Predicting soil properties using machine learning

: a smartphone solution for stakeholders

Abstract

Soil health is foundational to agricultural productivity, ecosystem sustainability, and climate regulation, yet traditional methods for assessing soil properties are labour-intensive, costly, and limited in spatial resolution. This thesis addresses these limitations by integrating machine learning and Digital Soil Mapping (DSM) with smartphone-based imagery to enhance the accuracy, accessibility, and affordability of soil property assessment. Using strategically sampled field data from diverse Scottish environments, detailed laboratory analysis established a robust baseline dataset for key soil attributes, including pH, organic carbon, moisture content, bulk density, and texture.

Advanced machine learning algorithms were developed and optimised to predict soil properties across Scotland, significantly increasing spatial resolution and capturing micro-scale variability. Incorporation of smartphone-captured soil images analysed through Convolutional Neural Networks (CNNs) further improved prediction accuracy, demonstrating the substantial value of visual soil characteristics in modelling efforts. The suitability of CNN-derived features was rigorously tested against traditional features, both with and without calibration, confirming that CNN features provided superior performance in predictive accuracy. The resulting high-resolution soil property maps, with quantified uncertainty intervals, represent a transformative improvement over conventional soil maps, enabling targeted agricultural and conservation strategies.

A user-friendly smartphone application was created to operationalise these advancements, empowering farmers, policymakers, and land managers with instant, site-specific soil information. Users can obtain site-specific soil property predictions by simply capturing a geotagged soil image. In cross validated tests, our best models achieved R² of 0.64 for bulk density, 0.61 for moisture content, 0.39 for pH, and 0.50 for LOI.

This research advances the science of digital soil mapping and directly contributes to sustainable land management practices, precision agriculture, and environmental conservation, offering a replicable and scalable framework suitable for global adoption in diverse geographical settings.

Date of Award	3 Nov 2025
Original language	English
Awarding Institution	Abertay University
Sponsors	Norman Fraser Design Trust
Supervisor	Ehsan Jorat (Supervisor), Thomas Howson (Supervisor) & Matt Aitkenhead (Supervisor)