Deep Learning Meets Process-Based Models: A Hybrid Approach to Agricultural Challenges

Yue Shi, Liangxiu Han*, Xin Zhang, Tam Sobeih, Amit Kumar Srivastava, Krishnagopal Halder, Frank Ewert, Thomas Gaiser, Nguyen Huu Thuy, Dominik Behrend · 2025

Read Deep Learning Meets Process-Based Models: A Hybrid Approach to Agricultural Challenges by Yue Shi, Liangxiu Han*, Xin Zhang, Tam Sobeih, Amit Kumar Srivastava, Krishnagopal Halder, Frank Ewert, Thomas Gaiser, Nguyen Huu Thuy, Dominik Behrend (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-2504-16141.txt.

Introduction

Yue Shi, Liangxiu Han, and a team of co-authors reviewed hybrid crop modeling in twenty twenty-five, joining process-based models with deep learning. Process models encode plant and soil physics, giving interpretable rigor but struggling to scale or adapt. Deep learning finds patterns in large data yet can be opaque and brittle where data are scarce.

The review sorts hybrids into two directions: networks that refine process models, and process knowledge that constrains networks. A case study on crop dry biomass compares hybrids against each parent approach across data quality, sample size, and location shifts. Hybrids won consistently, tolerating noise and generalizing to unseen places better. The authors close with challenges — interpretability, scale, data needs — and recommendations for agricultural modeling that weds domain knowledge to AI.

Catalog record

CREATORYue Shi, Liangxiu Han*, Xin Zhang, Tam Sobeih, Amit Kumar Srivastava, Krishnagopal Halder, Frank Ewert, Thomas Gaiser, Nguyen Huu Thuy, Dominik Behrend
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-16141

