Progress 06/01/24 to 05/31/25
Outputs
Target Audience:Our work primarily served four target audiences: academics and students (including postdocs), insurers, financial institutions, and policymakers. We engaged Ph.D. students and postdocs in statistics and agricultural economics through one-on-one mentoring that introduced them to agroecology, micro-economics for farmers, high-performance computing, and agricultural commodities trading. This audience was targeted to build interdisciplinary expertise and train the next generation of researchers who can bridge statistical methods and economics with agricultural applications. We engaged academic researchers in agriculture, agroecology, agricultural economics, conservation, and statistics more broadly by sharing research findings. We also expanded engagement with agricultural insurers in both the private sector and with policymakers who oversee the federal crop insurance program. Insurers were targeted because of their growing interest in making actuarially sound assessments of the risk-reducing benefits of conservation practices, and our work matters to them as they seek to fill gaps in federal coverage, overcome barriers, and innovate new products such as private captive risk pools. In addition, we continued some dialogue with private-sector lenders, who are increasingly expected to support growers in building soil health and diversifying operations; this audience was targeted because of their role in financing agricultural transitions that align with risk reduction and resilience goals. Finally, we focused outreach on Congressional offices in our nine-state Midwest study region and members of the House and Senate Agriculture Committees. Policymakers were targeted because of their role in shaping crop insurance policy and related programs, and our work provides them with science-based insights into the benefits of making actuarially sound assessments of conservation practices that reduce yield loss risk. Changes/Problems:During this reporting period, we encountered several challenges that led to delays in aspects of our research. Hiring delays, due to difficulty in identifying a suitable candidate, slowed progress on the economic analysis. In addition, the scale of our dataset and the computational demands of our Bayesian yield models required significant time on high-performance computing clusters. Even with well-tailored and efficient data and model processing protocols--developed by our highly skilled Ph.D. students in statistics--the fragility of these clusters posed persistent obstacles. Despite these challenges, we successfully ran four yield models across five states for corn by the end of the reporting period. Despite these challenges, the Bayesian computational methodology is desirable because it produces detailed posterior predictions, which allow us to communicate model uncertainty to stakeholders in terms of probabilities of occurrence--a format that is intuitive and directly relevant to our target audiences. What opportunities for training and professional development has the project provided?We engaged Ph.D. students and postdocs in statistics and agricultural economics through one-on-one mentoring that introduced them to agroecology, micro-economics for farmers, high-performance computing, and agricultural commodities trading. This project has supported four different PhD students and two postdocs in gaining skills in applied statistics, remote sensing and data science, micro-economics for farmers, high-performance computing, and agricultural commodities trading. More specifically, they have received both 1:1 mentorship from PIs in areas like 1) advancing their applied statistical methods in cutting-edge ways (e.g. causal inference, spatial statistics of large datasets); 2) knowledge of diversified agricultural systems and their interactions with climate; 3) how commodities markets and climate impacts on yields affect farmers' gross revenues; 4) methods for mapping conservation practices like cover crops. They have also received training and experience in communicating their work with a transdisciplinary group that includes non-profits, policymakers and lenders. They have presented their research at academic conferences. How have the results been disseminated to communities of interest?During this reporting period, we disseminated results to communities of interest through publications, academic presentations, outreach meetings, and targeted engagement with non-academic stakeholders. We also organized conference calls with financial institutions and agricultural insurers to share preliminary findings and data resources, reaching stakeholders who are not usually exposed to academic research but whose decisions directly shape agricultural risk management. Project personnel participated in academic conferences and high-profile events to broaden visibility and impact. PI Tim Bowles gave an invited presentation at the Ecological Society of America conference, and collaborators with Land Core gave a presentation at the Tri-Societies conference. Co-PI Frederi Viens served as a panelist for the National Academies of Science, Engineering, and Medicine Frontiers of Statistics in Science and Engineering consensus study (virtual, April 8, 2025) and for the American Statistical Association Climate and Risk Roundtable on Risk Management for Business, Finance, and Insurance (Alexandria, VA, Nov 14-15, 2024). We also presented results at the Conference on Applied Statistics in Agriculture & Natural Resources, the Conference on Market Microstructure, and the