Progress 09/01/23 to 08/31/24

Outputs
Target Audience:The Subterra Green addresses our customers' need to accurately, quickly, and cheaply measure soil organic carbon. We have found two potential markets for the Subterra Green: agricultural customers and environmental customers. Our initial target agricultural customers are large-scale farmers, cooperative representatives (sharing a device among small farmers), and soil measurement service providers. We have been in contact with dozens of farmers in order to secure access to their land for measurement and mapping. Our initial target environmental customers are soil researchers (private and governmental),companies providing carbon credit measurement services, and environmentalists and government regulators tasked with verifying carbon sequestration. We have several commercial environmental research companies that are intereested in evaluating our technology. Changes/Problems:Our goal was to have sampled 18 extensive sites and four intensive sites by this point in the grant performance period, but have in fact sampled 17 extensive sites and no intensive sites. Since progress in modeling depends on the data obtained by the sampling, progress towards these goals, and documentation using the target metrics, has also been delayed. In particular, the fact that we have not yet conducted sampling at an intensive site has prevented us from documenting model accuracy and precision for SOC stock (Objective #3). The delay in sampling is primarily due to problems with the availability of our prototype Subterra unit that occurred early in the performance period. We had contracted with a manufacturing partner to make two new prototype Subterra units in spring of 2023, prior to the beginning of the grant performance period in September 2023. We were promised that the units would be delivered to us by the beginning of July 2023. On the strength of that promise, our existing prototype was sold (as a Subterra Grey forensic model) in June 2023, leaving us without a working unit. However, the manufacturing partner was unable to satisfactorily complete our prototypes, so we retrieved them in an unfinished state in October 2023 and brought them for finishing to our original manufacturing partner. In addition, the manufacturer of the motor used to drive the probe ceased production of the motor. The newer Subterra prototypes therefore use a motor from a different manufacturer, which we needed to incorporate into our control application. We therefore did not have a functioning Subterra unit until after the beginning of 2024, which put us significantly behind the work plan described in the grant proposal. What opportunities for training and professional development has the project provided?The USDA SBIR Phase II project employs student and graduate students from the University of Akron to provide both field and lab work. All work done by students is done under close supervision and guidance of an S4 scientist. Field work teaches basic skills in laying out grids, using the Subterra Green to measure SOC, calibration techniques such taking core samples for data validation. Lab works includes the skills needed for soil analysis such as bulk density measurement. How have the results been disseminated to communities of interest?Dissemination of information is currently done by the S4 Mobile Laboratories web site (www.s4laboratories.com) and attendant social media (i.e., LinkedIn blogs). The majority of our discussions has been directly with researchers through academic channels, trade shows, etc. Since we are not actively selling the Subterra Green, these activities will increase in August 2025. What do you plan to do during the next reporting period to accomplish the goals?During the next reporting period, we will complete our field sampling campaigns. By the end of the next reporting period we will have sampled a total of 36 extensive sites (of which 17 have already been completed) and eight intensive sites. Using the data collected from these samples, we will improve our models for predicting SOC and BD using the Subterra Green spectral data. For the intensive sites, we will map SOC stock in three dimensions, and test our model's accuracy and precision for predicting SOC stock.

Impacts
What was accomplished under these goals? (1)To meet the requirements of extending generalizability of our SOC models while maintaining high accuracy and precision, our sampling choices are aimed at increasing both the quantity and the quality of the data available for training the machine learning models. Machine learning algorithms require large volumes of training data, but it is important that the training dataset encompass examples distributed throughout the range of target property variability. Soil spectroscopy requires a database that widely samples the soil variability within the study area. The relationships between spectra and soil properties can be both spatially dependent and highly non-linear, and it is difficult to construct a training set that adequately reflects the immense variation found in soils. We are therefore tracking the environmental factors that may influence spectral variation in the sites that we select, and are attempting to ensure that the training dataset fully encompasses the variability in our target population, defined as the soils of the cropland area of MLRA 111. In the two field campaigns of the project period, we have completed sampling at 17 extensive sites, distributed primarily in MLRA 111A . Two sites are located in MLRA 111B, and five in MLRA 139. We chose to sample in MLRA 139 in order to provide a coherent body of data from outside MLRA 111 that will act as a bridge between our newly-collected Phase II data and data previously collected during our Phase I activities, which were all from outside MLRA 111. We discuss our choice of site locations more fully under Objective #2, below. To date, we have not sampled at any intensive sites. Because the sampling activities at intensive sites are designed to test accuracy and precision of SOC stock estimation, we chose to delay these until we had developed good global models using data from extensive sites, which we have only recently had sufficient calibration data to develop (see discussion in section (4)). (2) Wetested a variety of soil and environmental datasets and different methods of combining them into a rubric that would be as comprehensive as possible and yet easy to visualize and understand, and would also be readily extensible into new target regions over time. Ultimately we settled on a primary rubric that allows for straightforward tracking of overall site type, while also retaining the option of using secondary attributes that can be visualized as necessary. (3)This objective depends on data collected from the intensive site types, which we have not yet sampled, as discussed above in sections (1) and (2). Challenges contributing to the slippage in progress toward the goal are discussed below. (4)The aim of the Phase II modeling activities is to develop models that are generalizable to novel data in a broad geographical area while maintaining high levels of accuracy. Towards that end we have engaged in a number of activities: We have collected new data for model training, as described above. We have monitored procedures and data throughout the data collection and modeling processes in order to maintain quality data and eliminate spurious data points. This has included obtaining new sampling equipment for collecting the soil cores, standardizing lab procedures for measuring BD and subsampling for OC measurement, and carefully examining spectra in order to eliminate noisy outlier spectra before they are added to our training dataset. We have expanded the set of land surface parameters (LSPs) calculated from ancillary data sources (primarily elevation grids) that can be included with the training data and have calculated these LSPs for a wider range of spatial scales than we did for our Phase I activities. In Phase I results, LSPs generally provided a slight performance boost, which we believe may be increased with a better match between spatial scale and soil processes. We have explored new machine learning model types. In particular, we are beginning to use artificial neural network (ANN) models with encouraging results, and when our dataset becomes large enough we will also try both 1D and 2D convolutional neural networks (CNN) model types. We have identified, obtained, and begun modeling with a large public soil spectral library that may also help to boost model performance. The chosen library is the Natural Resource Conservation Service's Rapid Carbon Assessment (RaCA) library (Wills et al. 2014; Staff and Loecke 2016; Wijewardane et al. 2016), which is appropriate for our purposes not only because it includes data from throughout the conterminous United States, but also because, unlike many other such libraries, it includes both surface and subsurface samples with information about the depth from which the samples were obtained. The main drawback of this library is that its samples were scanned in the laboratory in a dry, ground state, so its utility when combined with our moist, in-situ spectra is as yet unknown.

Publications