Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
An earth observation approach for mapping indigenous and plantation forests at regional scale
Thesis (MA)--Stellenbosch University, 2024.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2024
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613761042382848 |
|---|---|
| access_status_str | Open Access |
| author | Ndyafi, Vincent Chimbunde |
| author2 | Van Niekerk, Adriaan |
| author_browse | Van Niekerk, Adriaan Ndyafi, Vincent Chimbunde |
| author_facet | Van Niekerk, Adriaan Ndyafi, Vincent Chimbunde |
| author_sort | Ndyafi, Vincent Chimbunde |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MA)--Stellenbosch University, 2024. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/130286 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:41:16.700Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2024 |
| publishDateRange | 2024 |
| publishDateSort | 2024 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/130286 An earth observation approach for mapping indigenous and plantation forests at regional scale Ndyafi, Vincent Chimbunde Van Niekerk, Adriaan Stephenson, Garth Stellenbosch University. Faculty of Arts and Social Sciences. Dept. of Geography and Environmental Studies. Neural networks (Computer science) Deep learning (Machine learning) Forests and forestry Land cover Multispectral imaging Object-oriented programming (Computer science) Software compatibility High resolution imaging Forests and forestry -- Remote sensing UCTD Thesis (MA)--Stellenbosch University, 2024. ENGLISH ABSTRACT: The thesis research focuses on using very-high-resolution (VHR) red, green and blue (RGB) (low spectral resolution) imagery for quantitative image analysis. It explores the effectiveness of convolutional neural network (CNN) models for extracting forest and land-cover information from VHR RGB aerial imagery sourced from public institutions like the Chief Directorate: National Geospatial Information, Google Earth and Microsoft. The study aims to demonstrate cost-effective strategies for accurate mapping of forests and to evaluate the impact of various CNN parameters on classification accuracies. Existing studies using high-resolution imagery coupled with machine learning (ML) algorithms for mapping land cover are reviewed to reveal a dearth of research using RGB VHR imagery specifically for forest mapping in South Africa. The few studies that investigated the value of RGB VHR imagery for quantitative analysis have achieved high accuracies (~90%), indicating the potential of such data for mapping forests. The research objectives were to develop an objectbased CNN model for land-cover mapping; evaluate different CNN parameters’ impact on classification accuracy; and to compare object-based and per-pixel CNN approaches. Results reveal the significance of certain parameters on classification accuracy and demonstrate the superiority of object-based CNNs over per-pixel methods. The study further investigates signature extension across different climatic zones and biomes in South Africa. The diversity of sample sets was shown to significantly impact the model’s accuracy and suggest a combining of samples from diverse regions to enhance accuracy. The research underscores the potential of remote-sensing technologies and deep learning for cost-effective forest mapping and validates the operational use of VHR RGB imagery at regional and national scales. Despite the achieved objectives, there are evident needs for thorough tuning of parameters, additional methods for sample augmentation and the consideration of spectral bands that are sensitive to vegetation colour. The leveraging of publicly available online applications to reduce acquisition costs of imagery is recommended and the potential for future improvements is highlighted, including the fusion of higher spectral resolution images and the exploration of post-classification rule sets using vegetation indices. Ultimately, the research provides guidelines for operational forest mapping to enable regular updating of forest inventories at minimal cost. AFRIKAANSE OPSOMMING: Die tesisnavorsing fokus op die gebruik van baie-hoë-resolusie (BHR) rooi, groen en blou (RGB) (laag spektrale resolusie) beelde vir kwantitatiewe beeldanalise. Dit ondersoek die doeltreffendheid van konvolusionele neurale netwerkmodelle (KNN) vir die