Mapping surface water in complex and heterogeneous environments using remote sensing

Date
2019-04
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Global climate change characterised by rising temperatures and changes in the magnitude and intensity of precipitation is projected to affect the spatial and temporal distribution of land surface water (LSW) resources. Accurate and reliable information on the dynamics of LSW is valuable in understanding and monitoring the occurrence and impacts of floods and droughts. This knowledge is also critical for appropriate planning and impact assessment. Research has showed that droughts and floods are the two major hydrological disasters in developing countries such as southern Africa. This is mainly due to the lack of accurate and robust methods and reliable data sources necessary for monitoring the spatial and temporal dynamics of LSW resources. Satellite remote sensing (RS) technology is a promising primary data source and provides techniques suitable for repeated mapping water bodies and flood plains. However, many flood plains and water bodies are characterised by the presence of submerged vegetation, dissolved and suspended substances. These characteristics limit the application of RS in monitoring LSW resources. This study evaluated the potential of remotely sensed data with different temporal, spatial and radiometric properties to map LSW in such challenging environments. Three experiments were carried out. The first experiment evaluated a new spectral indices-based unmixing algorithm that uses a minimum number of spectral bands. The algorithm was applied to Medium Resolution Imaging Spectrometer Full Resolution (MERIS FR) imagery to map open water and partly submerged vegetation. MERIS FR imagery has high (three days) temporal, but low (300 m) spatial resolution. The quality of the flood map derived from MERIS data was compared to high (30 m) spatial, but low (16 day) temporal resolution Landsat Thematic Mapper (TM) images on two different flooding dates (17 April 2008 and 22 May 2009). The findings show that, despite the low resolution of MERIS, both the spatial and frequency distribution of the water fraction extracted from the MERIS data were in good agreement with the high-resolution TM retrievals. This suggests that the proposed technique can be used to produce reliable and frequent flood maps using low spatial resolution imagery. The use of synthetic aperture radar (SAR) has become increasingly relevant for mapping and monitoring flooded vegetation (FV). In a second experiment, a procedure was constructed and validated based on a time series of Sentinel-1 SAR data for mapping floods in a vegetated floodplain. For each newly available image, the probability of temporary flooded conditions is tested against the probability of not-flooded conditions. The changes in land cover characteristics are considered by the technique. The modelling and testing components were applied independently to the vertical transmit and horizontal receive (VH) polarisation, vertical transmit and vertical receive (VV) and VH/VV ratio. The resulting flood maps were compared to those obtained from Landsat-8 Operational Land Imager (OLI) and ground truthing. Overall classification accuracies showed that the maps produced from the fused Sentinel-1 products (VH and VH/VV) were most accurate (84.5%) and significantly better than when only the VH polarisation was used (78.7%). These results demonstrate that the fusion of VH/VV and VV polarisations can improve flood mapping in vegetated floodplains. The third experiment involved using automatic thresholding of near-concurrent normalized difference water index (NDWI) (generated from Sentinel-2) and VH backscatter bands (generated from Sentinel-1) to map waterbodies with diverse spectral and spatial characteristics. The resulting maps were compared to the classification performances of five machine learning algorithms (MLAs), namely decision tree (DT), k-nearest neighbour (k-NN), random forest (RF), and two implementations of the support vector machine (SVM). The results show that the combination of multispectral indices with SAR data is highly beneficial for classifying complex waterbodies and that the proposed thresholding approach classified waterbodies with an overall classification accuracy of 89.3%. However, the varying concentrations of suspended sediments (turbidity), dissolved particles and aquatic plants negatively affected the classification accuracies of the proposed method, whereas the MLAs (SVM in particular) were less sensitive to such variations. The LSW maps and techniques developed in this study are critical for flood status monitoring, water resources planning and disaster management, and will as such reduce the impact of floods and droughts on vulnerable communities living in southern Africa. Furthermore, the results of this study will hopefully inspire the remote sensing community to make use of the new generation of freely available multispectral and SAR data (such as those provided by the Sentinel constellations) for operational drought and flood monitoring.
