Object-based geomorphometry and soil spectroscopy in a fluvial terrace dominated landscape

Date
2021-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: The adaptation and mitigation of adverse global change necessitate better understanding and management of natural resources. Fluvial terrace staircases are a paleo-fluvial phenomenon critical to mining, agriculture, ecosystem services and the deciphering of paleo-tectonic and -climatic conditions.This study aims to develop methodologies for generating fine-scale object-based fluvial terrace maps. A brief treatise contextualises relevant concepts, techniques and methodologies related to object-based geomorphometry and soil spectroscopy.The paleo-fluvial landscape of theGreat Letaba River Catchment was investigated. An overview of the formation of the Great Escarpment and related tectonic uplift events provides Lowveld paleo-fluvial context. More over,the limited available evidence detailing the existence of fluvial terraces and planation surfaces along the Great Letaba River is reviewed.The combination of field reconnaissance and terrain analysis results with 1:250 000 scale soil data enabled the characterisation of extensive fluvial terrace and planation surface landforms and the collection of critical geomorphometric reference data. The findings expand existing morphological knowledge of the Great Letaba River Catchment fluvial terraces and planation surfaces and identify a distinct landscape, soil, climate and lithology paradigm that drives a complex cycle of landscape evolutionary processes. Specific geomorphometric land surface segmentation approaches delineate predefined landform units using largely scale-independent techniques that do not incorporate hierarchical landscape representations. This study details a new approach that delineates predefined landforms at multiple scale levels using discrete geomorphometric principles and local variance statistical techniques. First, land surface segmentation scale optimisation was exposed as an ill-structured problem, and a methodology for defining well-structured landform conditions was outlined. Next, an ensemble of scale optimisation techniques was constructed and applied to evaluate each well-structured condition at incrementally increased scale increments. Agreement between the scale optimisation techniques indicates scale levels at which the predefined landform conditions are met. The results indicate that ensemble scale optimisation techniques, unlike existing single technique approaches, produce refined scale selections that minimises analyst involvement in the scale optimisation process.The bulk of geomorphometric fluvial terrace mapping strategies employ per-pixel strategies that are either based on deterministic approaches that require extensive user parameterisation or black-box supervised classifiers that exclude expert knowledge from the classification process. A new methodology that couples object-based landform units with white-box decision tree (DT) classifiers was envisioned as an intermediary between knowledge-based and black-box classifiers. To evaluate the new methodology, an experiment was designed where both binary classes (terrace and non-terrace) and multiple fluvial terraces (non-terrace and ten fluvial terrace levels) were classified using per-pixel and object-based approaches. Through inductive DT classifier analysis, the object-based rule sets were shown to produce the most intuitively interpretable fluvial terrace rule sets. Moreover, qualitative and quantitative accuracy assessments reaffirm the superiority of object-based fluvial terrace classifications. The differentiation and correlation of fluvial terrace remnants in eroded landscapes often require resource intensive insitu stratigraphic and physiographic interpretations. Using discrete geomorphometric techniques and soil spectroscopy, the utility of the soil-landscape paradigm for fluvial terrace level mapping was illustrated. First, a chemometric principle component analysis (PCA) and partial least-squares regression (PLSR) approach was used to establish a good linear relationship between 88 spectral signatures and the relative height of fluvial terraces. Next, random forest (RF) was used to classify two sets of fluvial terrace level classes and to isolate the most important spectral wavelengths. Subsequent RF models indicated that individual wavelengths associated with smectite and kaolinite fractions provide sufficient variance to accurately classify fluvial terrace levels. The methodology introduced outlines a low-cost, semi-automated spectroscopic alternative to insitu fluvial terrace mapping approaches. Fluvial terrace maps are crucial forfacilitating–among other things –the production of fine-scale resource inventories, land use planning and global change adaptation and mitigation strategies.Yet, most southern African fluvial terraces remain largely unexplored.This research contributes new knowledge to the fields of southern African paleo-fluvial geomorphology, geomorphometry and fluvial terrace mapping through landscape characterisations of the Great Letaba River Catchment and the development of novel geomorphometric and soil spectroscopic methodologies.These contributions provide a sound foundation for further Lowveld paleo-fluvial investigations, the regional mapping of fluvial terraces and the development of transferable geomorphometric landform mapping methodologies.
