7

Breakout Sessions

After the fireside chat, the rest of the workshop’s second day was devoted to two sets of breakout sessions. For Breakout Session 1, the workshop’s panelists and participants divided into three groups, each led by a moderator, and focused on a particular aspect of dynamic soil information systems. Those three aspects were:

• Measurements, sampling, and archiving
• Collection and curation (including cyber infrastructure)
• Data analysis and models

Breakout Session 2 followed the same pattern and focused on the same three topics, but the panelists and participants (but not the moderator) were assigned to different topics so that each topic was investigated by two different sets of people. Designated rapporteurs took notes in each of the sessions.

On the workshop’s third day, the first session was devoted to summary reports from the breakout groups delivered by the moderators for each of the three topics. Charles Rice from Kansas State University synthesized the discussions on measurements, sampling, and archiving; Ranveer Chandra from Microsoft Azure Global summarized the session on collection and curation; and Bruno Basso from Michigan State University described the session on data analysis and models. These presentations were followed by a wide-ranging discussion among all of the workshop’s participants, which was moderated by Basso.

MEASUREMENTS, SAMPLING, AND ARCHIVING

Rice explained that what should be measured in soils depends on the questions that a project or group is trying to answer. Scientific researchers, policy makers, and growers will each require different measurements. In the case of scientific research, for example, measurements will be guided by hypotheses and developing theories, while the data can help to drive the development of new theories. In contrast, government agencies will have

different goals from researchers and from each other, and therefore will seek other types of data.

The types of measurements will differ not only by the methods and what is being measured, but also by temporal and spatial scales, Rice said. Spatially, for example, the measurements could be at the continental scale, regional scale, or, in the case of land management, down to hectares or acres. Similarly, the spatial scales of measurements will depend on the objectives. Studies of erosion may require data at the scale of decades, while studies of greenhouse gas flux may require measurements at the scale of minutes or even seconds because the types and amounts of gases emitted from the soil can vary dramatically over time. In the case of data collected for input into models, ideally there will be an interplay between the models and the data collection, with the models using existing data to explore various questions and then determining what types of data are needed to address unresolved issues.

In short, Rice said, the types and scales of soil measurements needed for a dynamic soil information system vary widely. The challenge, he said, is how best to integrate the different goals and objectives from all of the different projects and agencies interested in soil in order to create an optimal soil information system. Could such a system, for example, include data with temporal scales that range all the way from seconds to decades or spatial scales from microns to kilometers? What trade-offs will have to be made?

The breakout group discussions revealed several important soil characteristics to measure, Rice said. One was texture, which, surprisingly, many studies do not measure. Other physical characteristics were aggregate stability or some measure of soil structure, bulk density, water content, and color and aeration status. Rice observed that, unlike other physical characteristics that can be measured from archived soil samples, bulk density has to be measured on site. Color is another physical characteristic that indicates anaerobic or aerobic conditions in the soil, he noted.

The chemical measurements favored by the groups included soil organic carbon (SOC), pH, and various nutrients. They also discussed the value of knowing the carbon and nitrogen fractions of the organic matter.

Methods of measuring various biological properties are evolving, but there should be some measure of genomic diversity, which is becoming more widely available and less expensive, Rice said. Phospholipid analysis would also be valuable. Some participants suggested a need for simpler, more established techniques. Chloroform fumigation extraction, for example, is convenient, can measure characteristics such as microbial biomass carbon, and fits into many carbon models. Rice added that enzymes are another important biological characteristic to capture. Some measure of plant pathogens as a part of biological measurements was also suggested by participants in the breakout discussion.

The depth of samples, Rice said, also depends on the questions being addressed, but a measure of at least 1 meter of soil profile, if not 2 or 3 meters, would be valuable but perhaps not practical in every collection system.

