Feature Photo by Marie-Soleil Turmel

Background

When we started Our Sci in 2017, one of the first tools we built was a low-cost, handheld Reflectometer. The Reflectometer was used by the Quick Carbon project at Yale University to map soil organic carbon (SOC) at the landscape scale, by the Bionutrient Institute to predict nutrients in food crops (e.g. antioxidant content in lettuce) and by international research groups to assess soil carbon content on smallholder farms in Malawi.

Over the past several years, manufacturing of the Reflectometer has been stalled due to chip and other supply chain shortages. We used that time to address limitations observed in early versions of the Reflectometer. We added normalization factors to each device during the manufacturing process to reduce device-to-device variability and developed an in-field calibration to improve device consistency over time. This year we have been able to manufacture more than 80 new Reflectometers, with more to follow in the coming months.

Now that we have new Reflectometers available, we’re working with Catholic Relief Services (CRS) and the International Maize and Wheat Improvement Center (CIMMYT) to deploy and test the Reflectometer as a decision support tool in numerous countries in southern Africa and Central America. The Reflectometer is appealing to these organizations because its low cost ($500) enables scalability and it can return results in real time, even without internet connection. Real time, offline functionality is critical for stakeholders who may be in difficult-to-reach areas without any internet connection.

Specifically, our partners are interested in how well the Reflectometer can provide two outcomes:

• Classify soils – Readily identifying degraded soils, as well as soils with above-average soil carbon, can facilitate site-specific discussions about soil health and help development agency field staff and extension agents provide tailored recommendations to smallholder farmers.
• Estimate soil carbon content – Low-cost quantification of soil carbon can help organizations better evaluate the impacts of land restoration projects. 

How It Works

Like any spectroscopy tool, the Reflectometer correlates measured soil carbon values to spectral patterns (in this case reflection) from the Reflectometer in a training dataset. The first step is to develop a calibration dataset in the region of interest. In both peer-reviewed papers linked above, Reflectometer-based models performed better when paired with other soil and field metadata. We recommend the following data be included in a calibration dataset if possible:

1. Lab-derived soil organic carbon values.
2. Reflectance spectra measured with the Our Sci Reflectometer.
3. Soil texture data, either soil texture classes determined using hand texturing methods or lab derived clay, sand, and silt content.
4. Estimated or measured slope values of the sampled location.
5. GPS coordinates of the sampled location. These may be used to pull in metadata from digital soil maps and other resources.
6. Other metadata that may be critical to classifying soils include the region where the sample was collected (if there are significant regional differences in soil carbon), presence of carbonates, etc.

Once the training set is complete, we develop two distinct types of model outputs:

• Soil classification. Classification models can often be more accurate and simple than models that predict soil carbon content and are customizable to output messaging specific to the development program or stakeholders. We engage with local experts to develop locally appropriate soil classifications and recommendations (see table below). For example, in one area adding compost may be the appropriate recommendation for rehabilitating degraded soils while in another region/project green manures or doubled-up legumes may be the most appropriate practice.
• Soil carbon content. Soil carbon values (%) are generated using random forest, multi-linear regression, or other model types. These values are most useful to organizations tracking soil carbon change over time, often when evaluating the impacts of land restoration activities.

The Field Process

The Reflectometer connects to Android devices using the SurveyStack app. SurveyStack is an open-source survey platform designed to empower shared community knowledge. Partnering organizations use SurveyStack’s group management features to manage access to their surveys and data and use the question set library to develop project-specific surveys. SurveyStack’s custom javascript scripting environment enables project teams to deploy complex soil classification and soil carbon prediction models within the field survey, so that the entire process occurs in a simple linear workflow.

Recently, I had the pleasure of traveling to Guatemala City to meet with and train CRS staff and partners (pictured above and below). Prior to my trip, Our Sci and CRS had collaborated to develop a training dataset and build soil carbon models. The training focused on the Reflectometer, SurveyStack, and the in-field process:

1. When arriving at the 1^st field of the day, walk through the In-Field Calibration.
2. Answer questions about the field and estimate the slope.
3. Collect a soil sample.
4. Hand-texture the soil.
5. Scan the soil with the Reflectometer.
6. Run the script to classify the soil and predict soil carbon content based on local agronomic expertise.
7. Discuss the soil results and make tailored recommendations to smallholder farmers and land stewards.

The in-person training was a great opportunity to meet with the local experts who will be using the Reflectometer in the field. Together, we will customize the messaging output from the classification models and they will be ready to deploy the Reflectometer to offer decision support to smallholders farmers in Guatemala.

Next Steps

There are numerous areas where we hope to continue to improve the utility of the Reflectometer:

• Improve the accuracy of soil carbon prediction models. Currently, the focus is on fully offline compatible models. However, we may find that the best way to increase accuracy is to split up the field process so that only the classification model is run in the field, offline. Then, after the data has been submitted to SurveyStack, partners can use the same field Reflectometer scans to run a more sophisticated online model that can pull in additional metadata from digital soil maps and other resources.
• Develop open source, regional models. By sharing training datasets across organizations and countries, we can build regional models that empower groups to develop their own locally appropriate decision support tools tailored for many different stakeholders. These shared training datasets and open models may lower the barrier to expanding the Reflectometer into new regions by lowering the number of training samples needed for the model to be effective in that region. If you are interested in using the Reflectometer and contributing to regional datasets please contact me at dan@our-sci.net.
• Improve the user experience. We are currently investigating alternative methods of communication between the Reflectometer and android phone to simplify use of the Reflectometer and enhance the user experience.

If you are interested in learning more about the Reflectometer or about these initiatives, please check out our Reflectometer tutorials or email me at dan@our-sci.net.