AFRIKAANSE OPSOMMING: Die aanpassing en versagting van nadelige globale verandering noodsaak beter kennis en bestuur van natuurlike hulpbronne. Rivier terrastrappe is ʼn wêreldwye paleo-fluviale verskynsel wat van kritieke belang vir mynbou, landbou, ekostelseldienste en die ontsyfering van paleo-tektoniese en klimaatstoestandeis. Die doel van die studie is om nuwe metodologieë vir die skep van fynskaalse voorwerp gebaseerde kaarte van fluviale terrasse te ontwikkel. ʼn Kort verhandeling kontekstualiseer relevante konsepte, tegnieke en metodologieë wat met voorwerp gebaseerde geomorfometrie en grondspektroskopie verband hou. Die paleo-fluviale landskap van die Groot Letabarivier opvanggebied is ondersoek. ʼn Oorsig van die Groot Platorand formasie en verwante tektoniese opheffings gebeurtenisse verskaf konteks oor die Laeveldse paleo-fluviale prosesse. Daarbenewens word die beperkte beskikbare bewyse wat die bestaan van fluviale terrasse en planasie oppervlaktes langs die Groot Letaba rivier hersien. Die kombinasie van veldverkennings-en terrein analise-resultate met 1: 250 000 skaalgronddata het die karakterisering van wydverspreide fluviale-terras-en planasie-oppervlak landvorms en die inwinning van kritieke geomorfometriese verwysingsdata moontlik gemaak. Die bevindings brei die bestaande morfologiese kennis van die Groot Letaba rivier opvangsgebied fluviale terrasse en planasie oppervlaktes aansienlik uit en identifiseer ʼn duidelike paradigma vir landskap, grond, klimaat en litologie wat ʼn komplekse kringloop van evolusie prosesse in die landskap dryf. Spesifieke geomorfometriese landoppervlak-segmenterings benaderings definieer vooraf-gedefinieerde landvorm-eenhede met grootliks skaal-onafhanklike tegnieke wat nie hiërargiese landskapvoorstellings inkorporeer nie. Hierdie studie sit ʼn nuwe benadering uiteen wat vooraf-gedefinieerde landvorms op meervoudige skaalvlakke met behulp van diskrete geomorfometriese beginsels en statistiese tegnieke vir plaaslike variansie afbaken. Eerstens is die skaaloptimering van landoppervlak segmentering as ʼn ongestruktureerde problem blootgelê, en ʼn metodologie vir die definisie van goed gestruktureerde landvorm toestande is geïllustreer. Vervolgens is ʼn ensemble van skaaloptimaliseringstegnieke saamgestel en toegepas om elke goed gestruktureerde toestand met toenemende skaalstygings te evalueer. Ooreenkoms tussen die skaaloptimaliseringstegnieke dui skaalvlakke aan waar daar aan die vooraf-gedefinieerde landvorm voorwaardes voldoen word. Die resultate dui aan dat ensemble-skaaloptimaliseringstegnieke, in teenstelling met bestaande enkele tegniek benaderings, verfynde skaalkeuses oplewer wat ontleder betrokkenheid by die skaaloptimaliseringsproses beperk. Die meeste geomorfometriese fluviale-terraskaart strategieë gebruik per-piksel strategieë wat óf gebaseer is op deterministiese benaderings wat omvattende gebruikers parameterisering vereis, óf swart-boks gekontroleerde-klassifikasies wat kundige kennis van die klassifikasieproses uitsluit. ʼn Nuwe metodologie wat voorwerp gebaseerde landvorm-eenhede met witboks-beslissings boomklasseerders koppel, is as ʼn tussenganger tussen kennis gebaseerde-en swartboks-klassifiseerders beskou. ʼn Eksperiment,wat beide binêre klasse (terras en nie-terras) en meervoudige fluviale terrasse (nie-terras en tien fluviale-terrasvlakke) met behulp van per-piksel en voorwerpgebaseerde benaderings klassifiseer, is ontwerp om die nuwe metodologie te evalueer. Deur middel van induktiewe beslissingsboomklassifiseerder-analise is getoon dat die voorwerpgebaseerde reëlstelle die mees instinktiewe, interpreteerbare reëlstelle vir fluviale terrasse lewer. Daarbenewens bevestig kwalitatiewe en kwantitatiewe akkuraatheidsbeoordelings die superioriteit van voorwerpgebaseerde fluviale-terras klassifikasies.
Description
Thesis (DPhil)--Stellenbosch University, 2021.
Keywords
GIS (Information systems), Machine Learning, Fluvial geomorphology, Geoinformatics, Geology, Structural imaging -- Research, Soil structue -- Spectroscopic imaging, Remote sensing, Geographic Information Systems, Terraces (Geology), Soil science -- Spectroscopic imaging, Letaba River Catchment (South Africa), Lowveld Area (South Africa), UCTD
Citation