Metadata are important in collections, Rice said, and the lack of metadata may be the largest deficiency in soil datasets. Metadata on location (i.e., where a sample was taken), weather and climate, and land use history would be valuable. Land use history would include vegetation type and, in the case of agricultural systems, tillage, fertility, yield, and crop rotations. Rice added that information about root distribution depth would be useful. For urban environments, land use history is also important, particularly when examining soil contamination. Was the land used for housing, commercial space, industrial plants, or something else? Participants also raised the need for information about dust and particulates (particularly for areas affected by desertification, wildfire, and wind erosion) and soil salinity.

Aerial remote sensing data would be valuable, providing information on elevation, crop productivity, tillage, and types of vegetation or crops. The breakout groups discussed below-ground sensing, Rice said, because new sensors are being developed that are placed in the soil and can measure dynamic properties at the scale of seconds or minutes. Such new technologies will provide novel types of data, he said, but they will also add further to a large and growing collection. “We’re going to have a lot of data,” he said. Rice suggested consideration of techniques for measuring bulk density and SOC via below-ground sensors to minimize the need to collect and store soils.

The breakout groups discussed the need for standardization, Rice said. For example, available phosphorus measurements can vary based on the type of soil chemistry test method used. Several participants emphasized the importance of quality control for any effort to harmonize datasets. During the previous day’s discussion on Slack, Travis Nauman of the U.S. Geological Survey noted that harmonization is a large task that needs to be supported with appropriate time and budget allocations, and Shawn Shalley of the U.S. Department of Agriculture’s (USDA’s) Natural Resources Conservation Service added that harmonized data require people, infrastructure, and standards.

Rice reported that archiving samples is important but poses several challenges. One is simply having enough space to archive all of the samples of interest. To keep all of the information in the samples intact, it would be advantageous to have a large number of freezers around the country that could maintain samples at –80°C, he said, but it would be less expensive to extract DNA and then freeze those samples at a less extreme temperature. The advantage with enzymes, he said, is that some enzymes can be assayed from dried samples. A participant suggested that the equivalent of the Svalbard Global Seed Vault in Norway was needed for soils. Rice agreed that ultimately it will take a large global project to store significant amounts of soil samples.

Finally, he said, a network of reference landscape sites would be valuable. Here, newly developed measurement methods could be calibrated on these reference sites against existing methods to ensure their comparability. A participant suggested taking advantage of the National Ecological Observatory Network, the Long-Term Ecological Research, and the Long-Term Agroecosystem Research sites, and perhaps also have each land grant university in the United States dedicate a few acres to a management practice or a land use practice to serve as a reference site. Participants on Slack noted that existing rangeland and forest monitoring networks could be used as well. The number and diversity of sites would enable the study of soil and climate variability throughout the United States in many landscapes.

COLLECTION AND CURATION

Following Rice’s summation of the measurements, sampling, and archiving breakout sessions, Chandra spoke about collection and curation. One of the first questions his breakout groups addressed, he said, is what is needed to make soil data follow the findable, accessible, interoperable, and reproducible (FAIR) standards.

The participants discussed various angles of this issue, he said, beginning with ways to incentivize sharing. Building accurate soil information systems requires data from many different types of stakeholders—academic scientists, government, the private sector, and, most importantly, farmers—and convincing the owners of these data to share them can be challenging. The participants identified ways to incentivize sharing, including scientific journals requiring authors to provide the data underlying published studies. Data could be required to be shared in a particular format, Chandra said, so that it is easy to reference and to reproduce. Another incentive could be new business models, for example, carbon

markets or changes to crop insurance. In the case of carbon markets, Chandra said, farmers might be more willing to discuss their farming practices if there is financial incentive to do so. Participants noted that younger generations seem more willing to share data; however, people who are more senior and closer to retirement may also be more willing to share their data to ensure their legacy. To convince industry to share data, Chandra said, it is important to be consistent in terms of, for example, when the methodologies used in a study are shared.

A second topic was how to best find the data for a dynamic soil information system, given the many different potential data sources. One approach might be to leverage the semantic web, using knowledge graphs to link similar data, so that having one source of data leads to many others. Chandra added that collection of metadata will also be important. If journals required data sharing, then a DOI citation could be created to link datasets with the publications as well.

