Regional mapping of spekboom canopy cover using very high resolution aerial imagery

Date
2019-12
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Widespread degradation of subtropical thicket (South Africa) by poorly managed pastoralism has led to substantial decreases in ecological functioning and biodiversity. Once degraded, thicket does not recover after the removal of livestock, but requires active intervention for restoration. It has been shown that planting spekboom (Portulacaria afra), a dominant and keystone thicket species, increases biomass, improves soil health and creates conditions that support the natural regeneration of biodiversity. Spatial data, especially spekboom canopy cover maps, are required to inform and support large-scale restoration. This research aimed to develop and demonstrate a semi-automated spekboom canopy cover mapping method. A large study area in the Little Karoo in South Africa was selected to encompass the ecological heterogeneity of the wider region. Following a literature review, very high resolution (VHR) multi-spectral aerial imagery was identified as a viable data source for the fine-scale discrimination of spekboom. A set of 2228 aerial images covering the study area was subsequently acquired from Chief Directorate: National Geo-spatial Information (NGI). Techniques for (1) radiometric correction and (2) feature selection were devised to address specific challenges of regional canopy cover mapping. These techniques then formed components of the spekboom canopy cover mapping method. The need for the first technique, called radiometric homogenisation, arose from the presence of problematic radiometric variation in the aerial imagery. Radiometric homogenisation corrects for varying atmospheric and bidirectional reflectance distribution function (BRDF) effects by calibration with concurrent and collocated satellite surface reflectance data. In contrast to other radiometric correction methods, manual placement or acquisition of reflectance targets is not required. Moreover, it is not necessary to have detailed knowledge of atmospheric conditions at the time of capture. An experiment was conducted to establish the efficacy and accuracy of the technique. Homogenised images of the study area were validated by visual inspection and statistical comparison to surface reflectance reference data. Recognisable anomalies such as hot spots and seam lines were removed, and statistical results compared well to competing methods. While the technique was developed in the context of the spekboom canopy cover mapping problem, it could also be applied to general radiometric correction of VHR imagery. Radiometric homogenisation is especially applicable to large study areas where radiometric uncertainty can prevent accurate classification. The second technique, called feature clustering and ranking (FCR), was devised to address problems of sub-optimality and instability that often arise when applying feature selection to redundant data. Unlike other feature selection approaches, FCR allows for the optional inclusion of factors other than relevance (such as computation and measurement cost) in the selection criteria. An experiment was conducted to compare the effects of redundancy on popular feature selection approaches and FCR. Results confirmed that redundancy has a negative impact on commonly used ranking and greedy search (stepwise) feature selection methods. FCR provided the best accuracy and stability performance, confirming its value for selecting stable, informative features from high dimensional data containing redundancy. Finally, the radiometric homogenisation and FCR techniques were incorporated into a method for VHR spekboom canopy cover mapping. Per-pixel spectral, textural and vegetation index features were generated from imagery that had been processed with the radiometric homogenisation technique. FCR was subsequently used to select a reduced set of informative and computationally efficient features. The core of the spekboom mapping method consisted of supervised classification of selected features, followed by morphological post-processing of classifier output maps to remove noise and smooth boundaries. An experiment was carried out to test the accuracy of popular classifiers by comparing canopy cover estimates to ground truth data. A decision tree provided the best performance of the tested classifiers. Canopy cover maps exhibited some variation between different habitats, but provided good accuracy overall, with a mean absolute (canopy cover) error (MAE) of 5.85%. Regional vegetation maps are urgently required to inform responses to global issues, such as climate change. While there is a known operational need for large-area VHR vegetation maps, there are surprisingly few studies that address the cost, computation and classifier transferability challenges associated with large spatial extents. This research contributes to the important field of regional vegetation mapping through the development of the radiometric homogenisation and FCR techniques. In the context of thicket restoration, a viable method for regional mapping of spekboom canopy cover was demonstrated, providing a valuable foundation for future expansion of maps to the rest of the thicket biome. The techniques developed in this study will be useful for the mapping of other thicket vegetation traits, such as biomass.
