How Fix6 works

The problem at hand

Our mission is to make measuring and verifying carbon sequestration scalable so that land managers of any size can have access to carbon markets. At the moment, median project sizes for agriculture, forestry, and other land use (AFOLU) offset projects is above 30,000 acres. This fundamentally limits who has access to carbon markets.

To allow access to farmers and land managers across the world, we take an approach that allows us to use satellite data to create accurate models of earth and agriculture systems that sequester carbon.

Our approach was born out of our scientific research experience. In the last 20 years, earth system monitoring has gone from a data-scarce to a data-rich field, and we leverage this data to train models that simulate the complex physical processes that sequester carbon. This approach means we can simulate most biospheres, whether it be a tropical forest, arid watershed, or an agricultural field.

If you are starting a carbon removal project, you need to know:

  1. How much carbon could your project sequester over time?
  2. How much carbon did your project sequester since your last measurement?
  3. Do recent measurements change future sequestration rates?

Carbon removal information stack

We utilize three different information stacks in order to determine carbon sequestration:

  1. Data about the project and site. Things from weather, soil types, data previously collected on the site, climate projections, remote sensing data, etc.
  2. The physics controlling carbon sequestration in vegetation. With an understanding of the physics we can determine the amount of carbon being sequestered under different weather conditions, management practices, and project timeframes.
  3. Physics-informed machine learning. Using both the data we have about the project and our understanding of the physics, we can train a digital twin of the carbon removal project in order to look far into the future and understand sequestration over the lifetime of a project.

Data

We gather the following data about the project:

  • Historical weather
  • Climate projections
  • Seasonal forecast (7-months)
  • Soil parameterizations
  • Vegetation
  • Remote sensing data for vegetation and soil response

The reason we collect both historical weather and climate projections is because we simulate sequestration for the entirety of the project. So, in the case of a 30-year project, it can be important to account for potential temperature and precipitation changes on the vegetation that is sequestering carbon.

Physics

We use the collected data in order to simulate carbon sequestration over the project life from first principles. We use an ecohydrologic approach since water, energy, and vegetation cycles are dependent on each other.

For instance, plant growth is affected by the amount of energy and water available in the environment. In response, vegetation also has direct and indirect effects on carbon, water, energy, and nutrient cycling.

These interactions have been studied extensively and now we have an understanding of how these complex feedbacks play out from plant cells and the drought stress of a single tree, to watershed and regional interactions.

Carbon, water, and energy modeling

Through an understanding of the fundamental physics of these cycles, we can simulate the processes contributing to carbon sequestration. We use mathematical representations to model the interactions between soils, carbon, the atmosphere, and vegetation.

We model the water, carbon, and energy cycles from the top of the water table to the top of the canopy/vegetation (see image below). This means that we use atmospheric data (radiation, precipitation, temperature, windspeed, etc.) as inputs to our model, then simulate the sequestration response to that information over the course of years and decades.

For instance, in a drought year, we explicitly model out the reduction of water available in the soil. This lack of water availability decreases the efficiency of photosynthesis. If the drought is extreme enough, we will model out the wilting stresses and the loss of biomass in leaves, for example. Then, when the drought ends the plant will have less foliage which reduces its ability to sequester carbon in the near future.

These complex feedback dynamics can be vital to get an accurate accounting of carbon sequestration, especially when we take into account potential risks to a 10, 20, or 30-year removal contract.

There are multiple forms of these complex dynamics that can play out within a project. For example, the lower elevation areas can have higher soil water content based on subsurface flow, which can improve sequestration efficiency, or the opposite could happen and it could flood an area leading to waterlogged roots and potentially damaging the plant.

Understanding these relationships helps explain the why behind our sequestration figures. This will enable us to develop suggestions and guidance on improving sequestration as we provide MRV for more projects.

Machine learning

Although we can model these interactions, there is no guarantee that the model is accurate. In any modeling, garbage in means garbage out. This is where remote sensing data and machine learning helps us create accurate measurements of sequestered carbon.

When we run our model, we generate not only physical outputs that can be checked against data collected on the ground (soil carbon, vegetation height, tree diameters, etc.), but also indices (NDVI, LAI, soil indices, etc.) which can be compared against remote sensing data.

Traditional vs. physics-informed machine learning

In traditional machine learning, we would collect data that we could turn into features and try to predict the carbon sequestered by the project. Most data that has been historically collected was done so under non-regenerative practices. Also, the things that we can observe from remote sensing are usually only information about the top of the vegetation or soil. To estimate carbon being sequestered in areas not visible to remote sensing, models are made that often make assumptions about the efficiency of carbon sequestration.

In our physics-informed approach to machine learning, we avoid making assumptions about carbon sequestration and instead model out the response of the vegetation to the conditions in the project area. To do this, we treat many inputs and parameters of our physics models as uncertain, so we don't make explicit assumptions about values of important parameters and allow for uncertainty in the input data.

We compare the past five years of remote sensing data with outputs from our physics models. This then lets us determine the model parameters that produce what we actually observed from remote sensing data. This model training process gives us an understanding of the important physical characteristics of the soil, which then gives us better long-range estimates of sequestration.