Because a large amount of soil data are being generated and collected, people should be trained to use the data effectively, Chandra said. “How do we get not just soil scientists, but also the farmers, the public sector, [and] the government to use all of this data and make sense of a lot of this data?” Participants offered several approaches, including use of Wolfram’s natural language for non-coders to facilitate extraction of information from the data available in different data sources and use of data visualization. A challenge exists in finding the best way to present soil data to someone who wishes to use the data to make decisions or to answer scientific questions. How does one present the physical, chemical, and biological properties of soil in visualizations so that they make sense to people both inside and outside the field? Much work remains to be done here. Another approach to making the data accessible to all users is the application of natural language processing technologies. With the right software, for example, user questions, posed conversationally, could be converted into a query to a database. In short, Chandra said, several advances might help train people to use soil data.

In addition to the soil data, soil information systems also need metadata that provide information about how the data were collected and what they correspond to, Chandra said. But who should perform this data annotation? This effort becomes more complicated when the data are collected by farmers and citizen scientists. Crowdsourced data will likely become increasingly important, but who will annotate those data? Should scientists be paid to do it? Can annotation be automated, perhaps using artificial intelligence? Again, several challenges remain.

Next, Chandra summarized the groups’ discussion on data fidelity and data veracity. There is a trade-off between sharing all of the data that have been collected and sharing only the part of the dataset that has undergone strict cleaning procedures, such as calibration of the sites and the data. Different organizations have made different choices, he said. One advantage of making all of the data available is that users can apply their own filters to the entire dataset.

A related question pertains to the appropriate level of data to share, particularly in the context of privacy concerns. Farmers, for instance, might be concerned about revealing their farming practices if all of the data in a dataset are shared. One potential solution is to use recent advances in cryptography that make it possible to run analyses on encrypted data so that conclusions can be made based on the data without compromising the privacy of the people who contributed the data. These are new technologies, Chandra said, but they might be useful in developing soil information systems that contain sensitive data.

Chandra reported that the groups also discussed the pros and cons of various storage options. Data can be stored on local servers, the Cloud, or perhaps in one centralized location. Cloud storage offers several benefits, he said. Data uploaded to the Cloud can be shared

easily, and software applications and computing power facilitate data analysis in the Cloud. However, Chandra said, downloading large amounts of data from the Cloud can be expensive.

Another consideration is that some soil data could have national security implications, and therefore regulations may prevent sharing or storage of the data on a centralized server. Thus, at least some data will have to be shared via distributed datasets, which will require standardized metadata and a knowledge graph that displays the location of particular data and the relationships among datasets. Because of these considerations, Chandra said, “we believe the future of soil datasets will be distributed, especially across countries.”

Finally, Chandra summarized discussions concerning the use of artificial intelligence (AI) on soil data. At present, most of the AI-related work on soil datasets involves understanding the current state of soils; little work is being done to predict future soil conditions. AI-enabled forecasting would be an important capability, but various challenges to its use exist. A major challenge is data quality and fidelity, he said: “How do you make sure that the data are correct?” The use of AI in soil science is in an early stage compared to other areas of science, which presents the opportunity, Chandra said, “to make sure that we use AI the right way.” Doing this, he said, will require ensuring that the data are not biased, that any uncertainty is clearly communicated, that the AI methods are explainable—“Whatever the models predict, you should be able to say, ‘Why did it predict what it predicted?’”—and that the AI methods are applied ethically.

DATA ANALYSIS AND MODELS

The breakout group discussions on data analysis and models considered the promise of the current machine learning and AI methods for working with dynamic soil information systems. First, participants agreed that the promise depends on what is being modeled and that any modeling approach will be shaped by the questions being asked. Second, machine learning can help generate new hypotheses that can then be tested through experiments or with process-based modeling. Third, the field is moving toward multiple models, both machine learning models and process-based models, to create metamodels. This approach can be very valuable in characterizing uncertainties for different sources, Basso said. Fourth, it is important to triangulate from different models—machine learning, process-based models, remote sensing, and measured data—to better verify the answers generated by AI.