onttrekking van woud- en grondbedekkingsinligting uit BHR RGB-lugfoto’s wat van openbare instellings soos die Hoofdirektoraat: Nasionale Geo-ruimtelike Inligting, Google Earth en Microsoft verkrygbaar is. Die studie doel is om koste-effektiewe strategieë vir akkurate kartering van woude en die evaluering van die impak van verskeie parameters van KNN-modelle op klassifikasie akkuraatheid te demonstreer. Bestaande studies wat hoë-resolusie-beelde tesame met masjienleer-algoritmes vir die kartering van grondbedekking word bestudeer, maar 'n tekort aan navorsing oor die gebruik van RGB BHR-beelde spesifiek vir woudkartering in Suid-Afrika is onthul. Die paar studies wat die waarde van RGB BHR-beelde vir kwantitatiewe analise ondersoek het, het hoë akkuraatheid behaal (~90%), wat die potensiaal van sodanige data vir die kartering van woude aandui. Die navorsingsdoelwitte was om 'n objekgebaseerde KNN-model vir die kartering van landbedekking te ontwikkel; verskillende KNN-parameters se impak op klassifikasieakkuraatheid te evalueer; en om objekgebaseerde en per-beeldelement KNN-benaderings te vergelyk. Resultate dui op die belangrikheid van sekere parameters op klassifikasie-akkuraatheid en demonstreer die superioriteit van objekgebaseerde KNN bo per-beeldelement metodes. Die studie ondersoek verder kenmerkende uitbreiding oor verskillende klimaatsones en biome in Suid-Afrika. Dit het geblyk dat die diversiteit van steekproefstelle die akkuraatheid van die model beduidend beïnvloed, en 'n kombinasie van steekproewe uit verskillende streke word voorgestel om die akkuraatheid te verbeter. Die navorsing beklemtoon die potensiaal van afstandswaarnemingstegnologieë en diep leer vir koste-effektiewe woudkartering en valideer die operasionele gebruik van BHR RGB-beelde op streeks- en nasionale skaal. Ten spyte van die doelwitte wat bereik is, is daar duidelike behoefte aan deeglike aanpassing van parameters, bykomende metodes vir steekproeftoevoeging en 'n oorweging van spektrale bande wat sensitiefis vir die kleur van plantegroei. Die gebruik van openlik-beskikbare aanlyn toepassings om die verkrygingskoste van beelde te verminder word aanbeveel, en die potensiaal vir toekomstige verbeterings word uitgelig, insluitend die samesmelting van hoër ruimtelike resolusie beelde en die verkenning van naklassifikasie- reëlstelle wat van plantegroei-indekse gebruik maak. Hoofsaaklik bied die navorsing riglyne vir operasionele woudkartering om gereelde opdatering van woudvoorraad teen minimale koste moontlik te maak. Masters 2024-03-01T05:33:03Z 2024-04-26T12:04:54Z 2024-03-01T05:33:03Z 2024-04-26T12:04:54Z 2024-03 Thesis https://scholar.sun.ac.za/handle/10019.1/130286 en Stellenbosch University xiv, 126 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Neural networks (Computer science) Deep learning (Machine learning) Forests and forestry Land cover Multispectral imaging Object-oriented programming (Computer science) Software compatibility High resolution imaging Forests and forestry -- Remote sensing UCTD Ndyafi, Vincent Chimbunde An earth observation approach for mapping indigenous and plantation forests at regional scale |
| title | An earth observation approach for mapping indigenous and plantation forests at regional scale |
| title_full | An earth observation approach for mapping indigenous and plantation forests at regional scale |
| title_fullStr | An earth observation approach for mapping indigenous and plantation forests at regional scale |
| title_full_unstemmed | An earth observation approach for mapping indigenous and plantation forests at regional scale |
| title_short | An earth observation approach for mapping indigenous and plantation forests at regional scale |
| title_sort | earth observation approach for mapping indigenous and plantation forests at regional scale |
| topic | Neural networks (Computer science) Deep learning (Machine learning) Forests and forestry Land cover Multispectral imaging Object-oriented programming (Computer science) Software compatibility High resolution imaging Forests and forestry -- Remote sensing UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/130286 |
| work_keys_str_mv | AT ndyafivincentchimbunde anearthobservationapproachformappingindigenousandplantationforestsatregionalscale AT ndyafivincentchimbunde earthobservationapproachformappingindigenousandplantationforestsatregionalscale |