Advancing Soil Carbon Quantification with Remote Sensing and Machine Learning 

We’re excited to release a newly developed soil organic carbon model to Australian users for use in stratification and sample planning.

This marks a significant milestone in our efforts to reduce the cost of soil carbon measurement as part of the National Soil Carbon Innovation Challenge (NSCIC). The ‘KTG’ or ‘Kick the $3/Ha Goal’ project is leveraging remote sensing and machine learning (RS/ML) to provide accurate, cost-effective soil organic carbon (SOC) quantification across Australia’s vast landscapes, reducing the reliance on extensive physical soil sampling.

Addressing Traditional Challenges

Traditional soil carbon stratification and sampling methods are variable, costly (often requiring a ‘pre-baseline’ to determine accurate stock quantities), and limited in scope. Stratification using layers that don’t compare well to soil carbon stock can result in poor sampling performance, and a reduction in overall carbon offsets generated under both voluntary and regulatory carbon offset protocols. In some cases, poor stratification may entirely negate any increase in soil carbon stocks due to a high measurement error.

The initial “Zero point” model overcomes these challenges by integrating extensive soil sample results from the NSCIC program with advanced RS/ML models, significantly enhancing SOC prediction accuracy and efficiency. Using high-resolution Sentinel-2 satellite data, the model captures fine-scale variations in SOC at a 10-meter resolution, providing detailed insights into the top 30 cm layer, which is highly susceptible to climate change and land use practices.

The first major milestone developed two initial “Zero Point” models for percentage soil organic carbon (SOC %) across Australia for initial Beta release. These models are now available to FarmLab users for stratification and sample planning in soil carbon projects.

Accuracy and Efficiency

Since the development of the model in December 23, the FarmLab project team have been reviewing it, in conjunction with the Australian Clean Energy Regulator’s random stratification sample method, to assess whether the model is successful in identifying areas with higher, medium, and lower levels of soil carbon. This method allows for the calculation of SOC stock variance, providing a measure of the model’s accuracy for random stratified sampling – the mandatory method for soil carbon offset projects in Australia.

The results are promising:

Over the course of the review the model has shown to reduce soil carbon variance by 50% compared to traditional stratification using the Topographical Wetness Index (TWI).

Farm Number of Samples Size of Area (ha) Number of  strata 0-30cm SOC Stock (t/ha) 0-30cm Variance (SOC t/ha) 
Farm A18 726.96 34.847 8.625 
Farm B18 627.29 29.461 2.05 
Farm C18 210.57 27.188 2.77 
Farm D17 1140 43.406 8.66 
Farm E18 353.75 35.334 2.176 

The results from the variance analysis can be found in the above table; the average variance over five farms using our model for random stratification is 4.85 tonnes of SOC per hectare to a depth of 30 cm, with a sampling density of 1 sample per 35 hectares.

In particular the model seems to appear better for areas less than 600 hectares (~1200 acres). In contrast, farms stratified with TWI show an average variance of 9.7 tonnes of SOC per hectare. It is expected the project will sample a further 25 farms using the model, with further analyses shared following the return of the results.

The model appears to perform on farms under 500 hectares, requiring only 18 samples across these sites. The next phase of the project involves translating this increased confidence into reduced costs for farmers and carbon project developers. For a typical 500-hectare farm (based on 18 soil cores down to 100 cm), this efficiency translates to significant cost savings, reducing the sampling cost to $10 per hectare for a 500-hectare project, which includes the cost of stratification and sampling/lab testing.

Innovative Methodology

Our methodology involves a two-stage modeling process:

  • Stage 1: Sentinel-2 data is used to generate a high-resolution temporal and spectral dynamics model, predicting SOC distribution before physical sampling.
  • Stage 2: These predictions are refined by incorporating localized spatial data, using a hybrid AI and deep learning approach to account for spatial autocorrelations and improve accuracy.

This approach not only enhances prediction precision but also offers valuable uncertainty quantification, crucial for effective decision-making.

Practical Applications and Benefits

Operational results have been integrated into the FarmLab platform, providing consultants with a powerful tool for stratifying and assessing soil carbon. By stratifying farms based on predicted SOC content, our project facilitates targeted and efficient soil sampling, reducing costs and improving the representativeness of samples. This strategic approach aligns with the Clean Energy Regulator’s mandatory soil carbon measurement requirements and provides a significant step forward in reducing the cost of soil carbon measurement.

Future Outlook

Our project demonstrates the potential of combining remote sensing, machine learning, and robust fieldwork. Future project developments include using the model to conduct region-wide quantification, which is currently underway across Kangaroo Island in South Australia.

As we continue to receive data from the KTG project, we will release further updates and improvements to the model. With ongoing support from the Australian Government, we are poised to make a significant impact on soil carbon quantification and climate change mitigation efforts.

Stay tuned for more updates as we continue to develop new methods for soil carbon measurement in our support of a more sustainable future for agriculture.

Acknowledgements:

  • Dale Roberts (FarmLab Chief Data Scientist) for your work developing the models, sorting through the data and processing over 774 billion pixels across Australia!
  • Our incredible project partners: UTAS, Perennial (formerly CloudAg), APAL Labs and all of our amazing soil samplers!
  • Senani Karutnaratne (CSIRO Senior Research Scientist) for your guidance on modeling SOC stocks and percentages.
  • Department of Primary Industries and Regions, South Australia for your support in sampling, particularly across Kangaroo Island.
  • The Department of Industry, Science, and Resources for sponsoring and funding this groundbreaking work.
  • All the farmers involved in the NSCIC challenge for lending their data to science!
Scroll to Top