Machine learning has several potential applications, Basso said. It can be used, for example, to predict management practices based on detailed farmer surveys by converting qualitative information into quantitative information. It can search for missing data and fill in data gaps, and it can assist in determining where and when measurements should be taken, based on what is already known and what is missing from existing datasets. Optimization techniques can be augmented with machine learning and enable more efficient exploration of the parameter space.

Next, Basso addressed the concerns related to machine learning and AI. To begin, he said, they are black boxes—how they arrived at a particular decision can be unclear. Second, the methods have to be matched with the right questions in order to arrive at the best answers. Furthermore, machine learning cannot extrapolate beyond the training data that are supplied. “That’s certainly a big thing to consider,” he said. Some participants added that overfitting is a frequent problem with machine learning.

One important issue, Basso said, is how machine learning and AI could be used to derive insights. If machine learning is to guide decision making, the people involved need to have enough domain knowledge to use that guidance effectively. That need, in turn, points to the need for fully integrated multidisciplinary teams.

In addition, there are limited high-quality data for use with machine learning. A process-based model can help by generating data in order to fill gaps, he said, but that model must be validated. “We need to be sure that we understand the process and serve as an independent validation.”

Machine learning should not be used alone with sparse data, Basso said, because the results end up modeling the noise rather than the data. In such cases, the more promising approach is to combine machine learning with a process-based model or models. It is important to recognize that machine learning is just one more tool—it does not replace other tools.

Modelers need to be cognizant of information such as site history, past management practices, water table levels, and in-season heterogeneity, Basso said. Too often, these types of information are not considered in the models. However, integrating models with data poses various challenges, Basso said. His breakout groups considered various ways to improve such integration. Bayesian methods, for instance, offer a way to account for expectations based on prior experiences and to achieve better parameter estimation in models, characterizing uncertainties and past management practices. Process-based models will likely be better at predicting the future than machine learning and AI models because a systems approach is better able to capture the interaction between soil, plants, climate, management practices, and other factors. AI, by contrast, depends on the amount and type of data used for training, and the lack of a proper training set is often a limiting factor. During the previous day’s discussion on Slack, Michael Dietze of Boston University noted the untapped potential to leverage iterative state data assimilation approaches to continually pull data into models in almost real time. Such approaches would update forecasts on a rolling basis and would provide a way to reconcile different data types into biogeochemical reanalysis products (e.g., for carbon pools and fluxes).

When certain inputs for models are missing or incomplete, plants can sometimes be used as indicators of variability over space and time, Basso said. That is, plants can be used as proxies to improve the available input information. For example, thermal analysis of vegetation could serve as a proxy for soil depth.

Modeling soil organic carbon, nutrient, and water dynamics requires the proper simulation of crop yields, biomass, roots partitioning, uptake, water balance, and other factors, he said. The variability in crop history from yield maps and remote sensing is captured via yield stability maps and thermal stability maps.

Basso next spoke about lessons learned from the Agricultural Model Intercomparison and Improvement Project (AgMIP), a program that he has been involved with for 10 years. The soil community could learn a great deal from AgMIP’s experiences, he said, particulary about the use of ensembles of models, the medians of models, and the minimum number of models needed to capture model structures. The program compared the uncertainties when running 1 crop model with 20 climate models versus 1 climate model with 20 crop models, and found more uncertainty in the results of the crop models. It also examined the tradeoffs between the added complexity with additional models versus the added accuracy of the results that comes with the extra models; the ideal is to capture the system well enough while maintaining as much simplicity as possible.

The scales at which data are captured will vary according to the purpose, Basso said. Experimentalists, for instance, often make measurements at scales that differ from those used by modelers to make predictions.

Several new data types are becoming available that may inform understanding of soils. For example, some researchers are using qualitative trait loci for gene modeling, which in turn is embedded into crop simulation models. Spectroscopy is another measurement technique that is not yet fully used in soil science models but could add important details.

A dynamic soil information system needs to include site history metadata in order to help modelers, Basso said. Information about just the soil is not enough—it is important to know what happened at the site as well. Ideally, he said, modelers would use a nested hierarchy of models in order to capture details at different scales. It may be a challenge to combine a number of models in this way, however, because they will likely have different designs.