AFRIKAANSE OPSOMMING: Wydverspreide agteruitgang van subtropiese struikgewas (Suid-Afrika) as gevolg van swak bestuurde weiding het gelei tot aansienlike afname in ekologiese funksionering en biodiversiteit. Wanneer dit eers gedegradeer het, herstel struikgewasse nie sonder aktiewe ingryping ná die verwydering van lewende hawe nie. Daar is getoon dat die plant van spekboom – 'n dominante en sleutelstruikgewasspesie – biomassa verhoog, die grondgesondheid verbeter en toestande skep wat die natuurlike regenerasie van biodiversiteit ondersteun. Ruimtelike data, veral spekboomblaardakdekkingskaarte, word vir inligting oor en ondersteuning van grootskaalse restourasie benodig. Hierdie navorsing se doel was om 'n semi-geoutomatiseerde karteringsmetode vir spekboomblaardakdekking te ontwikkel. 'n Groot studiegebied in die Klein Karoo in Suid-Afrika is gekies om die eienskappe van die ekologiese heterogeniteit van die breër streek in te sluit. Na 'n literatuuroorsig is baie hoë resolusie (BHR) multispektrale lugfoto's as 'n lewensvatbare databron vir die fynskaalse diskriminasie van spekboom geïdentifiseer. 'n Stel van 2228 lugfoto's wat die studiegebied dek, is daarna van die Hoofdirektoraat: Nasionale Georuimtelike Inligting (NGI) verkry. Tegnieke vir (1) radiometriese korreksie en (2) eienskapseleksie is bedink om spesifieke uitdagings van die kartering van streeksblaardakdekking aan te spreek. Hierdie tegnieke het dan komponente van die karteringsmetode van spekboomblaardakdekking gevorm. Die behoefte aan die eerste tegniek, genaamd radiometriese homogenisering, het as gevolg van die teenwoordigheid van problematiese radiometriese variasie in die lugfoto’s ontstaan. Radiometriese homogenisering korrigeer vir verskillende atmosferiese en effekte van tweerigtingweerkaatsingverspreidingsfunksie (TWVF) deur kalibrasie met samelopende en gegroepeerde data oor satelliet-oppervlakweerkaatsing. In teenstelling met ander radiometriese regstellingsmetodes, is dit nie nodig om teikens vir weerkaatsing per hand te plaas of te verkry nie. Daarbenewens word gedetailleerde kennis oor atmosferiese toestande ten tye van die vaslegging nie benodig nie. 'n Eksperiment is uitgevoer om die doeltreffendheid en akkuraatheid van die tegniek te bepaal. Gehomogeneerde beelde van die studiegebied is deur visuele inspeksie en statistiese vergelyking met verwysingsdata van oppervlakweerkaatsing gevalideer. Herkenbare onreëlmatighede soos verligte punte en naatlyne is verwyder, en statistiese resultate het goed met mededingende metodes vergelyk. Al is die tegniek in die konteks van die probleem van die kartering van spekboomblaardakdekking ontwikkel, kan dit ook op algemene radiometriese regstelling van BHR-beelde toegepas word. Radiometriese homogenisasie is veral van toepassing op groot studiegebiede waar radiometriese onsekerheid akkurate klassifikasie kan voorkom. Die tweede tegniek, genoem kenmerksaamgroepering en -rangordening (KSR), is ontwerp om probleme van suboptimaliteit en onstabiliteit aan te spreek wat dikwels voorkom wanneer kenmerkseleksie op oortollige data toegepas word. In teenstelling met ander eienskapseleksiebenaderings, maak KSR voorsiening vir die opsionele insluiting van ander faktore buiten relevansie (soos berekenings- en metingskoste) in die siftingskriteria. 'n Eksperiment is uitgevoer om die effekte van oortolligheid op gewilde eienskapseleksiebenaderings en KSR te vergelyk. Resultate het bevestig dat oortolligheid 'n negatiewe impak op die algemeen gebruikte rangskikking en gulsige soektog (stapsgewys) eienskapseleksiemetodes het. KSR het die beste akkuraatheids- en stabiliteitsprestasie gelewer, wat sy waarde vir die kies van stabiele, insiggewende eienskappe van hoë-dimensionele data met oortolligheid bevestig. Ten slotte is die radiometriese homogenisasie- en KSR-tegnieke in 'n metode vir die kartering van BHR-spekboomblaardekking opgeneem. Per-piksel spektrale, tekstuur en plantegroei-indekskenmerke is van beelde wat met die radiometriese homogenisasie tegniek verwerk is, geskep. KSR is gevolglik gebruik om 'n verkorte stel informatiewe en berekeningsdoeltreffende kenmerke te kies. Die kern van die spekboomkarteringsmetode het uit gekontroleerde-klassifikasie van geselekteerde kenmerke bestaan, gevolg deur morfologiese naverwerking van klassifiseerderuitsetkaarte om geraas en gladde grense te verwyder. 'n Eksperiment is uitgevoer om die akkuraatheid van gewilde klassifiseerders te toets deur blaardakdekkingskattings met veldwaarnemings te vergelyk. 'n Beslissingsboom het die beste prestasie van die getoetsde klassifiseerders gelewer. Blaardakdekkingskaarte het 'n mate van variasie tussen verskillende habitatte getoon, maar het goeie algehele akkuraatheid behaal met 'n gemiddelde absolute (blaardakdekking) fout (GAE) van 5,85%. Streeksplantegroeikaarte word dringend benodig om reaksies op globale kwessies soos klimaatsverandering in te lig. Alhoewel die operasionele behoefte aan grootskaalse BHR-plantegroeikaarte bekend is, is daar verrassend min studies wat die uitdagings met betrekking tot koste, berekeningstyd en klassifiseerderoordraagbaarheid wat met groot ruimtelike areas verband hou, aanspreek. Deur die ontwikkeling van die radiometriese homogenisasie- en KSR-tegnieke dra hierdie navorsing by tot die belangrike veld van die kartering van streeksplantegroei by. In die konteks van die restourasie van struikgewasse is 'n lewensvatbare metode vir spekboomblaardekking op streeksvlak gedemonstreer. Dit bied 'n waardevolle grondslag vir toekomstige uitbreiding van kaarte na die res van die struikgewasbioom. Die tegnieke wat in hierdie studie ontwikkel is, sal ook vir die kartering van ander struikgewasplantegroei-eienskappe, soos biomassa, van nut wees.
Description
Thesis (PhD)--Stellenbosch University, 2019.
Keywords
Biodiversity conservation, Vegetation mapping, Portulacaceae -- South Africa -- Karoo -- Remote sensing, Spekboom -- South Africa -- Karoo -- Remote sensing, Plant canopies -- Remote sensing, UCTD
Citation