International Conference on Probability and Stochastic Finance at ENSA, Cadi Ayyad University in Marrakech, Morocco. PI Bowles and Co-PI McLauchlan led our first virtual informational session and tool demonstration for Congressional offices, showcasing the risk model research, preliminary findings, and policy-relevant takeaways. Co-PI McLauchlan was also a panelist on a Foundation for Food and Agriculture Research webinar, A Decade of Soil Health Research at FFAR, where she shared insights from our interdisciplinary team. In addition, we attended and presented at the 2024 Regenerative Food Systems Investment Forum, engaging directly with financiers and practitioners in sustainable agriculture. Beyond formal presentations, Land Core worked to further socialize the project through its monthly newsletter, which reaches conservation and agriculture professionals across the country. Collectively, these activities allowed us to reach both traditional academic audiences and professional communities such as insurers, lenders, and policymakers, while also engaging practitioners and broader public audiences. By sharing results in scientific, financial, policy, and practitioner settings, our efforts enhanced understanding of agricultural risk and resilience and generated new interest in careers and learning at the intersection of science, technology, finance, and agriculture. We gave the following oral presentations during this project period: Bowles, T.M. Diversifying cropping systems to mitigate climate risk. Ecological Society of America Conference, August 2024, Long Beach, CA. Owen, R. The "Good Soil Discount": Creating the economic incentive for soil health practice adoption [Paper presentation]. ASA-CSSA-SSSA Annual Conference, Alliance of Crop, Soil, and Environmental Science Societies (ACSESS), November 21, 2024, San Antonio, TX. Pizzo, G., Viens, F. Big data and market microstructure in farm operations: a Bayesian study of the predictive distribution of corn revenue in the US Midwest. Market Microstructure, Quantitative Trading, High Frequency, and Large Data, May 15-17, 2025, Chicago, IL. Viens, F. Bayesian uncertainty quantification in agro-ecology, paleoclimatology and other applications. International Conference on Probability and Stochastic Finance. National School of Applied Sciences (ENSA), Cadi Ayyad University, Marrakech, Morocco, May 23 2025. What do you plan to do during the next reporting period to accomplish the goals?During the next reporting period, we will continue advancing our core objectives with a focus on expanding crop-specific modeling and farm-level economic analysis. Specifically, we will model crop-specific returns and net present value (NPV) for farms in the Midwest, providing clearer insights into the long-term economic impacts of conservation practices. We will also expand our work on soy yield modeling to address Objectives 2 through 4. We will use the results of the casual inference model comparison study to assess the drivers of spatial and temporal heterogeneity of crop rotation effects, using a novel meta-analytic approach. We will also use similar modeling approaches to understand the risk reducing implications of cover crop adoption using the nearly completed custom dataset on cover cropping.
Impacts
What was accomplished under these goals?
For Objective 1, we have nearly completed our goals. We now have a dataset consisting of approximately 10M field-level observations of corn yields between 2008-2022 across nine states in the U.S. Corn Belt, and a slightly lower number of observations of soybean yields. The yields are remotely-sensed and modeled pixel-level yields from a dataset recently released and described in Ma et al. (2024) using a method called "Quantile loss Domain Adversarial Neural Networks" (QDANN). QDANN-estimated yields achieve strong agreement with ground-based yield measures, with accuracy metrics higher than benchmark approaches. We have aggregated pixel-level yields to the field-scale using remotely-sensed field boundaries. Coupled with these yield observations are comprehensive data on monthly climate conditions (e.g. temperature, precipitation, vapor pressure deficit, growing degree days, etc) and soil characteristics. We have also compiled data on the 6-year cash crop history for each field using data from the Cropland Data Layer, and applied various metrics of rotational complexity. This dataset is foundational for the other objectives. Importantly, we have also made substantial progress producing a new dataset on the presence and absence of winter cover crops in the same region as above from 2013-2023. This is a major accomplishment because we had previously encountered delays due to a) companies unwilling to share cover crop datasets and b) challenges obtaining sufficient ground truth data on cover cropping to train an algorithm for cover crop detection. We combined three different machine learning models into an ensemble to improve predictive performance and accuracy metrics. For Objective 2, we have taken two approaches. First, we have compared several statistical methods for causal inference, which accounts for the fact that our data are observational and decisions to rotate could be confounded with other factors (like land quality) that also affect yield. Using the dataset assembled in Objective 1, and a simplified metric of rotation (prior crop history as corn vs. soybean) as treatments, we compared four causal inference approaches--structural equation models with spatially varying coefficients, matching, inverse probability weighting, and causal