Research & Development

Societal Grand Challenge

Agriculture and forestry have unrealized potential to reduce Greenhouse Gas (GHG) emissions and become a larger GHG sink to slow the changing climate. While our forests and farms have significant untapped potential for net GHG (e.g., atmospheric CO2) reduction, realizing this potential requires addressing competition between land for biomass production and ecosystem services, scientific and technological challenges of actionable (e.g., cheaper and more accurate) sequestered carbon verification, computational integration of the incentive structures of land managers that implement agriculture and forest management practices, and policy innovations to encourage large-scale equitable adoption of climate-smart practices.

Research Mission and Objectives

AI-LEAF is transforming the science of AI and decision support tools for climate-smart practices in agriculture and forestry to co-create solutions for previously unsolvable problems, and accelerate adaptation to and mitigation of the changing climate, while informing policy and empowering carbon markets. We are growing a national and global community of AI researchers and practitioners, and become a trusted, integrated entity that delivers, partners, and collaborates on science-based solutions to address complex agriculture and forestry challenges. We are creating new sectors of the economy and revitalizing industries with significant economic impact and quality-of-life improvements to grow productivity. We are also re-envisioning U.S. agriculture for sustaining international competitiveness, with competitors (e.g., China) outspending the U.S. on both public and private agricultural and AI research and development.

Sensing is a powerful tool for monitoring spatial and temporal variability in crop and animal stress, but has limited ability to penetrate soil, clouds, or animal hides. Process-based (PB) crop and forest simulation models are powerful tools for understanding how biophysical processes affect temporal patterns in growth, but a lack of fine-grained pervasive CSAF data limits scaling up PB models from plots to fields to watersheds and larger scales. AI is a powerful tool for describing and predicting complex agricultural and forestry data relationships but is limited for modeling physical processes that may be evolving and interacting at multiple spatial and temporal scales. Addressing these challenges thus requires integrating the unique strengths of these technologies, creating vast synergies that drive innovation in CSAF and AI.

Research Themes

Expand all

Knowledge-Guided Machine Learning (KGML)

Lead: 

Goal: Develop advanced machine learning methods, specifically Knowledge-Guided Machine Learning (KGML), to address challenges in modeling complex ecosystem carbon-nutrient-water-thermal processes for greenhouse gas and facilitate many other downstream tasks, such as risk analyzing, management optimization and decision making. 

Building upon prior work, the research aims to provide:

  • Accurate predictions,
  • Generalize across different environments, and
  • Improve predictions by assimilating remote sensing data, thereby advancing the field of physical science applications.

 

Four main sub-working groups:

1. Building Benchmark Dataset for AI Algorithm Development (BBD Group)
The BBD Group is dedicated to creating standardized, high-quality benchmark datasets that integrates both observational (e.g. in-situ, remote sensing) and simulation data (e.g., DayCent, Ecosys, LDNDC, Climate Land Modeling) for various interested domain (e.g. carbon cycle, nitrogen cycle, greenhouse gas (GHG) emissions, drone images, satellite images, extremes, disease). By unifying data from various models and field measurements, the group aims to support the development and testing of AI algorithms—especially those within KGML. In addition to curating relevant agricultural and environmental datasets, the team will produce ready-to-use code pipelines to ensure seamless adoption by researchers within AI-LEAF and beyond.

Potential Impact:
Robust benchmark datasets will accelerate algorithmic innovation and model validation, ultimately leading to more reliable climate and agricultural predictions. It will foster collaboration across AI-LEAF groups and beyond, enabling reproducible research and enhancing the scalability of solutions in monitoring greenhouse gas emissions, predicting soil carbon changes, increasing climate resilience, and refining climate adaptation strategies.

2. Knowledge Guided Foundation Models
This project focuses on developing and applying large-scale foundation models tailored to agricultural and climate domains. By infusing domain-specific knowledge (e.g., agronomic, biophysical) into advanced deep learning architectures, the team seeks to improve the accuracy and interpretability of models used for land cover classification, soil parameter estimation, and emission projections. A key emphasis is on leveraging graphics processing unit (GPU) resources to train models that are both flexible and responsive to multi-modal data inputs (e.g., remote sensing, simulation outputs).

Potential Impact:
Knowledge-guided foundation models promise to enhance both the performance and trustworthiness of AI systems across AI-LEAF initiatives. These models can streamline data-driven decision-making, reduce computational costs, and expand the applicability of AI techniques to diverse real-world challenges, such as precision agriculture, carbon sequestration, and climate risk assessments.

3. Quantify and Reduce Uncertainty in Emissions Estimation (Q Group)
The Q Group targets integrating process-based modeling (e.g., DayCent, Ecosys, LDNDC) with AI-driven approaches to improve estimates of greenhouse gas (GHG) emissions and soil carbon dynamics. Their work centers on creating KGML “emulators” capable of matching or outperforming traditional models, while offering uncertainty quantification. Through shared data repositories and close collaboration with AI experts, the group iteratively refines these emulators to capture the complexity of soil, weather, and management interactions.

