Learn about Regrow's methodology for quantifying uncertainty
Uncertainty in MonitorML Practice Determinations
What is Uncertainty and How Do We Measure It?
Uncertainty in data refers to the degree of variability and unknowns in the information generated. In the context of agricultural practice determinations, uncertainty arises from multiple sources and affects the confidence we have in the reported metrics.
We categorize uncertainty into three primary types:
-
Input Uncertainty – Stemming from the raw data sources used in the model. This includes satellite-derived observations, which can be affected by factors such as cloud cover, resolution limitations, and signal inconsistencies.
-
Parametric Uncertainty – Related to the internal parameters used to process and interpret data. Examples include the applied thresholds, constants, factors, and coefficients used to classify practices like cover cropping from remote sensing imagery, and residue thresholds used to classify tillage practices.
-
Structural Uncertainty – Emerging from the methodology and model design itself. This refers to ambiguity in how different models might interpret the same data differently or the limitations of the chosen modeling approach.
By acknowledging and quantifying these uncertainties, we enhance decision-making by providing a clearer representation of possible outcomes rather than treating data as an absolute truth.
How is This Approach Different from Our Current Methodology?
Today, our MonitorML operates in a deterministic manner—meaning that the same inputs always produce the same outputs. While this ensures reproducibility, it does not account for uncertainty. With this new effort, we are introducing uncertainty quantification and propagation to enrich our data interpretations. This shift enables us to:
-
Capture and express confidence levels around practice determinations.
-
Provide a probabilistic view of field-level determinations rather than a binary classification.
-
Enable users to make more informed decisions by understanding the variability in data-driven insights.
How Do We Quantify MonitorML Uncertainty?
We are quantifying uncertainty in the practice determinations made by MonitorML at the field level and propagating that uncertainty to select Sustainability Insights (SI) KPIs.
We recognize that the view of a field from a satellite is imperfect and that certain pixel-level features in satellite imagery may be difficult to interpret. Furthermore, it’s possible that only certain parts of the field may be visible when the satellite makes an overpass due to, for example, cloud cover. These satellite views are directly used in determining the regenerative agricultural practice potentially performed on a given field. These practice determinations are then used in various reporting metrics surfaced through SI.
There are other sources of uncertainty that affect the metrics surfaced through SI, but are not addressed with this initial work. Some examples include uncertainty in the crop identification for a field, the actual field boundaries, and thresholds/parameters used in the practice determinations. Additionally, uncertainty in Monitor data is not currently propagated as input uncertainty to DNDC.
To capture uncertainty in practice determinations, we leveraged existing data archives and intermediate calculations:
-
Green Cover/Cover Crop: We use the monthly mean and standard deviation of monthly NDVI values to estimate the variability in field-level vegetation cover, treating these as inputs to a truncated normal distribution.
Fields with a small standard deviation imply that the satellite observations were consistent across the observed pixels within the field boundaries. A large standard deviation would imply greater variability across those pixels and thus a greater potential for a different cover crop classification, which is reflected by higher uncertainty.
-
Tillage Practices: We interpolate weekly percentile-based residue cover distributions to infer variability, leveraging multiple percentile values to construct a reasonable approximation of field conditions.
Similar to green cover, a smaller spread in the percentile values indicates less variability in the residue cover observed on the field that week. A larger spread in the percentile values would indicate less consistency in the observed residue cover and thus higher uncertainty in the reported tillage practice classification.
How Should Users Interpret This Data?
Users should view uncertainty as an added layer of insight. Key takeaways include:
-
Confidence Ranges: Higher uncertainty suggests greater variability in practice determinations, which may warrant closer inspection.
-
Comparative Insights: Uncertainty levels can help compare fields or regions, prioritizing areas where more data validation may be required.
-
Decision Support: Rather than viewing data as binary (true/false), users should consider the probability associated with each determination to guide strategic planning.
Uncertainty in DNDC Outcome Quantification
A few words about Regrow's DNDC
Regrow models on-farm emissions using the Denitrification-Decomposition (DNDC) model, which is a biogeochemical model that is based on first principles of soil biogeochemistry and estimates nutrient cycling in the soil, including dynamic soil responses to practice change adoption. The model predicts greenhouse gas emissions, soil carbon sequestration, and other environmental outcomes and feedbacks of crop production, such as crop growth and yield, based on a series of environmental drivers (crop management, weather, and soil data, cultivar etc.).
Calculation of emissions begins with monitoring a continuous series of agricultural practices at the crop field scale that both inform and drive DNDC. For more information on MonitorML, Regrow's proprietary technology to identify field-level historical events and create crop field boundaries, please see here. Earth observation satellite data along with local climate and weather inputs are compiled over each and every field boundary and implemented in Regrow algorithms to identify agricultural practices or events.
Each practice or event, and the compounding factor of multiple events, will affect emissions. For example, adoption of multiple practices such as crop rotations, planting consistent cover crops, and limited to no tillage will have positive effects on soil health and be reflected in the estimated emissions. This comprehensive field level history is passed to the DNDC soil biogeochemical model, which quantifies the effects of these practices on critical soil properties and processes, so that Regrow can simulate GHG emissions (measured in tonnes of carbon dioxide equivalent, or CO2e) consisting of the following components: CH4, N20 (indirect/direct), and soil organic carbon (SOC). Components are converted to CO2e by their global warming potential following IPCC 2014 AR5 values.
N2O emissions from DNDC cannot be explicitly traced to a specific nitrogen source (i.e. fertilizer N vs mineralized N). As such, direct and indirect N2O emissions as displayed in the product are inclusive of all nitrogen sources in the soil, not exclusively that from fertilizer application.
How Do We Quantify DNDC Outcome Uncertainty?
Regrow has taken a standard error approach to calculating uncertainty for emissions and removals values in Sustainability Insights. Within emissions, standard error is provided for each constituent component of GHG emissions (CH4, direct N2O, indirect N2O) for every region and crop of interest. We surface uncertainty in the form of 1 standard deviation from the mean.
This is a measure of the variability of the estimated emissions that indicates how much the estimated emissions deviate from a true population regional mean. The variability of standard error will increase as supply shed size decreases.
This calculation methodology is a type of scenario uncertainty, as described in Section 16.6.1 of the draft GHGp Land Sector and Removals Guidance (LSRG). It is important to note that SI uncertainty does not provide an estimate of DNDC structural uncertainty that Regrow uses in the MRV product.