forests. Average treatment effects (ATEs) of prior soybean versus corn rotations were modest at the region-wide scale (around 1-3% of yield), but local effects showed clear heterogeneity, particularly during stressful years such as 2012. Spatial modeling improved the precision of local effect estimates, while structural equation models proved especially efficient and interpretable for handling large datasets. Overall, these analyses demonstrate that even simple rotations can provide yield benefits, though the magnitude and spatial distribution of those benefits depend strongly on weather, soil, and local context. All four methods produced highly concordant results, giving confidence in their robustness and establishing a strong methodological foundation for next steps. We have also expanded prior statistical modeling based on Bayesian predictive analysis, which excels at uncertainty quantification (though with tradeoffs for computational time). Expanding from one state in last year's project report to five (IL, MN, OH, WI, and SD), our results revealed that diversification effects vary substantially across space and time. Models were calibrated using thirteen years of data, which provided robust representation of common weather and management conditions but limited the ability to extrapolate to rare or extreme scenarios. The methodology allowed us to quantify both mean yield effects and downside risk (i.e. the probability of falling below a yield threshold based on historical yields in a given field) across different levels of rotational complexity, demonstrating that cash crop rotation diversification generally reduces downside risks compared to continuous corn. These effects were particularly consistent across the study region for moving from one to two crops in rotation, with lesser but generally positive or neutral effects from two to three crops, and stronger effects from two to four crops. Downside risk reduction was generally strongest in IL, which coincided with where opportunities for yields higher than thresholds also increased with diversification. However, depending on which climate variables were included, downside risk could sometimes increase with diversification, especially in WI and SD, showing the importance of accounting for spatial variation in both drivers and outcomes. For Objective 3, we examined which rotation patterns lead to more stable corn yields. We modeled both the first and second moments of corn yield distribution across millions of field-years, allowing us to separate effects on average yield from effects on yield variance. Importantly, we only focused on rotations including corn and/or soybeans. A key finding is that only a few sequences outperform corn monoculture in terms of higher mean/lower variance. Higher performing sequences have common elements including a recent soy rotation, no consecutive soy years, and a low impact of crops prior to 1-3 years before the focal year. Perfect corn/soy rotation has positive but nonsignificant effects on yields, and reduces yield variance. We also built on and expanded Bayesian modeling results from Objective 2, and tested how these findings could be applied in actuarial contexts. By comparing model-estimated risk distributions to government-set crop insurance premiums, we identified mismatches caused by current reliance on lagged county-level yield data, and how risk-reducing benefits of diverse rotations are not reflected in current rating practices. This work provides a technical foundation for incorporating rotation metrics into actuarial models, offering a more precise and probabilistic approach to aligning crop insurance premiums with on-farm risk reduction. We also began examining future risk (Pottinger et al 2025, included as a product). Specifically, we predicted the impacts of climate-driven crop loss at a policy-salient "risk unit" scale. Built with a neural network Monte Carlo method, simulations predict both more frequent and more severe losses that would result in a costly doubling of the annual probability of maize Yield Protection insurance claims at mid-century. We fill an important gap in current understanding for climate adaptation by bridging existing historic yield estimation and climate projection to predict crop loss metrics at policy-relevant granularity. For Objective 4, we quantified farm-level economic impacts of cropping system diversification by modeling the uncertainty in Midwestern corn farmers' annual revenues. Using our Bayesian predictive yield framework, calibrated at the county level, and incorporating price uncertainty via a geometric Brownian motion process based on corn price data from grain elevators (collected through Bloomberg), we generated distributions of farm revenue for every field across 90 Illinois counties. Monte Carlo simulations allowed us to combine yield and price variability, and in a preliminary analysis under an independence assumption, we found that yield variability accounted for a little more than one-third of total revenue uncertainty. When explicitly incorporating the negative correlation between price and yield--estimated from county-level data and consistent with national patterns--we found that yield variability represents closer to one-quarter of total variability, with the remainder driven by price fluctuations. These findings advance our understanding of how diversification and climate variability influence farm-level financial outcomes, directly addressing Objective 4's goal of quantifying net returns and variance of net returns under current and future conditions.
