Machine learning for regional modelling of soil depth in the Western Cape of South Africa

Date
2024-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Soil depth is a property of soils that has a profound impact on numerous processes, natural and human-related. Soil depth itself is also impacted by various external factors. Currently, in South Africa, soil databases and knowledge suites contain many year’s worth of soil data collected over the course of various studies, however, these information repositories lack the ability to be updated frequently, and cost-effectively. Traditionally, these measurements are taken in-field and require lengthy and expensive excavations to be able to extract the necessary information. The data also mostly consists of low-resolution data that is not always appropriate for every application, especially smaller-scale, local operations. Digital soil mapping (DSM), along with freely available, high-resolution imagery and data offers a cost-efficient alternative to traditional soil mapping methods and has the potential to be automated, resulting in continuously updated datasets. These continuously updated datasets are needed increasingly as climate change affects the environment in which humans operate, specifically agriculture (though its reach extends to many other industries). They will contribute to decision-making assistance systems that are focussed on mitigating the effects of climate change. DSM, however, has not been widely adopted in South Africa, therefore there is a growing need for research relating to these issues. This study aims to evaluate digital soil mapping (DSM) and machine learning (ML) methodologies to map soil depth at a much higher resolution than what is currently available in South Africa, as well as address a few additional concerns when applying these methodologies, such as input data, sampling techniques etc. The two main objectives were divided into two experiment groups. The first experiment group evaluated the modelling ability of a per-pixel approach to soil depth modelling. Covariates relating to terrain variables, Sentinel-2 optical imagery and climatic data were used in various combinations to determine their influence on soil depth. It was found that these models were able to model soil depth with moderate performance, leaving room for improvement. The per-pixel approach also appeared to prefer higher resolution data (10 m) as opposed to lower resolution (30 m). The second experiment group followed a similar methodology, however, instead opting for an object-oriented approach. These models were found to be marginally more reliable and robust than the per-pixel-based analyses and appeared to have had less of a reliance on higher resolution data to produce models with comparable performance. The results show that soil depth classification models can be produced with overall accuracies of 66%. They also show that soil depth regression models can be produced with relatively low mean absolute error values of 288.81 mm and are able to explain 38% of the variability of soil depth distribution. Recommendations have also been made for future research to aid in the improvement of these performance metrics.
AFRIKAANSE OPSOMMING: Gronddiepte is 'n eienskap van grond wat 'n diepgaande impak het op talle prosesse, beide natuurlik en menslik. Gronddiepte self word ook beïnvloed deur verskeie eksterne faktore. Tans bevat gronddatabasisse en kennisversamelings in Suid-Afrika baie jare se gronddata wat oor die loop van verskeie studies ingesamel is. Hierdie inligtings stoorplekke het egter nie die vermoë om gereeld en koste-doeltreffend opgedateer te word nie. Tradisioneel word hierdie metings in die veld geneem en vereis dit langdurige en duur uitgrawings om die nodige inligting te verkry. Die data bestaan meestal uit lae resolusie data wat nie altyd geskik is vir elke toepassing nie, veral vir kleiner skaal, plaaslike operasies. Digitale grondkartering, saam met gratis beskikbare, hoë resolusie beeldmateriaal en data, bied 'n koste-doeltreffende alternatief vir tradisionele grondkarteringsmetodes en het die potensiaal om geoutomatiseer te word, wat lei tot deurlopend opgedateerde datastelle. Hierdie deurlopend opgedateerde datastelle is toenemend nodig omdat klimaatsverandering die omgewing waarin menslike bedrywighede gebeur, veral landbou (alhoewel dit strek na vele ander industrieë), beïnvloed. Dit sal bydra tot besluitnemingsondersteuningstelsels wat gefokus is op die versagting van die gevolge van klimaatsverandering. Digitale grondkartering is egter nog nie wyd aanvaar in Suid-Afrika nie, daarom is daar 'n toenemende behoefte aan navorsing oor hierdie aangeleenthede.Hierdie studie beoog om digitale grondkartering en masjienleer metodologieë te evalueer om gronddiepte in Suid-Afrika te karteer met 'n baie hoër resolusie as wat tans beskikbaar is. Dit poog ook om 'n paar addisionele bekommernisse aan te spreek wanneer hierdie metodologieë toegepas word, soos insetdata, data insamel tegnieke ens. Die twee hoof doelwitte is verdeel in twee eksperimentgroepe. Die eerste eksperimentgroep het die modelleringsvermoë van 'n ‘per-pixel’ benadering tot gronddieptemodellering geëvalueer. Invoerveranderlikes wat verband hou met terrein, Sentinel-2 optiese beeldmateriaal en klimaatdata is in verskeie kombinasies gebruik om hul invloed op gronddiepte te bepaal. Daar is bevind dat hierdie modelle in staat was om gronddiepte met matige prestasie te modelleer, met ruimte vir verbetering. Die ‘per-pixel’ benadering het ook klaarblyklik 'n voorkeur gehad vir hoër resolusie data (10 m) eerder as laer resolusie (30 m). Die tweede eksperimentgroep het 'n soortgelyke proses gevolg, maar het 'n objek-georiënteerde benadering gebruik. Daar is bevind dat hierdie modelle effens meer betroubaar was as die ‘per-pixel’ analises en het minder afhanklikheid van hoër resolusie data getoon om modelle met vergelykbare prestasie te produseer. Die resultate toon dat gronddiepteklassifikasiemodelle met algehele akkuraatheid van 66% geproduseer kan word. Dit toon ook dat gronddiepteregressiemodelle met relatief lae gemiddelde absolute foutwaardes van 288.81 mm geproduseer kan word en in staat is om 38% van die variasie in die verspreiding van gronddiepte te verklaar. Aanbevelings is ook gemaak vir toekomstige navorsing om by te dra tot die verbetering van hierdie prestasie maatstaf.
Description
Thesis (MA)--Stellenbosch University, 2024.
Keywords
Citation