Machine learning and uncertainty quantification in soil science and agricultural land economic

Li, Shuai

Machine learning and uncertainty quantification in soil science and agricultural land economic

Li, Shuai

Permalink

Publication Type:: Thesis
Issue Date:: 2025

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download thesisAdobe PDF (14.3 MB)

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	Li, Shuai
dc.date.accessioned	2026-04-20T04:21:31Z
dc.date.available	2026-04-20T04:21:31Z
dc.date.issued	2025
dc.identifier.uri	http://hdl.handle.net/10453/194768
dc.description	University of Technology Sydney. Faculty of Engineering and Information Technology.	en_US.UTF-8
dc.description.abstract	This thesis integrates three research components using machine learning and spatial modelling to address major challenges in agricultural prediction and land valuation. Although data-driven methods are increasingly common, few studies jointly model uncertainty and spatial heterogeneity across soil, crop yield, and farmland markets. This research develops a unified probabilistic and spatial framework to fill that gap. The first component predicts Soil Organic Carbon (SOC) using five models—linear regression, neural networks, XGBoost, variational Bayesian linear regression, and variational Bayesian neural networks—based on accessible soil properties such as bulk density, pH, and texture. Models were evaluated with and without nitrogen data, showing nitrogen is essential and that most models become overconfident when key inputs are missing. Variational Bayesian neural networks offered the best balance between accuracy and uncertainty quantification. The second component predicts maize yields for 842 U.S. Corn Belt counties (2014–2023) using soil, climate, remote-sensing, and phenological variables. Four models were assessed: linear regression, GWR, VB-GWR, and a neural-augmented VB-GWR. VB-GWR performed best, capturing spatially varying relationships and predicting yields within 150–220 bu/acre. pH emerged as the strongest predictor, while precipitation had limited impact. The third component examines the spatial determinants of farmland rental prices using GWR with macroeconomic drivers such as GDP, oil prices, and Treasury yields. Oil prices showed strong positive effects in ethanol-producing regions, while maize yield had weaker influence in core production areas. Uncertainty analysis indicated limited spatial variation, suggesting economic factors dominate rent dynamics. The model projects rental prices for 2026–2027, with an average annual increase of about 10.40%. Overall, the thesis demonstrates the value of integrating machine learning, spatial analysis, and probabilistic modelling, offering a coherent framework for predicting soil carbon, crop yield, and farmland rent under uncertain and spatially heterogeneous conditions.	en_US.UTF-8
dc.format	Thesis (ME)
dc.language.iso	en_US	en_US.UTF-8
dc.relation	https://opus.lib.uts.edu.au/bitstream/10453/194768/1/thesis.pdf
dc.rights	info:eu-repo/semantics/openAccess
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	© 2025 Shuai Li
dc.rights	au.edu.uts.lib/cph
dc.title	Machine learning and uncertainty quantification in soil science and agricultural land economic	en_US.UTF-8
dc.type	Thesis
utslib.copyright.status	open_access	*

Abstract:

This thesis integrates three research components using machine learning and spatial modelling to address major challenges in agricultural prediction and land valuation. Although data-driven methods are increasingly common, few studies jointly model uncertainty and spatial heterogeneity across soil, crop yield, and farmland markets. This research develops a unified probabilistic and spatial framework to fill that gap. The first component predicts Soil Organic Carbon (SOC) using five models—linear regression, neural networks, XGBoost, variational Bayesian linear regression, and variational Bayesian neural networks—based on accessible soil properties such as bulk density, pH, and texture. Models were evaluated with and without nitrogen data, showing nitrogen is essential and that most models become overconfident when key inputs are missing. Variational Bayesian neural networks offered the best balance between accuracy and uncertainty quantification. The second component predicts maize yields for 842 U.S. Corn Belt counties (2014–2023) using soil, climate, remote-sensing, and phenological variables. Four models were assessed: linear regression, GWR, VB-GWR, and a neural-augmented VB-GWR. VB-GWR performed best, capturing spatially varying relationships and predicting yields within 150–220 bu/acre. pH emerged as the strongest predictor, while precipitation had limited impact. The third component examines the spatial determinants of farmland rental prices using GWR with macroeconomic drivers such as GDP, oil prices, and Treasury yields. Oil prices showed strong positive effects in ethanol-producing regions, while maize yield had weaker influence in core production areas. Uncertainty analysis indicated limited spatial variation, suggesting economic factors dominate rent dynamics. The model projects rental prices for 2026–2027, with an average annual increase of about 10.40%. Overall, the thesis demonstrates the value of integrating machine learning, spatial analysis, and probabilistic modelling, offering a coherent framework for predicting soil carbon, crop yield, and farmland rent under uncertain and spatially heterogeneous conditions.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/194768