Knowledge Guided Encoder-Decoder Framework: Integrating Multiple Physical Models for Agricultural Ecosystem Modeling

Qi Cheng, Licheng Liu, Yao Zhang, Mu Hong, Shiyuan Luo, Zhenong Jin, Yiqun Xie, Xiaowei Jia · 2025

Read Knowledge Guided Encoder-Decoder Framework: Integrating Multiple Physical Models for Agricultural Ecosystem Modeling by Qi Cheng, Licheng Liu, Yao Zhang, Mu Hong, Shiyuan Luo, Zhenong Jin, Yiqun Xie, Xiaowei Jia (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06266.txt.

Introduction

Qi Cheng, Licheng Liu, and their co-authors present a knowledge guided model for farm ecosystems that blends physics based models with learning. Their 2025 paper notes that crop monitoring matters for food security and emissions policy. Physical models are built for narrow settings and carry uncertain parameters, while data driven models act as black boxes and need large training sets.

Their encoder decoder design draws process knowledge from several physical models at once, so it can predict key crop variables in new settings. A language model handles inputs that arrive in mixed and inconsistent forms, and also picks which physical model to lean on for each case. The method aims at one model that transfers across tasks and regions rather than many narrow ones.

Catalog record

CREATORQi Cheng, Licheng Liu, Yao Zhang, Mu Hong, Shiyuan Luo, Zhenong Jin, Yiqun Xie, Xiaowei Jia
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-06266