A dynamic soil information system should have data from repeated measurements in order to reduce uncertainties, he said. New types of data could be generated by underground sensors for nitrous oxide and volatile organic compounds. Basso said that he is excited by the potential of these new sensors to provide better data for models. In the past, he said, the general consensus in the field was that models fail to make accurate predictions because their inputs were of insufficient quality, rather than because researchers’ understanding of the system was poor. On a related note, he said, there is untapped potential in linking remote sensing with process-based models and using the remote sensing data to inform the model processes and inputs on the fly. Advances are being made in remote sensing, including those that improve resolution and those that offer new capabilities, such as new types of thermal imaging, and the resulting data will inform more accurate models.

American modelers could benefit from a survey such as the European Union’s Land Use and Coverage Area frame Survey (LUCAS), Basso said. LUCAS collects data on a large number of variables at high resolution, and its annual operating costs are $10 million, which is inexpensive when compared to the value of the data. Another important step, he said, would be to organize soil databases to facilitate their use by modelers. Modelers need a system that helps them synthesize information from several different sources, and the information system should be able to link multiple databases to make it easier for modelers to contextualize the data. Basso said that he works with global assessments, and he has found it difficult to cobble together data from different sources at different resolutions; he must create a custom data stream each time. Also valuable would be better soil ontologies and better communication of the uncertainties and measure of risks. For example, communication about the sample size is important because the smaller the sample, the larger the uncertainty.

Basso concluded by reiterating six key points from the breakout groups that discussed data analysis and models:

• Setting the scale and the objective of any modeling approach is critical for choosing the right type of model system.
• Machine learning presents great promise as well as concerns.
• Trade-offs may be made between model complexity and simplicity because it is important to find the right balance. However, many of today’s models are not balanced, with the level of detail varying greatly across mechanisms.
• The integration and fusion of domain knowledge is necessary for making sound assumptions; in particular, the integration of machine learning, remote sensing, observed data, and process-based models is key.
• Better soil ontologies and FAIR standards and file formats are needed.
• The resolution of data needs to be improved at both the spatial and temporal scales.

DISCUSSION

A lengthy discussion followed the reports from the breakout sessions. Phil Robertson of Michigan State University wondered whether it would be worth identifying minimum dataset expectations. Even though different stakeholders need different measurements, some

soil properties are likely to be of interest to all. He suggested that a tier-one set of variants that all database contributors could strive to provide could include, for example, taxonomic information, bulk density, texture, pH, and SOC. A participant on Slack suggested the addition of inorganic carbon to a tier-one list. Stephen Wood of The Nature Conservancy also noted on Slack that USDA, Cornell University, and the Soil Health Institute have already created tiered soil indicators. In his opinion, the lists have had limited effect because they have not identified how the indicators provide insight into land management practices. Basso added that the indicators identified are not necessarily scalable. He noted that, to gain insights, long-term observations are needed, including of the past management practices and of the initial condition of the soil. This information can be difficult to impossible to collect; however, models can help fill the gap, allowing measurements and modeling to come together, he said, to provide insights.

Stephen Ogle from Colorado State University emphasized points raised in Alison Hoyt’s presentation from the first day of the workshop: there are datasets in the published literature that need to be brought into a soil information system. He also noted that systematic sampling, as is occurring through LUCAS in the European Union, is important because it can be used to draw inferences across the continent. That kind of statistically-based sampling could be incredibly important for places in the world for which data are lacking. The field will benefit from thinking globally and engaging organizations such as the United Nations in promoting global assessment of soils.

Michael Young from The University of Texas at Austin reiterated the importance of ontology. When creating federated datasets, harmonizing the language among datasets is a challenge. The geoscience community has made some progress in this area, particularly work at the University of Colorado Boulder (Stoica and Peckman, 2019). The soil sciences could consider drawing on that framework, he said.