Publications
- Type:
Peer Reviewed Journal Articles
Status:
Published
Year Published:
2025
Citation:
Pottinger, S., Connor, L., Guzder-Williams, B., Weltman-Fahs, M., Gondek, N., Bowles, T.M. 2025. Climate-driven doubling of U.S. maize loss probability: Interactive simulation with neural network Monte Carlo. Journal of Data Science, Statistics, and Visualisation
- Type:
Peer Reviewed Journal Articles
Status:
Other
Year Published:
2024
Citation:
Manski, S., Socolar, Y., Goldstein, B., Pizzo, G., Ahmed, Z., Connor, L., Cross, H., Fettes, K., McLauchlan, A., Pham, L., Viens, F., & Bowles, T. M. (2024). Diversified crop rotations mitigate agricultural losses from dry weather. Preprint. agriRxiv. https://doi.org/10.31220/agriRxiv.2024.00244
|
Progress 06/01/23 to 05/31/24
Outputs
Target Audience:The key target audiences during this reporting period included undergraduate and graduate students and postdocs, academic researchers, as well as agricultural lenders and U.S. federal policymakers. Engagement in applied statistical research provided experiential learning opportunities for undergraduate and graduate students and postdocs. These trainees are a key target audience because it is important they gain domain knowledge of agriculture and diversified cropping systems as well as experience communicating their findings to knowledgeable laypeople. Academic researchers were another key audiences via academic audiences including the Conference on Applied Statistics in Agriculture and Natural Resources (attended by students, professors, and contractors working in fields related to agriculture and natural resources) and the American Geophysical Union Conference (attended by a wide array of academic disciplines as well as representatives from large agriculture / natural resources NGOs). Academic professionals are a key audience in order to receive feedback on our research methods and spark potential for collaboration. During this period, we also focused on engaging agricultural lenders, primarily in the private sector. Since this research can play an important role in helping the institutions who are in the business of assessing agricultural risk understand the risk-mitigating impacts of diversified cropping systems, early engagement is important so they might consider how to develop pricing or incentives that reflect the impact on production risk and on-farm profitability. In turn, the lenders we engaged also provided their expertise on loan pricing, underwriting, appraisal, crop insurance, and legal requirements to help inform the development of our beta risk model tool and the underlying research. To a lesser degree during this period, we also engaged agricultural insurers, both private and public/federal, to share our project's approach, and understand their current considerations, constraints, availability of data, etc. when assessing yield risk in agriculture, developing actuarial tables and writing or updating policies of insurance. ?We have also engaged in educational outreach with Congressional offices, particularly those within the nine state region of the US Midwest where the Risk Model focuses, as well as with members on the House and Senate Agriculture Committees, and USDA Risk Management Agency staff. Our outreach to policymakers allows us to share our research approach, partnerships, and preliminary findings on our risk-reduction potential of diverse crop rotations, and discuss potential key takeaways for federal policymakers. It also allows us to better understand informational needs for developing new FCIP policies that reflect risk profiles of diversified cropping systems within current statutory limits. Changes/Problems:A delay that occurred early in the award period came from difficulty in acquiring ground truth data for developing a dataset on the presence and absence of cover crops. We had to pivot to creating our own cover-cropping data due to a challenge with the partner we had identified in the grant proposal. While we have largely resolved the technical challenges, especially through a partnership with the Eric and Wendy Schmidt Center for Data Science for the Environment at UC Berkeley, we are still looking for ways to obtain more ground truth data that is required to develop detection models for cover crops from remotely sensed data. The change that may result from this challenge is narrowing the geographic scope at which we are able to assess the risk mitigating effects of cover crops, from nine states to possibly three or four. What opportunities for training and professional development has the project provided?This project has supported multiple PhD students and a postdoc in gaining skills in applied statistics in agricultural systems. They have received both 1:1 mentorship from all PIs from both angles of 1) advancing their applied statistical methods in cutting-edge ways (e.g. causal inference, spatial statistics of large datasets) and 2) knowledge of diversified agricultural systems and their interactions with climate. They have also received training and experience in communicating their work with a transdisciplinary group that includes non-profits, policymakers