AFRIKAANSE OPSOMMING: Globale klimaatsverandering gekenmerk deur stygende temperature en veranderinge in die grootte en intensiteit van presipitasie word geprojekteer om die ruimtelike en temporale verspreiding van hulpbronne vir grondoppervlakwater (GOW) te beïnvloed. Akkurate en betroubare inligting oor die dinamika van GOW is nuttig om die voorkoms en impak van vloede en droogtes te verstaan en te monitor. Hierdie kennis is ook van kritieke belang vir toepaslike beplanning en impakbepaling. Navorsing het getoon dat droogtes en vloede die twee grootste hidrologiese rampe in ontwikkelende lande, soos Suider-Afrika, is. Dit is hoofsaaklik te wyte aan die gebrek aan akkurate en robuuste metodes, tesame met ‘n tekort aan betroubare databronne wat vir die monitering van die ruimtelike en temporale dinamika van GOW-hulpbronne benodig word. Satelliet afstandswaarneming (AW)-tegnologie is 'n belowende primêre databron en bied tegnieke wat vir herhaalde kartering van waterliggame en vloedvlaktes geskik is. Baie vloedvlaktes en waterliggame word egter deur die teenwoordigheid van ondergedompelde plantegroei en opgeloste en gesuspendeerde stowwe gekenmerk. Hierdie eienskappe beperk die toepassing van AW in die monitering van GOW-hulpbronne. Hierdie studie het die potensiaal van afstandswaarnemingdata met verskillende tydelike, ruimtelike en radiometriese eienskappe geevalueer om GOW in sodanige uitdagende omgewings te karteer. Drie eksperimente is uitgevoer. Die eerste eksperiment het 'n nuwe spektrum indeks-gebaseerde ontmenging-algoritme geëvalueer wat gebruik maak van 'n minimum aantal spektrale bande. Die algoritme is toegepas op Medium-Resolusie Beeldvormende Spektrometer Volle Resolusie (MERBS VR) beeldmateriaal om oop water en plante wat gedeeltelik gedompel is te karteer. MERBS VR beeldmateriaal het 'n hoë (drie dae) temporale resolusie, maar 'n lae (300 m) ruimtelike resolusie. Die kwaliteit van die vloedkaart wat afgelei is van die MERBS-data is teen hoë (30 m) ruimtelike resolusie, maar lae (16 dae) temporale Landsat Tematiese Karteerder (TK) beelde van twee verskillende datums (17 April 2008 en 22 Mei 2009) waartydens oorstromings plaasgevind het, geëvalueer. Die bevindings toon dat, ten spyte van die lae resolusie van MERBS, beide die ruimtelike en frekwensieverspreiding van die waterfraksie wat vanuit die MERBS-data verkry is goed ooreengestem het met die hoë-resolusie TK-herwinnings. Dit dui daarop dat die voorgestelde tegniek gebruik kan word om betroubare en gereelde vloedkaarte te produseer deur van lae-ruimtelike-resolusie-beelde gebruik te maak. Die gebruik van sintetiese diafragma-radar (SDR) het toenemend relevant vir die kartering en monitering van oorstroomde plantegroei (OP) geword. In 'n tweede eksperiment is ’n prosedure, gebaseer op 'n tydreeks van Sentinel-1 SDR-data, vir die kartering van oorstromings in 'n vloedvlakte met plante ontwikkel en gevalideer. Vir elke nuwe beskikbare beeld word die waarskynlikheid van tydelik-oorstroomde toestande getoets teen die waarskynlikheid van nie-oorstroomde toestande. Veranderinge in grondbedekkingseienskappe word deur die tegniek oorweeg. Die modellering- en toetskomponente is onafhanklik op die vertikale transmissie en horisontale ontvangs (VH), vertikale transmissie en vertikale ontvangs (VV) en VH/VV verhouding polarisasies toegepas. Die resulterende vloedkaarte is met dié van Landsat-8 Operasionele-grondbeelder (OGB) en grondslag-getrouheid vergelyk. Algehele klassifikasie-akkuraatheid het getoon dat die kaarte wat uit die aaneengesmelte Sentinel-1 produkte (VH en VH/VV) vervaardig is, die akkuraatste (84,5%) was en aansienlik beter was as wanneer slegs die VH polarisasie gebruik is (78,7%). Hierdie resultate toon dat die samesmelting van VH/VV en VV-polarisasies die vloedkartering in beplante vloedvlaktes kan verbeter. Die derde eksperiment het die gebruik van outomatiese drempelbepaling van naby-gelyktydig genormaliseerde verskil-natheid-indeks (GVNI) (gegenereer met Sentinel-2 beelde) en VH-terugverspreidingbande (gegenereer met Sentinel-1 data) behels om waterliggame met uiteenlopende spektrale en ruimtelike eienskappe te karteer. Die resulterende kaarte is vergelyk met die klassifikasieprestasies van vyf masjienleer-algoritmes (MLAs), naamlik besluitboom (BB), k-naaste buurman (k-NN), ewekansige woud (EW) en twee implementasies van die ondersteuningsvektormasjien (OVM). Die resultate toon dat die kombinasie van multispektrale indekse met SDR data uiters voordelig vir die klassifikasie van komplekse waterliggame is en dat die voorgestelde drempelbepalingbenadering waterliggame met 'n algehele klassifikasie-akkuraatheid van 89,3% geklassifiseer het. Die wisselende konsentrasies van gesuspendeerde sedimente (turbiditeit), opgeloste deeltjies en waterplante het egter die klassifikasie-akkuraatheid van die voorgestelde metode negatief beïnvloed, terwyl die MLAs (OVM in die besonder) minder sensitief vir sodanige variasies was. Die GOW-kaarte en -tegnieke wat in hierdie studie ontwikkel is, is van kritieke belang vir vloedstatusmonitering, waterhulpbronbeplanning en rampbestuur en sal sodanig die impak van vloede en droogtes op kwesbare gemeenskappe in Suider-Afrika verminder. Daarbenewens sal die resultate van hierdie studie hopelik die afstandswaarneminggemeenskap inspireer om van die nuwe generasie, vrylik-beskikbare multispektrale en SDR-data gebruik te maak om operasionele droogte en vloede te monitor (soos die wat deur die Sentinel-konstellasies verskaf word).
Description
Thesis (PhD)--Stellenbosch University, 2019.
Keywords
Remote sensing, Water supply, Agricultural -- Remote sensing, Water in agriculture -- Remote sensing, Irrigation efficiency -- Remote sensing, Land use mapping -- Remote sensing, Land use, Urban -- Remote sensing, Multi-spectral imaging -- Remote sensing, Spectrum analysis -- Remote sensing, Spectro chemistry -- Remote sensing, Machine learning -- Remote sensing, Heterogenous catalysis -- Remote sensing, Biodiversity -- Remote sensing, Ecological heterogeneity -- Remote sensing, Land use mapping -- Remote sensing, UCTD
Citation