description
Thesis (MA)--Stellenbosch University, 2021.
Keywords
Alluvial fans, Landsat satellites, Remote sensing, Wetland conservation, Ecotones, Machine learning, Land-water ecotones, Data mining, Image processing -- Digital techniques, Wetlands -- South Africa -- Western Cape -- Classification, Fynbos ecology, UCTD
Citation