Potential Impact:
More accurate and transparent GHG and soil carbon estimates are critical for national inventories, decision support tools, and agricultural stakeholders. By reducing uncertainties, the Q Group’s work can guide policy development, support carbon credit programs, and help validate sustainable land management practices, ultimately contributing to global efforts in climate mitigation.

4. Verification of Regional Greenhouse Gas Emissions: A Multiscale Observational and KGML Modeling Approach
This team employs top-down atmospheric measurements (e.g., tall-tower data) and bottom-up modeling (e.g., land surface models, KGML-based simulations) to validate regional GHG flux estimates. By combining high-resolution observational footprints with state-of-the-art modeling approaches, they aim to disentangle anthropogenic and natural emission sources. The group plans to refine bottom-up inventories by introducing KGML methods, enhancing the reliability of emission estimates at local, regional, and interannual timescales.

Potential Impact:
Accurate verification of regional GHG emissions is essential for gauging progress toward climate targets and informing mitigation policies. By pinpointing high-uncertainty sources and improving model fidelity, this project advances scientific understanding of climate–anthropogenic interactions and facilitates evidence-based strategies to reduce greenhouse gas emissions across scales.

Climate Risks and Adaptation

The Climate Risks and Adaptation team examines how diverse climate-related stressors impact crop and forestry systems, as well as the effects of various adaptation efforts to minimize risks through shifts in system vulnerability and exposure. In crop systems, the team is building advanced machine learning models to better understand the impacts of hydroclimatic stressors on crop yields and examining adaptation-mitigation tradeoffs with irrigation. In forest systems, the team is building benchmark datasets, improving estimates of forest biomass, and examining risks from fires and other disturbances.

Combined Learning and AI Reasoning (CLeAR)

CLeAR integrates reasoning and learning for various applications, including estimating uncertainties in greenhouse gas flux, dynamic forest productivity modeling, and stochastic optimization in Earth-Economy scenarios.

(AI-MOD) Multi-Criteria Optimization Decision-Making for Climate Adaptation

The research focuses on AI-guided multi-objective optimization for CSAF decision-making. This involves optimizing mitigation practices and bio-productivity across scales with multiple criteria, including equity and GHG reduction. The challenges involve complex problems necessitating fast response times, handling large areas and addressing spatial fragmentation. The research outlines three tasks: large-scale stochastic optimization for sequential decision-making, developing methods to compute Pareto frontiers for trade-offs and designing spatial optimization techniques to ensure contiguous solutions. Researchers are drawn on prior work in complex network optimization and spatial dependency modeling, particularly in the Amazon basin, showcasing their expertise in addressing CSAF challenges.

Digital Twins

The research focuses on AI-aided Digital Twins (AIDT) to enhance resilience planning for climate scenarios in CSAF. Digital twins are necessary for evaluating CSAF and AI concepts under diverse climate scenarios, considering impacts on agriculture and forestry.  The digital twin is an in-silico representation of phenomena. One focus is on the creation of a digital for soils encompassing the contiguous United States. Soils play a critical role in agriculture settings including: (1) providing essential nutrients to plants for their growth and development, (2) Supporting water retention and drainage, and thus helping to regulate soil moisture levels, and (3) serving as a habitat for beneficial microorganisms that contribute to soil health.

Our work on digital twins involves modeling encompassing diverse meteorological and environmental phenomena to estimate critical soil properties such as moisture levels, salinity, and nutrient concentrations. By integrating diverse data sources alongside these models, we create a highly detailed and dynamic representation of real-world conditions. Users can interact with this digital twin through a visualization engine that enables the layering of multiple datasets, real-time animations, and the rendering of complex environmental processes, providing deeper insights and facilitating informed decision-making.

Expand all

Advances in GHG flux estimation and verification.

We hypothesize that AI-augmented CSAF tools can help lower cost and improve accuracy to operationalize the next generation of measurement, reporting, and verification (MRV) systems for agricultural and forestry GHG fluxes across multiple decision-relevant scales.

Climate risks, adaptation, and geographic shifts in cropland-forest transitions.

We hypothesize that AI-driven tools will accelerate our ability to model geographic shifts in cropland forest transitions and climate risks, and to assess the efficacy and feedback between adaptation actions and GHG fluxes, carbon sequestration, or productivity across a range of agriculture and forestry systems.

Multi-criteria optimization of mitigation practices and productivity.

We hypothesize that AI-driven advances in multi criteria optimization can be used to identify where and when to implement climate smart GHG mitigation practices, while maintaining or increasing biomass productivity and considering human decision-making priorities.

AI-guided advancements that emulate Earth-Economy macroeconomic ecosystem service payment markets.

We hypothesize (H4) that coupling advanced deep neural network (DNN) approaches (e.g., KGML and DRNet detailed in 3b3 themes 1 and 2) with Earth-Economy models will drastically reduce computation time and allow for more efficient consideration of how land use change affects ecosystem services and national and global economies.

Multi-scale multi-criteria GHG decision support tools.

We hypothesize (H5) that developing an AI enhanced, multi-scale, multi-criteria decision support tool will help collaboration and social learning among stakeholders to increase the chances of identifying consensus solutions by evaluating tradeoffs between and profitability of alternate CSAF practices in fields and parcels.