and lenders. They have also presented their research at academic conferences. How have the results been disseminated to communities of interest?Our team has given multiple conference presentations, including at the Conference on Applied Statistics in Agriculture and Natural Resources and American Geophysical Union. Co-PI Aria McLauchlan and other members of her non-profit, Land Core, have led efforts to communicate early findings with agricultural lenders and insurers, as well as members of Congress and their staffers. For example, we created a tool that allows for examining outputs from our statistical model of yields, crop rotation and climate in IL, showing patterns of risk mitigation and its magnitude. We provided in-depth demos of the tool's functionality to our flagship banking partner, Compeer Financial, and to pioneering lender, Fractal Ag, and we presented the tool and preliminary findings to audiences of agricultural lenders, investors, insurers & policymakers. These demos and conversations have been instrumental in providing user feedback and allow us to refine aspects of the tool. Co-PI McLauchlan also met with over 20 Congressional offices in October 2023 to share the project research, preliminary findings and the beta tool. These offices expressed great interest in the tool, and were curious to learn more about how the tool could be used to scale adoption of conservation practices in their districts and inform crop insurance. PI Bowles met multiple times with staff from the Risk Management Agency to share results and obtain feedback on the research. What do you plan to do during the next reporting period to accomplish the goals?Our main effort in the next reporting period will be to expand the work assessing how diversifed crop rotations and cover cropping affect maize and soybean yields across the full nine state region. We are first going to complete a comparison of causal inference methods to see which performs best with the less computational intensity. This is an important step to dealing with the confounding factors present in our observational data. We will also continue mapping cover crops, using remotely sensed data based on machine learning methods. This will include improving the accuracy of our current map (which only covers Indiana due to the presence of a lot of ground truth data) and expanding the geographic scope of the maps (to the best of our ability as more ground truth data is obtained). We will also continue the initial work on farm revenue prediction, including a full analysis of the correlation structure between various types of corn price (country elevator prices, grain terminal prices, national-level futures contracts), and how field yield predictions are correlated on a given farm. We will also begin to assess the uncertainty due to the prices of inputs and other farm operation costs such as grain storage, and how farmers can optimize their cropping and marketing decisions based on all these uncertainty assessments.
Impacts
What was accomplished under these goals?
We have made excellent progress toward several objectives. For objective 1, we have assembled a large geospatial dataset of high resolution (30m for most datasets) on crop rotational complexity, maize and soybean yields, soil characteristics, and climate information, which we have aggregated to the field level across five of our nine focus states in the Corn Belt. The remaining four states are currently in progress. One major effort in parallel was creating a field-level dataset on the presence and absence of cover cropping. A geospatial data scientist, Kangogo Sogomo, working with co-PI Land Core, developed custom pipelines to rapidly process Harmonized Landsat and Sentinel imagery to 1) generate cloud-free gap filled daily annual data at 30m resolution; 2) identify "green up" dates for cash and putative cover crops; and 3) efficiently extract candidate features from these NDVI time series that could be used to distinguish cover crops leveraging UC Berkeley's High Power Computing resources. This approach is based on a novel strategy combining methods from Zhou et al (2022) and Gao et al (2020) which distinguishes presence or absence of cover in fields. In tandem, we have collected >36,000 field points ground-truth cover crop observations from 2014 to 2019 for model training and validation, though greater geographic scope is ideally needed. The next steps are to determine the dynamic thresholds that define cover cropped and non cover cropped fields and develop a model that can accurately predict fields in the entire state of Indiana and then the additional 8 midwestern project states. Regarding Objective 2, we have also nearly completed a study of how crop rotational diversity affects corn yields during stressful weather in two states, IL and MN. Specifically, we used Bayesian modeling to show spatial patterns of yield benefits that result from increasing rotational diversity in a range of weather conditions. Data from over 2.2 million field-years revealed that diverse rotations decreased the risk of corn yield losses in dry years in areas that experience these