Mark Bradford from Yale University said that many of the properties that researchers might want to capture in a soil information system, such as SOC, have effects, for example, on water-holding capacity. From a policy maker’s or practitioner’s standpoint, what needs to be known is not how much SOC there is, but how that SOC amount translated into higher crop yields because of improved water-holding capacity. Much work remains to be done before those types of insights can be drawn from the data, he noted. Machine learning and AI do not provide those types of insights. Chandra agreed that more work is needed in the area of causal inference in machine learning, but that as it develops, there could be interest in applications for soil data.

Colin Averill of ETH Zürich commented on the importance of learning more about the biology of the systems. “In plant biology,” he said, “if you didn’t know whether you were looking at a tree or at moss, that’d be a big problem and you learn a lot by gaining that information.” LUCAS has invested a great deal in collecting biological information from soils, he said, and is now conducting genetic sequencing on samples from around 1,500 sites. Macro-scale data are needed so that spatially distributed data can be paired with functions, such as soil fractions and productivity. Industry, too, has recognized the value of this biological data and understanding and is investing extensively in microbial product development. “I think if we miss that, we’ll miss an opportunity,” he said.

Basso commented that biology is indeed critical and reported on his conversation with a biology colleague who pointed out that soils are involved in genes and species, ecosystem, climate landscape, and management interaction. Unfortunately, he said, the potential contributions of biology are often underestimated.

Basso then expressed his opinion that the breakout groups that discussed measurements, sampling, and archiving missed an important issue. Regarding standardization among labs,

he said, the key is to agree on procedures. It is not about whether some labs are good and some are not so good—all of their measurements may very well be correct, but the procedures need to be the same across labs. Rice agreed and explained that his group focused its discussion on some of the methods and the ancillary data. Many of the lab procedures are standard, he said, but some of them do vary; therefore, it is important to determine whether the different procedures are comparable, and if so, how to make comparisons between them.

Another issue that the breakout group seemed to miss, Basso said, was the possibility of capturing heterogeneity. Rice answered that the groups did not even approach that issue. “Maybe nobody wanted to tackle that,” he said.

Michelle Wander highlighted a 2019 editorial by Baveye et al. (2019) that she believes would help to direct investment in microbial community research and maybe even the way in which information on biological response is interpreted. With regard to marrying theory with statistical fitting, the field needs to be cognizant of using data from soil cores that average lots of microbial environments. To really understand nutrient use efficiency or the co-location of the microbes, the enzymes, and the roots for breeding, the field must acknowledge, she said, that marrying information will not be a good use of time and energy. Digital soil approaches have been criticized because of these scaling issues. This point was made in the early soil quality work, she said, and it is important to remember that the field is still trying to sort out what the best investment will be.

Basso passed along a comment from Slack that next-generation biology must be linked to micro-scale responses to be relevant to policy makers and others. Averill agreed, adding that the biology must be connected with crop productivity or something similar. “How many datasets do you know where those observations are paired? I know very few,” he said, and yet it is datasets with linked biological and crop data that are most likely to lead to breakthroughs. Basso responded that several private-sector companies are trying to relate genomic data to yields. However, relating the genomic data of the microbes with the yield may be too great a leap, he said, so understanding of some intervening variables may be needed. Averill agreed and said that the right approach may be to link the metagenome profile of the fungi or other relevant microbes and translate that linked profile into features that can be input into a simulation model or even a statistical model of productivity. Ultimately, though, success will depend on having datasets with all of the necessary data.

Mark Bradford asked how the field should interpret microbial indicators and enzymes. For example, in agricultural soils, an increase in enzymes is perceived as good because that indicates increased microbial activity. However, in other soils, an increase in enzymes is an indication of nutrient limitation. This issue of interpretation relates to his earlier point on the broad question around insights, information causation, and how they can be used. Mary Firestone from the University of California, Berkeley, responded, first by offering a thought on Averill’s comment that fungi are permanent and bacterial are ephemeral. She agreed to some extent, but the functionality of bacteria are not ephemeral, and the functionality is fundamental, so the next question is how to establish that functionality. She said that some people look to enzymes to establish functionality, but enzymes can be ephemeral as well. What, then, is the cause and the indicator? What is the cause of those enzymes? What can researchers hope will be semi-permanent and deterministic? That, she said, will be the genetic capacity, and to understand genetic capacity, metagenomics must be used. However, researchers are still learning how to piece this information together in order to understand how genes work to, for example, cause denitrification. The potential to use metagenomics as a tool for understanding fundamental soil characteristics is moving quickly, but there is still much to learn.