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Knowledge Guided Encoder-Decoder Framework: Integrating Multiple Physical Models for Agricultural Ecosystem Modeling Qi Cheng, Licheng Liu, Yao Zhang, Mu Hong, Shiyuan Luo, Zhenong Jin, Yiqun Xie, Xiaowei Jia Abstract Agricultural monitoring is critical for ensuring food security, maintaining sustainable farming practices, informing policies on mitigating food shortage, and managing greenhouse gas emissions. Traditional process-based physical models are often designed and implemented for specific situations, and their parameters could also be highly uncertain. In contrast, data-driven models often use black-box structures and does not explicitly model the inter-dependence between different ecological variables. As a result, they require extensive training data and lack generalizability to different tasks with data distribution shifts and inconsistent observed variables. To address the need for more universal models, we propose a knowledge-guided encoder-decoder model, which can predict key crop variables by leveraging knowledge of underlying processes from multiple physical models. The proposed method also integrates a language model to process complex and inconsistent inputs and also utilizes it to implement a model selection mechanism for selectively combining the knowledge from different physical models. Our evaluations on predicting carbon and nitrogen fluxes for multiple sites demonstrate the effectiveness and robustness of the proposed model under various scenarios. 1 Introduction Agricultural ecosystems are facing increasing pressures due to the growing population and the changing climate. Effective modeling of agricultural ecosystems can provide critical and timely insights into the growth of crops, soil health, water conditions, and impacts of environmental changes. These insights are important for ensuring food security, maintaining sustainable farming practices, and creating policies to mitigate natural disturbance-incurred food shortage (e.g., distribution of subsidies). Moreover, such models can help provide information needed for managing greenhouse gas (GHG) emissions. According to the statistics from the U.S. Department of Agriculture (USDA), around 10% of GHG emissions are related to agriculture management in the United States Sewadeh (2015) . Hence, the effective monitoring of agricultural ecosystems is indispensable for making decisions and policies for mitigating agricultural emissions and hence climate change. Conventional process-based physical models (PBMs) have been widely used to simulate GHG emissions from agricultural land. While grounded in detailed scientific principles, these physical models often involve complex parameterization Parton and others (2018) ; Zhou et al. (2021b) ; Markstrom (2012) . Such parameterization approximates the reality under certain conditions and observations and thus lacks flexibility in adapting to real-world scenarios with varying environmental conditions, management practices, and modeling scales. Moreover, multiple PBMs have been developed for distinct tasks using different datasets. As a result, different PBMs may show advantageous performance in modeling different processes in agricultural ecosystems. Machine learning (ML)-based data-driven models, on the other hand, are capable of learning patterns automatically from large data. Hence, they are increasingly considered as alternatives to PBMs for scientific modeling Reichstein et al. (2019) ; Xu and Valocchi (2015) ; O’Gorman and Dwyer (2018) ; Willard et al. (2022) . However, traditional ML models are not designed to capture complex physical and biochemical processes in scientific systems. Moreover, training these models usually requires large training data, which is often unavailable in real agricultural monitoring tasks due to the substantial human labor and material costs needed for data collection. Recent works have shown the potential of enhancing ML with scientific knowledge in agricultural modeling Liu et al. (2024a) ; Liu et al. (2022b) . Despite their promise, these methods require extensive expertise to design models for a specific task and are not easily adaptable to other tasks. To address these challenges, this paper proposes a knowledge-guided foundation model (KGFM) that leverages complementary strengths of PBMs and ML models. Such novel combination of physics and ML within the KGFM method provides threefold benefits. First, it can better capture complex physical processes by using a modular encoder structure that replicates different components in the agricultural ecosystems, including the carbon cycle, water cycle, energy cycle, and nitrogen cycle. This encoder is trained with sufficient simulated data produced by PBMs to capture dynamics specific to each component. It also enhances the interpretability by providing the simulation of flux variables related to intermediate processes. Second, it embraces multiple PBMs and introduces a model selection mechanism to intelligently identify PBMs that are most relevant to the downstream task. Hence, this significantly reduces the need for domain expertise for common users or stakeholders to manually select suitable models. The knowledge of the relevant PBMs is then leveraged to enhance ML model prediction. Finally, KGFM is able to simulate a range of physical variables related to the agricultural ecosystems through a decoder structure. In particular, it leverages language model-based structures to enable flexibility in handling partial observations so that KGFM can still perform model selection and get fine-tuned when we just observe certain variables (e.g., CO 2 or N 2 O) in a downstream task (e.g., for a specific monitoring location). We evaluate the proposed KGFM model using true observations of GPP, N 2 O, and CO 2 from different regions. The results demonstrate the superiority of KGFM in predicting different target variables compared to other baseline methods. We also validate the effectiveness of the model selection mechanism in identifying relevant PBMs. Our implementation has been released 1 1 1 https://anonymous.4open.science/r/KGFM-7264/ , and the dataset will also be made publicly available upon acceptance. This work represents a significant step toward more robust, generalizable, and interpretable modeling for scientific systems, paving the way for future research that integrates accumulated knowledge into ML for advancing scientific discovery. 2 Related Work The proposed method can be generally applied to modeling environmental ecosystems in many domains, such as climate science, hydrology, and geology. Traditional scientific modeling often relies on PBMs that are built based on a series of mathematical or physical equations to simulate the different processes. These models are necessarily approximations of reality, and their parameterization components make them limited in adapting to different scenarios, as highlighted by prior studies Osborne et al. (2010) ; Wang and Smith (2023) . Here we introduce recent works that use ML models as an alternative to modeling environmental ecosystems. ML methods for modeling environmental ecosystems. ML-based data-driven approaches have gained popularity due to their ability to handle large datasets and uncover complex patterns without requiring explicit process understanding. ML models, such as neural networks and random forests, have been applied to predict carbon fluxes and other ecological variables, as demonstrated by studies from Carbonneau et al. Carbonneau and others (2020) and Sun et al. Sun and others (2022) . More recently, advanced models, such as graph neural networks and transformer-based models, have shown further performance improvement in certain environmental applications Topp et al. (2023) ; Ye et al. (2021) . However, these methods are generally limited in their interpretability and generalizability, especially given limited and less representative training data Schneider and others (2017) ; Read et al. (2019) . Knowledge Guided ML Models. Knowledge guided ML (KGML) has emerged as a new hybrid paradigm to leverage complementary strengths of ML models and physical models Willard et al. (2022) ; Karpatne et al. (2024) . In particular, KGML aims to incorporate physical knowledge into different components of ML models, including training objective, model architecture, and learning algorithms. For example, previous works have shown that using the unsupervised mass

Read the complete text