Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
The effect of measurement error and land management covariates on machine learning derived digital soil maps
Malatji, Boitumelo Raymond. 2024. The effect of measurement error and land management covariates on machine learning derived digital soil maps. Unpublished masters dissertation. Stellenbosch : Stellenbosch University [online]. Available: https://scholar.sun.ac.za/handle/10019.1/131815 Thesis (MScA...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613964926451712 |
|---|---|
| access_status_str | Open Access |
| author | Malatji, Boitumelo Raymond |
| author2 | Van der Westhuizen, Stephan |
| author_browse | Malatji, Boitumelo Raymond Van der Westhuizen, Stephan |
| author_facet | Van der Westhuizen, Stephan Malatji, Boitumelo Raymond |
| author_sort | Malatji, Boitumelo Raymond |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Malatji, Boitumelo Raymond. 2024. The effect of measurement error and land management covariates on machine learning derived digital soil maps. Unpublished masters dissertation. Stellenbosch : Stellenbosch University [online]. Available: https://scholar.sun.ac.za/handle/10019.1/131815
Thesis (MScAgric)--Stellenbosch University, 2024. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/131815 |
| institution | Stellenbosch University (South Africa) |
| last_indexed | 2026-06-10T12:44:30.757Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/131815 The effect of measurement error and land management covariates on machine learning derived digital soil maps Malatji, Boitumelo Raymond Van der Westhuizen, Stephan Clarke, Catherine E. Stellenbosch University. Faculty of AgriSciences. Dept. of Soil Science. Digital soil mapping -- Data processing Machine learning Errors -- Measurement Soil organic carbon Errors-in-variables models Land use -- Management UCTD Malatji, Boitumelo Raymond. 2024. The effect of measurement error and land management covariates on machine learning derived digital soil maps. Unpublished masters dissertation. Stellenbosch : Stellenbosch University [online]. Available: https://scholar.sun.ac.za/handle/10019.1/131815 Thesis (MScAgric)--Stellenbosch University, 2024. ENGLISH ABSTRACT: Digital soil mapping (DSM) has become a fundamental tool used to create detailed spatial predictions of soil properties such as soil organic carbon (SOC) and clay content. Despite its advantages, its accuracy can be hindered by various sources of error, such as measurement errors. When these errors are not accounted for, they propagate to the DSM outputs, resulting in soil maps that are not measurement error-free. Since these soil maps help in making informed decisions in the agricultural sciences, and in policy making, it is important to account for these errors. Moreover, these soil properties vary across depths, and as depth-specific soil data is crucial for a thorough understanding of soil formation processes and for assessing potential environmental impacts, these properties are predicted at different soil depths. Therefore, models that are better in predicting these properties are necessary for obtaining accurate predictions. In addition, land management practices have a significant effect on the spatial distribution of SOC in farmlands. Therefore, incorporating this information as covariates for SOC predictions can be vital for improving the prediction accuracy of SOC maps. This study investigated the effect of accounting for measurement errors on digital maps of SOC and clay prediction, and the effect of land management covariates on SOC prediction using machine learning (ML). Moreover, this study compared the model performance between a depth-specific modelling and a three-dimensional (3D) modelling that uses depth as a covariate. This study also investigated how incorporating land management covariates influenced the prediction of SOC. The effect of measurement error was accounted for by incorporating measurement error variances (MEVs) derived from the mid-infrared (MIR) repeats, as weights in the calibration of a ML model. In this way measurements with larger MEVs receive less weight when a ML model is fitted. This was done based on SOC and clay, for 0-10 cm, 10-30 cm, and 30-50 cm depths, by comparing random forest model (RF) with measurement error-filtered RF model (FRF). The results showed that FRF outperformed RF for both SOC and clay. Moreover, a depth-specific model was compared with a 3D model, and the results showed that both modelling approaches produced comparable results for both SOC and clay. However, the results also showed that model performance decreased with increasing soil depth for both SOC and clay using both modelling approaches. The effect of land management covariates on the prediction of SOC was assessed by comparing different sets of covariates with and without land management covariates. Three land management covariates used were land management practices, land use, and crop ecotope. The results showed that land management covariates improved model performance in the 0-10 cm depth. Farm and land use were among the most important covariates in predicting SOC at this depth. However, no model improvement was observed in the 10-30 cm and 30-50 cm. The overall results of this study showed that accounting for measurement errors are vital for improving model prediction. In addition, incorporating land