P.PSH.2103 - Utilising hybrid-approaches of hyperspectral remote sensing for soil organic carbon monitoring
Using just six soil samples and hyper-spectral imagery, a machine learning model can be trained to estimate soil carbon at any point on the landscape within a regional carbon estimation area.
Project start date: | 07 May 2021 |
Project end date: | 30 April 2023 |
Publication date: | 23 April 2024 |
Project status: | Completed |
Livestock species: | Grain-fed Cattle, Grass-fed Cattle, Sheep, Goat, Lamb |
Relevant regions: | National |
Download Report
(3.1 MB)
|
Summary
Carbon sequestration in livestock producers’ soils has the potential to be a significant contributor to the red meat industry achieving carbon neutrality. A critical first step in achieving this is to rigorously baseline individual farms’ current soil carbon stocks in a cost-effective way that scales to an industry wide level. The objective of the project was to use data from space to reduce the number and spatial density of costly physical soil samples required to still have confidence in the accuracy of the results.
The project demonstrated that in target regions, by combining and ground truthing data from space with as few as 6–9 physical soil samples per farm, soil carbon stock levels can be accurately quantified to provide results that producers and the red meat industry can have confidence in reporting.
Objectives
Identify lower cost pathways for measuring soil carbon at scale.
Validate US-based Cloud Agronomics hybrid-modelling technology in eastern Australia. The hybrid approach uses local soil samples, hyper-spectral imagery and machine learning to create regional soil carbon estimation maps with a high degree of accuracy.
Conduct validation trial in a manner that complies with the Measure-Model-Measure requirements of the Emissions Reduction Fund.
Key findings
The optimal number of physical soil samples needed to train a local model per CEA are six, with diminishing returns on accuracy beyond nine samples.
Benefits to industry
Hybrid remote and local soil carbon measurement technology seeks to combine the scale of remote sensing, combined with the accuracy of physical soil sampling. By using remote-imagery, in-field soil samples together with the power of machine learning, a hybrid approach could deliver large scale soil carbon mapping at a fraction of the cost of traditional measurement and with enhanced accuracy on purely remote-sensed methods.
The benefits to industry are more complete and comprehensive soil carbon measurement, with the potential for scale across all land under livestock management. The would allow for a more refined measure of carbon stored in agricultural soils, for industry level greenhouse gas reporting, payment for environmental services schemes and on-farm management decisions.
MLA action
The key learning of this project was to determine the number of data points (physical soil samples) required to optimally train a machine learning model to estimate soil carbon. This project has indicated that six samples are optimal, with diminishing returns after nine samples. This is important guidance as technology providers seek to develop scalable digital methods of SOC estimation.
Future research
The project team is pursuing funding through the Federal Government National Soil Innovation Challenge to map 1 million hectares of grazing land using their hybrid modelling. This seeks to develop a comprehensive, accurate and cost-effective soils map for the red meat industry.
The project has also created a subsequent pathway to prove and establish this approach at a nationwide scale in alignment with Australia’s soil carbon methodology and work towards achieving the Commonwealth Governments target of $3/ha quantification.
More information
Project manager: | Julia Waite |
Contact email: | reports@mla.com.au |
Primary researcher: | Cloud Agronomics Inc |