Vanessa Bailey from the Pacific Northwest National Laboratory offered a similar comment. After noting the importance of time-series data—“I think a lot of these data are use-

less as single points in time,” she said—she spoke of the value of developing models that combine genomic data and enzyme optimization conditions with the physical and chemical characteristics of soil and with site history information. “I think novel models that bridge biology and physics, biology and chemistry all together … would be a way to give us more robust information that’s useful for future predictions,” she said.

Wood commented that much of the conversation had focused on how best to understand changes in soil properties across space and time. This is a good goal, he said, but it is also important to talk about what can be done with these measures. “Generally it’s been our experience that even if you can measure something across space and time, we just don’t know how to interpret that and what to do with that,” he said. “So what does a change in a soil protein index mean, or a change in enzyme potential, [or] even microbial biomass?” Many of these properties have been measured for a long time, but no solid data exist that show how changes in those properties have translated to changes in agronomic outcomes or environmental outcomes. “So even if we go all in on a dynamic soil system that tells us fine-scale detail about how these things change,” he said, “it’s not going to be hugely practical and useful until we can say what those changes mean.”

Melissa Ho from the World Wildlife Fund built on Wood’s comments, saying that it is crucial to first understand the objectives of a study and then have those objectives determine what is being measured and the appropriate scale. The objectives will depend on the needs of clients and users, the cost-effectiveness of various approaches, and so on, and will likely differ by stakeholder (e.g., policy maker, market-driven actor). “So,” she said, “I would frame the conversation on the why and then talk about the appropriate metrics.”

Basso responded by asking how, given these points, a dynamic soil information system should be structured. Perhaps subsets of teams should be created to tackle different questions, Ho responded. “I think I would recommend we organize by what are the questions we’re trying to ask and then what sort of soil data can we collect.” Other considerations include cost efficiency and leveraging of citizen science and existing field sites with long-term datasets to collect at least the minimum data needed.

“I think also we should be thinking about the environmental challenges out there,” Ogle added, such as food security and climate change. Researchers should identify the knowledge gaps and the research questions requiring answers to address those issues with the help of dynamic soil information systems.

Kathryn Elmes from Indigo Ag agreed with these suggestions to start with the overarching objectives and then focus on the research questions that lead to achieving those objectives; however, once those questions have been identified, the field should issue as broad a call for research as possible. “These are the questions. Who has data already that might help us answer these questions? At what scale? And what are the gaps that exist?” she asked.

Julie Jastrow from the Argonne National Laboratory offered that a tractable approach to a dynamic soil information system that could deal with varied questions, interests, agency missions, and funding could be one that is hierarchical. A widespread georeferenced system with a set of basic characterization measurements would provide all users with greater context of their measurements. More targeted and comprehensive measurements on regional or even site-specific locations and addressing specific needs and questions could be contributed with specific geolocations overlaid on this system and built over time. Then, efforts for system-wide systematic measurements could be added later, she said.

Rodrigo Vargas from the University of Delaware noted that a wide variety of tools is needed to understand soils and that many different communities are interested in soils. The challenge is to bring together all of these data, concepts, and perspectives to increase knowledge of soils for specific applications.

Chandra noted that more breakthrough research is needed on new sensing methods to democratize the technology. Sensors remain expensive, and taking the kind of measurements that were discussed by the measurements, sampling, and archiving breakout group is nontrivial. The vision is the use of phones to measure some soil properties, but as was discussed by the collection and curation breakout group, challenges with data fidelity and data calibration remain. However, the increase in the amount of data collected helps to lower measurement costs.

This page intentionally left blank.