Leveraging DPIRD's WA Soil Data for Improved Machine Learning-Based Soil Mapping and Suitability Assessment
Tracks
| Monday, July 21, 2025 |
| 11:44 AM - 11:56 AM |
Overview
Speaker
Dr Jakob Petereit
Research Scientist
Department of Primary Industries and Regional Development
Leveraging DPIRD's WA Soil Data for Improved Machine Learning-Based Soil Mapping and Suitability Assessment
Abstract
Digital soil mapping (DSM) initiatives, such as the Soil and Landscape Grid of Australia, have significantly advanced our understanding of soil properties at national scales. However, the unique characteristics of Western Australia's ancient, sandy, and deeply weathered soils, combined with limited training data availability, mean there is significant uncertainty when using the national predictions in Western Australia.
Since 1937, the Department of Primary Industries and Regional Development (DPIRD) has surveyed over 80,000 soil sites across Western Australia. Leveraging this repository, we employed machine learning algorithms—XGBoost and Random Forest—to predict soil types pertinent to land suitability assessments.
Our methodology involved training individual models using XGBoost and Random Forest, followed by the development of ensemble models that combine the strengths of both algorithms. As a demonstration, we predict soil classifications and intersect them with landscape positions, which we anticipate will be a precursor to land suitability assessment in Western Australia's diverse environments.
The results demonstrate that these machine learning approaches, underpinned by DPIRD's extensive soil dataset with nested SQL queries to calculate soil types plus input from local knowledge, can produce soil type maps tailored to Western Australia's unique landscapes. This work shows the potential for integrating legacy soil site data with advanced modelling techniques to extend land suitability evaluations. These maps will contribute to the assessment of land management practice sustainability in the rangelands of Western Australia.
Since 1937, the Department of Primary Industries and Regional Development (DPIRD) has surveyed over 80,000 soil sites across Western Australia. Leveraging this repository, we employed machine learning algorithms—XGBoost and Random Forest—to predict soil types pertinent to land suitability assessments.
Our methodology involved training individual models using XGBoost and Random Forest, followed by the development of ensemble models that combine the strengths of both algorithms. As a demonstration, we predict soil classifications and intersect them with landscape positions, which we anticipate will be a precursor to land suitability assessment in Western Australia's diverse environments.
The results demonstrate that these machine learning approaches, underpinned by DPIRD's extensive soil dataset with nested SQL queries to calculate soil types plus input from local knowledge, can produce soil type maps tailored to Western Australia's unique landscapes. This work shows the potential for integrating legacy soil site data with advanced modelling techniques to extend land suitability evaluations. These maps will contribute to the assessment of land management practice sustainability in the rangelands of Western Australia.
Biography
Jakob studied Plant Biotechnology in Hannover, Germany, and completed a PhD in Plant Energy Biology with a focus on biochemistry. He went on to work as a postdoctoral researcher in crop genomics, specialising in bioinformatics, followed by a year in the pharmaceutical industry applying human bioinformatics. He now works in geospatial coding, data science, and machine learning. Jakob has extensive experience in coding, pipeline development, and high-performance computing. His work bridges molecular biology and spatial prediction, with a focus on developing scalable, reproducible solutions for complex biological and environmental data challenges.