conditions more frequently, while simultaneously increasing yields under favorable conditions across the region. This resilience effect was particularly clear across Illinois, where increasing rotational complexity typically does not negatively impact cash crop yields, and helps limit losses in drier years. There is also some indication that, in certain counties in Illinois, these more complex rotations can increase the chance of a bumper crop in good years (years with average rainfall). The potential for yield risk mitigation underscores the critical need for crop rotation adoption as changing climates threaten yield stability in the US and across the globe. We highlight areas where diverse rotations are most likely to improve yields and mitigate risk, amidst spatial heterogeneity in the magnitude of these benefits. For Objective 3, we have completed and submitted a paper examining future risk. In this first study for this project, we were interested in understanding how climate change might affect crop insurance indemnity claims for corn, and in subsequent studies we will focus on the role of diversified systems in mitigating these claims. This research area recognizes that climate change not only threatens agricultural producers but also strains financial institutions. We use an artificial neural network to predict future maize yields in the U.S. Corn Belt, finding alarming changes to institutional risk exposure within the Federal Crop Insurance Program. Specifically,our machine learning method anticipates more frequent and more severe yield losses that would result in the annual probability of Yield Protection (YP) claims to more than double at mid-century relative to simulations without continued climate change. Furthermore, our dual finding of relatively unchanged average yields paired with decreasing yield stability reveals targeted opportunities to adjust coverage formulas to include variability. This important structural shift may help regulators support grower adaptation to continued climate change by recognizing the value of risk-reducing strategies such as regenerative agriculture. Altogether, paired with open source interactive tools for deeper investigation, our risk profile simulations fill an actionable gap in current understanding, bridging granular historic yield estimation and climate-informed prediction of future insurer-relevant loss. Finally, we have begun research under Objective 4. Initial results focus on predicting uncertainty in farm revenue from cash crops. Our preliminary results show (1) cost uncertainty is a significantly larger share of variability than weather uncertainty or idiosyncratic (unexplained) variability at the field level; (2) revenue uncertainty is significantly smaller at farm level than field level; and (3) there is evidence of revenue risk reduction at the field level when going from 2 crops to 3 crops in rotation. This is work in progress since the analysis still needs to incorporate the fact that when crop yields are expected to go down, crop prices tend to start rising
Publications
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2023
Citation:
Cross, H. (2023). Quantifying and Reducing Production Risk in Agriculture: Key Takeaways from a New Soil Health Risk Model , Extension Risk Management Education National Conference (ERME), Chicago, IL.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2023
Citation:
McLauchlan, A. (2023). How Policy is Shaping Finance and Investment in Regeneration. Regenerative Food Systems Investment Forum, 2023, Denver, CO.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2023
Citation:
Cross, H. (2023) Rethinking Risk in Agricultural Portfolios: Adapting to Climate Volatility in Farmland Management and Underwriting. CREO Global Meeting, Montreal, QC, Canada.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2023
Citation:
Sogomo, K., Kang, Y., Bowles, T.M., Gao, F. (2023). Developing a Remote Sensed Cover Crop Dataset for the U.S. Midwest using Harmonized Landsat Sentinel (HLS) poster presentation, American Geophysical Union Conference, San Francisco, CA.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2023
Citation:
Bowles, T.M., Vendig, I., Socolar, Y., Manski, S., Cross, H., Fettes, K., Goldstein, B., McLauchlan, A., Pizzo, G., Viens, F. (2023) Diversifying cropping systems to build soil carbon and adapt to climate risks . Invited oral presentation, American Geophysical Union Conference, December, 2024, San Francisco, CA.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2024
Citation:
Gina Pizzo, Lawson Connor, Eunchun Park, Zirong Liu, and Frederi Viens. (2024) Farm Revenue Uncertainty: a Bayesian study for Corn in the US Midwest", oral presentation, Conference on Applied Statistics in Agriculture and Natural Resources, Iowa State University.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2024
Citation:
Tyler Bagwell, Gina Pizzo, Sam Manski, and Frederi Viens. (2024) Constructing interpretable statistical crop yield prediction models using field-level data, oral presentation , Conference on Applied Statistics in Agriculture and Natural Resources, Iowa State University.
|