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

TL;DR

KGFM tackles universal agroecosystem forecasting under data scarcity by fusing physics-based surrogates with data-driven learning in a knowledge-guided encoder–decoder. It introduces a modular encoder with $N$ PBM surrogates across carbon, nitrogen, water, and thermal cycles and a language-model–assisted decoder that handles inconsistent inputs and long-term temporal dynamics. A PBM-weighting mechanism automatically selects and weights surrogates for a downstream task, with training on synthetic PBM data and fine-tuning on real observations. Across GPP, CO$_2$, and N$_2$O, KGFM outperforms baselines, captures temporal changes more accurately, and provides interpretable intermediate fluxes and site-level generalization, offering a scalable tool for agricultural monitoring and policy support.

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.

Figures (5)

• Figure 1: (a) The overall encoder-docoder architecture of KGFM. The encoder extracts intermediate flux variables, which are combined with input drivers in the decoder to predict final outputs. (b) Adaptation of KGFM to the downstream task. Different PBM surrogates are weighted and the decoder is fine-tuned given true observations.
• Figure 2: Training Process of the model selection module.
• Figure 3: Model performance for predicting GPP and CO$_2$ across sites with observation data.
• Figure 4: Predictions of scaled N$_2$O values for a specific year across six different sites in the testing set of the observation data. Each subplot represents one site, with the x-axis indicating the day of the year and the y-axis showing the scaled N$_2$O values. From this plot, it is observed proposed KGFM model consistently outperforms the other models, particularly in accurately capturing the peak N$_2$O values, which are the most critical for N$_2$O predictions.
• Figure 5: Comparison between predicted and ground truth weights for the two encoder modules: carbon, and water. Each subplot presents the weight variation of the Ecosys for the specific module being tested, while the weights for the other three modules remain constant across the four combinations. The figure for experiments on nitrogen and thermal module are in Appendix