Periodic Reporting for period 2 - SEQANA (SEQANA GmbH)
Reporting period: 2023-03-01 to 2024-02-29
Industrial farming and other anthropogenic land management practices have caused widespread land and soil degradation around the world, resulting in greenhouse gas emissions, loss of biodiversity, damage to soil health and fertility, and decreased soil water retention capacity. Ultimately all this is threatening food security. However, this trend can be stopped and even reversed by increasing soil organic carbon levels. Soil organic carbon decline is a key indicator of land degradation, and its levels can be increased through traditional and regenerative approaches to farming and land management. Additionally, soil organic carbon sequestration can be monetized as a carbon offset on the voluntary carbon market. This will provide a financial incentive for landowners and farmers to adopt more regenerative practices. However, this process requires accurate monitoring, reporting and verification of soil organic carbon sequestration. Seqana offers cost-effective monitoring, reporting and verification of soil organic carbon sequestration at scale through machine learning models trained on a large database of existing soil organic carbon samples and satellite-based remote sensing data.
We have assembled a large database with millions of soil data records describing soil organic carbon levels in various locations in the world at different points in time. In our machine learning models we combine these target reference data, which derive from taken soil samples of the ground that were analyzed in the laboratory for their soil organic carbon content, with predictors such as satellite data and other environmental covariates of soil organic carbon, such as topography, weather and climate data. To access these predictors we have built a predictor data access pipeline that enables us to acquire predictor data for any particular location and time. With both target and predictor data we can calibrate and evaluate our machine learning models. To properly evaluate the models, we need a benchmark to compare our model performance against. In our case this benchmark is set by traditional soil sampling. Our model performance needs to achieve similar levels of accuracy and uncertainty in quantification as the traditional approach, while offering a faster and more cost-effective way of quantifying the changes. No test environment existed to be able to meaningfully compare the traditional method of soil sampling with other methods of quantifying soil organic carbon stock changes in this way. We have developed a framework with a set of metrics to directly compare the accuracy and uncertainty of traditional sampling with modeling-based quantification approaches. We continuously test our models in this test environment against the traditional sampling benchmark in an experimentation framework which we have developed to quickly test hypotheses with respect to model performance and to iteratively improve our models.
With these continuous improvements we were able to already deploy our soil organic carbon stock quantification approaches in pilot projects with clients. In comparison with traditional sampling approaches, those pilot projects showcased the high levels of accuracy in combination with the quick accessibility and low-cost of our solution.
With these continuous improvements we were able to already deploy our soil organic carbon stock quantification approaches in pilot projects with clients. In comparison with traditional sampling approaches, those pilot projects showcased the high levels of accuracy in combination with the quick accessibility and low-cost of our solution.
Prior to our project, there was no automated, reproducible way to quantify changes in soil organic carbon stock at scale using remote sensing data. We have developed a novel approach that surpasses the previous state of the art. Additionally, we have created a method for directly comparing the accuracy and uncertainty of measurements obtained through traditional soil sampling of soil organic carbon with those obtained through regression modeling of soil organic carbon stock levels and changes. We have also introduced the concept of "domains" to distinguish subpopulations of soil organic carbon regimes based on geographic scope, soil depth, soil organic carbon range, and other factors such as land cover and use. We have integrated all of these findings to develop new machine learning models that use satellite data and environmental covariates to quantify soil organic carbon stock levels and changes, and evaluate these models against the benchmark of traditional sampling in different soil organic carbon domains. Pilot projects have demonstrated that this approach can be a breakthrough for monitoring, reporting, and verification of carbon credits earned for soil organic carbon sequestration.