Browse the full library · About the Tower of Babel

Full text

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

Deep Learning Meets Process-Based Models: A Hybrid Approach to Agricultural Challenges Thanks: Corresponding Author : Liangxiu Han Yue Shi, Liangxiu Han*, Xin Zhang, Tam Sobeih, Amit Kumar Srivastava, Krishnagopal Halder, Frank Ewert, Thomas Gaiser, Nguyen Huu Thuy, Dominik Behrend Abstract Process-based models (PBMs) and deep learning (DL) are two key approaches in agricultural modelling, each offering distinct advantages and limitations. PBMs provide mechanistic insights based on physical and biological principles, ensuring interpretability and scientific rigour. However, they often struggle with scalability, parameterisation, and adaptation to heterogeneous environments. In contrast, DL models excel at capturing complex, nonlinear patterns from large datasets but may suffer from limited interpretability, high computational demands, and overfitting in data-scarce scenarios. This study presents a systematic review of PBMs, DL models, and hybrid PBM-DL frameworks, highlighting their applications in agricultural and environmental modelling. We classify hybrid PBM-DL approaches into DL-informed PBMs, where neural networks refine process-based models, and PBM-informed DL, where physical constraints guide deep learning predictions. Additionally, we conduct a case study on crop dry biomass prediction, comparing hybrid models against standalone PBMs and DL models under varying data quality, sample sizes, and spatial conditions. The results demonstrate that hybrid models consistently outperform traditional PBMs and DL models, offering greater robustness to noisy data and improved generalisation across unseen locations. Finally, we discuss key challenges, including model interpretability, scalability, and data requirements, alongside actionable recommendations for advancing hybrid modelling in agriculture. By integrating domain knowledge with AI-driven approaches, this study contributes to the development of scalable, interpretable, and reproducible agricultural models that support data-driven decision-making for sustainable agriculture. K eywords Hybrid PBM-DL Modeling \cdot Process-based Modeling \cdot Deep Learning \cdot Computational Sustainability \cdot agricultural Modeling 1 Introduction The world population was projected to reach 9 billion by 2020 [ 1 ] . Agricultural production must not only meet increasing food demands but also enhance resource use efficiency while minimising negative environmental impacts. These challenges are further exacerbated by the increasing frequency and severity of climate change effects. Therefore, adopting sustainable agricultural practices is both crucial and necessary to ensure long-term productivity and environmental resilience. Process-based models (PBMs) have long been essential tools [ 2 ] for studying and analyzing the interactions between agricultural production outputs and environmental variables [ 3 ] . Recent advancements in machine learning, deep learning, and artificial intelligence (AI) [ 4 ] have significantly enhanced research capabilities in agriculture. The integration of PBMs with these advanced technologies has opened new possibilities, enabling more accurate, scalable, and actionable analyses of agricultural and environmental dynamics. Such synergies are driving innovation toward more sustainable and resilient farming systems. PBMs encompass a wide range of disciplines with expanding boundaries and increasing societal relevance, particularly in agricultural analysis [ 5 , 6 , 7 ] . In agriculture, these models forecast crop productivity and evaluate environmental impacts [ 8 , 9 ] . They simulate dynamic system responses to temporal forcings influenced by static landscape attributes. Many PBMs share commonalities across various processes due to their focus on temporal dynamics, often represented by nonlinear equations, including ordinary differential equations (ODEs) and partial differential equations (PDEs). The understanding of vegetation and crop growth processes varies, with some being well understood and others relying on assumptions or empirical representations [ 10 , 11 ] . Originally, the spatial scale of process-based cropping system models PBMs used for agro-ecosystem analysis is either a single plant or a point in a field with homogenous soil and canopy conditions. In this case, PBMs use parameterizations to represent processes at much finer scales than the computational grid of climate [ 12 , 13 ] . However, these representations and parameterizations involve significant uncertainty due to complexity of the model. Within the broader frame of agricultural evaluations, these models are synthesized to deliver critical climatic forecasts and inform decision-making processes for resource custodians and policymakers [ 14 , 15 ] . In addition to PBMs, the rapid growth of Artificial Intelligence (AI), particularly within the realm of deep learning (DL), presents novel prospects for extracting insights from large datasets and bridging the knowledge gaps inherent in process-based models [ 16 , 17 ] . While numerous DL methodologies have been introduced, the intrinsic differentiable advantage of DL has not been fully acknowledged. Recent development with synthesizing elements from PBMs and DL-driven models (so called hybrid PBM-DL models) has brought about a revolutionary advancement in agricultural science. Fig 1 illustrates the search results of PBM, DL hybrid PBM-DL models in agriculture modeling on Web of Science platform. It shows that, since 2021, there has been a rapid development in the application of hybrid PBM-DL models in agriculture application. Figure 1: Number of publications related to the agriculture modeling of hybrid PBM-DL models, comparing purely data-driven DL models and process-based models in the domain of agricultural analysis. In this study, we present a systematic overview of modelling approaches in agri-ecosystem research, focusing on process-based models (PBMs), deep learning (DL), and hybrid PBM-DL frameworks. We evaluate their respective strengths and limitations, particularly in the context of agricultural and environmental applications. The paper is structured as follows: Section 2 provides an overview of existing PBM and DL methodologies in agricultural modeling. Section 3 reviews hybrid PBM-DL models and their applications in agricultural systems. Section 4 details an experimental evaluation of the hybrid PBM-DL model’s efficiency in agricultural science. Finally, Section presents the conclusions. 2 Process-Based and Deep Learning Models in Agricultural Systems PBMs and DL methodologies represent two important but distinct approaches for modeling agriculturals, each possessing inherent limitations (see Table 1 ). The comparative analysis of these two modeling paradigms reveals both congruities and divergences, suggesting that the benefits they offer can be complementary rather than mutually exclusive. Table 1: A Comparative Perspective: PBM vs. DL models Aspect Process-Based Models Deep Learning-Based Models Modeling Function Mechanistic & Rule-Based : Explicitly represent the biological, chemical, and physical processes involved in crop growth (e.g., photosynthesis, transpiration, nutrient uptake). Data-Driven : Utilize large volumes of data to learn complex, nonlinear relationships between inputs (such as weather, soil properties, and remote sensing imagery) and outputs (like yield). Process Representation : Use differential equations and empirical relationships derived from experimental and field research to simulate crop development over time. Pattern Recognition : Rely on neural network architectures that automatically capture interactions among diverse variables without embedding explicit rules. Scientific Foundation : Built upon established agronomic and physiological principles that can be related directly to plant behavior under different environmental conditions. Adaptability : Can integrate a wide range of heterogeneous data sources, making them suitable for rapidly changing or novel scenarios. Interpretability High Transparency : Each component of the model is linked to a known physiological or ecological process, making it easier for expert to interpret and validate model behavior. Black-Box Nature : The internal workings of neural networks are often not intuitive, making it difficult to understand exactly how input variables interact to produce the output. Scientific Insight : Changes in simulated processes (like reduced photosynthesis under drought) can be directly tied to observed plant behavior. Post-hoc Interpretability : Techniques such as SHAP values, saliency maps, or feature importance rankings are needed to derive explanations. Diagnostic Capability : Researchers can tweak individual

Read the complete text