management covariates in predicting topsoil SOC is important for obtaining better results. AFRIKAANSE OPSOMMING: Digitale grondkartering (DGK) het 'n fundamentele hulpmiddel geword wat gebruik word om gedetailleerde ruimtelike voorspellings van grondeienskappe soos grondorganiese koolstof (GOK) en klei-inhoud te skep. Ten spyte van die voordele van DGK, kan die akkuraatheid daarvan deur verskeie bronne van foute, soos meetfoute, belemmer word. Wanneer hierdie foute nie in ag geneem word nie, dra dit oor na die DGK-afvoer, wat lei tot grondkaarte wat nie meetfoutvry is nie. Aangesien hierdie grondkaarte gebruik word om ingeligte besluite in die landbouwetenskappe en beleidsvorming te neem, is dit belangrik om hierdie foute in ag te neem. Verder wissel hierdie grondeienskappe oor diepte, en aangesien dieptespesifieke gronddata noodsaaklik is vir 'n deeglike begrip van grondvormingsprosesse en vir die assessering van moontlike omgewingsimpakte, word hierdie kenmerke op verskillende grond-dieptes voorspel. Daarom is modelle wat beter is in die voorspelling van hierdie kenmerke nodig om akkurate voorspellings te verkry. Boonop het grondbestuurspraktyke 'n beduidende invloed op die ruimtelike verspreiding van GOK in landerye. Daarom kan die inkorporering van hierdie inligting as kovariate vir GOK-voorspellings van kritieke belang wees vir die verbetering van die voorspelling-akkuraatheid van GOK-kaarte. Hierdie studie het die effek van die inagneming van meetfoute op digitale kaarte van GOK- en klei-inhoud voorspellings ondersoek, asook die invloed van grondbestuur-kovariate op GOK-voorspellings deur middel van masjienleer (ML). Verder het hierdie studie die modelprestasie vergelyk tussen 'n dieptespesifieke modellering en 'n driedimensionele (3D) modellering wat diepte as 'n kovariaat gebruik. Die studie het ook ondersoek ingestel na hoe die inkorporering van grondbestuur-kovariate die voorspelling van GOK beïnvloed het. Die effek van meetfoute is in ag geneem deur meetfoutvariansies (MFV's), afgelei van midinfrarooi (MIR) herhalings, as gewigte in die kalibrasie van 'n ML-model in te sluit. Op hierdie manier ontvang metings met groter MFV's kleiner gewig wanneer 'n ML-model gepas word. Dit is gedoen vir GOK en klei, vir dieptes van 0-10 cm, 10-30 cm, en 30-50 cm, deur 'n ewekansige ‘random forest’ (RF) te vergelyk met 'n meetfout-gefiltreerde RF-model (FRF). Die resultate het getoon dat FRF beter gevaar het as RF vir beide GOK en klei. Verder is 'n dieptespesifieke model vergelyk met 'n 3D-model, en die resultate het getoon dat beide modelleringsbenaderings vergelykbare resultate opgelewer het vir beide GOK en klei. Die resultate het egter ook getoon dat modelprestasie afgeneem het met toenemende grond-diepte vir beide GOK en klei met beide modelleringsbenaderings. Die effek van grondbestuur-kovariate op die voorspelling van GOK is beoordeel deur verskillende stelle kovariate met en sonder grondbestuur-kovariate te vergelyk. Drie grondbestuur-kovariate wat gebruik is, was grondbestuurspraktyke, grondgebruik en gewasekotoop. Die resultate het getoon dat grondbestuur-kovariate modelprestasie verbeter het in die 0-10 cm diepte. Plaas- en grondgebruik was onder die belangrikste kovariate in die voorspelling van GOK op hierdie diepte. Geen modelverbetering is egter waargeneem in die 10-30 cm en 30-50 cm dieptes nie. Die algehele resultate van hierdie studie het getoon dat die inagneming van meetfoute noodsaaklik is vir die verbetering van modelvoorspellings. Daarbenewens is die inkorporering van grondbestuur-kovariate in die voorspelling van bogrondse GOK belangrik om beter resultate te verkry. Masters 2025-03-31T07:02:03Z 2025-03-31T07:02:03Z 2024-12 Thesis https://scholar.sun.ac.za/handle/10019.1/131815 Stellenbosch University xiii, 85 pages : illustrations, maps application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Digital soil mapping -- Data processing Machine learning Errors -- Measurement Soil organic carbon Errors-in-variables models Land use -- Management UCTD Malatji, Boitumelo Raymond The effect of measurement error and land management covariates on machine learning derived digital soil maps |
| title | The effect of measurement error and land management covariates on machine learning derived digital soil maps |
| title_full | The effect of measurement error and land management covariates on machine learning derived digital soil maps |
| title_fullStr | The effect of measurement error and land management covariates on machine learning derived digital soil maps |
| title_full_unstemmed | The effect of measurement error and land management covariates on machine learning derived digital soil maps |
| title_short | The effect of measurement error and land management covariates on machine learning derived digital soil maps |
| title_sort | effect of measurement error and land management covariates on machine learning derived digital soil maps |
| topic | Digital soil mapping -- Data processing Machine learning Errors -- Measurement Soil organic carbon Errors-in-variables models Land use -- Management UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/131815 |
| work_keys_str_mv | AT malatjiboitumeloraymond theeffectofmeasurementerrorandlandmanagementcovariatesonmachinelearningderiveddigitalsoilmaps AT malatjiboitumeloraymond effectofmeasurementerrorandlandmanagementcovariatesonmachinelearningderiveddigitalsoilmaps |