Exploring Physics-Informed Neural Networks for Crop Yield Loss Forecasting
Miro Miranda, Marcela Charfuelan, Andreas Dengel · 2025
Read Exploring Physics-Informed Neural Networks for Crop Yield Loss Forecasting by Miro Miranda, Marcela Charfuelan, Andreas Dengel (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-2501-00502.txt.
Introduction
Miro Miranda, Marcela Charfuelan, and Andreas Dengel forecast crop losses in a paper whose text is dated 2024. Extreme weather puts harvests and food supply at risk. Classic crop models follow plant physics but often miss the mark. Pure machine learning models score well yet hide their reasoning.
The team blends both paths in a physics informed neural network. Their model estimates water use and water stress at each map pixel. It then solves the standard yield response relation step by step. Inputs came from satellite images, climate records, and field yield data. Accuracy matched or beat strong rival models in their tests. Outputs stayed true to plant physics, so users can read the reasons. Farmers and officials could use such forecasts to plan ahead.
Catalog record
| CREATOR | Miro Miranda, Marcela Charfuelan, Andreas Dengel |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-00502
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exploring Physics-Informed Neural Networks for Crop Yield Loss Forecasting Miro Miranda, Marcela Charfuelan, Andreas Dengel Abstract In response to climate change, assessing crop productivity under extreme weather conditions is essential to enhance food security. Crop simulation models, which align with physical processes, offer explainability but often perform poorly. Conversely, machine learning (ML) models for crop modeling are powerful and scalable yet operate as black boxes and lack adherence to crop growth’s physical principles. To bridge this gap, we propose a novel method that combines the strengths of both approaches by estimating the water use and the crop sensitivity to water scarcity at the pixel level. This approach enables yield loss estimation grounded in physical principles by sequentially solving the equation for crop yield response to water scarcity, using an enhanced loss function. Leveraging Sentinel-2 satellite imagery, climate data, simulated water use data, and pixel-level yield data, our model demonstrates high accuracy, achieving an R 2 R^{2} of up to 0.77—matching or surpassing state-of-the-art models like RNNs and Transformers. Additionally, it provides interpretable and physical consistent outputs, supporting industry, policymakers, and farmers in adapting to extreme weather conditions. Keywords Informed Learning \cdot Yield Prediction \cdot Remote Sensing 1 Introduction (a) Physics-Informed Model (b) Prediction, target and yield loss distribution. Figure 1: (a) Framework for physics-informed yield loss forecasting. Data modalities, including Sentinel-2 imagery and climate variables, are used to train an RNN that predicts biophysical properties for each time point (actual water use ( ET_{a} ), and the crop susceptibility to water scarcity ( K_{y} )). A simulation model predicts the maximum water use ( ET_{x} ). The predicted biophysical properties are used to calculate the actual yield loss ( Y_{l} ). We leverage prior knowledge describing the relationship between water use and the relative yield loss. (b) Performance plots for visual inspection of a single field. Yield data from cereals in Switzerland is shown, harvested in 2020. In response to climate change, closing the gap between potential and actual yields is an urgent task to sustain food security ( Fischer et al. 2012 ) . Extreme weather conditions like droughts and floodings are pressing challenges in the agricultural sector, directly affecting productivity, and causing yield losses ( Arora 2019 ; Molotoks et al. 2021 ) . Digital agriculture has emerged as a key strategy to address these challenges, providing tools for timely, informed decision-making by policymakers, industry stakeholders, and farmers ( Roscher et al. 2023 ) . The response of crop yields to water availability has been a central focus of research for decades, serving as a critical parameter in assessing crop resilience under extreme climate conditions ( Pereira et al. 2015 ) . Traditionally, simulation models, also known as process-based models, have been employed to capture this relationship. These models build upon biological and physical principles to offer high explainability, supporting decisions in areas such as irrigation, fertilization, pest management, and disease control. However, crop simulation models often struggle with high-resolution, multidimensional data, are computationally intensive, and require calibration, rendering their applicability to high spatial resolution (pixel-level) unfeasible. Furthermore, they are often simplified representations of reality, relying on approximations to maintain computational efficiency ( Kang et al. 2009 ) , oftentimes resulting in inaccurate performances ( Leng and Hall 2020 ) . Therefore, machine learning (ML) models are increasingly utilized for crop productivity estimation ( Van Klompenburg et al. 2020 ) . Recent studies demonstrate impressive scalability and accuracy, even at pixel resolution ( Helber et al. 2023 ; Pathak et al. 2023 ) , by handling complex and multidimensional data ( Miranda et al. 2024 ) . However, ML models are often criticized for their black-box characteristics, limiting their interpretability ( Castelvecchi 2016 ) . Additionally, ML are seldomly designed to follow underlying physical principles of plant growth ( He et al. 2023 ) , thereby hardly ever providing meaningful intermediate outputs, which are essential for tracking crop responses to changing environmental conditions over time. There is a growing demand to integrate the strengths of data-driven ML approaches with the interpretability of simulation models ( Dash et al. 2022 ; Kang et al. 2009 ) . While some recent studies attempt to combine ML and prior knowledge for yield prediction, also under water-limited conditions ( Shuai and Basso 2022 ) , these approaches generally focus on data enrichment rather than on producing interpretable and physical consistent sequential outputs of crop physiology, similar to simulation models. Consequently, they often retain the black-box limitations of conventional ML models, providing limited insight into crop physiology by potentially violating governing physical principles. This study seeks to bridge the gap between interpretable simulation models and high-performance ML models by introducing a physics-informed approach to crop yield loss forecasting. We argue that we can learn more precisely the water use and the crop susceptibility to water scarcity, by building upon Sentinel-2 (S2) multispectral imagery, climate data, and simulations, coupled with pixel-level yield data from cereal crops harvested between 2017–2021 in Switzerland. We demonstrate yield loss estimation at the pixel level, grounded on physical principles by sequentially solving the equation for crop yield response to water scarcity ( Doorenbos and Kassam 1979 ) , using an enhanced loss function. Initial results indicate high potential, achieving an R^{2} of up to 0.77, matching or surpassing the performance of state-of-the-art (SOTA) ML models for yield prediction, including RNNs and Transformers, while providing interpretable and physical consistent outputs, supporting industry, policymakers, and farmers in adapting to extreme weather conditions. 2 Methodology Problem Definition The water use, also called evapotranspiration (ET), is the sum of all biophysical processes in which liquid water is converted to water vapor from various surfaces, including topsoil and vegetation. For a specific crop type, the Food and Agriculture Organization (FAO) differentiates between the maximum ( ET_{x} ) and the actual ET ( ET_{a} ). The former represents the water use under standard and non-limiting environmental conditions and is solely impacted by climate conditions and crop specifics, achieving full productivity, i.e., disease-free, well-fertilized, and under optimum soil water conditions. In contrast, the ET_{a} represents the water use under limiting environmental conditions, resulting in a reduction in water use and ultimately causing a reduction of crop productivity. This reduction in productivity is defined as yield loss and typically expressed in relative terms. Various factors cause productivity limiting conditions, including soil infertility, soil salinity, limited soil water content, diseases, and poor management. Especially, in the light of extreme weather conditions, the frequency of severe droughts and floodings is expected to increase, causing either water scarcity or water abundance, negatively impacting ET and crop productivity ( Kang et al. 2009 ; Arora 2019 ) . Assessing the reduction in ET, is a major challenge in agriculture. A detailed description of biophysical processes of the ET is described by the FAO-56 method ( Allen et al. 1998 ) , using the Penman–Monteith equation, recommended for daily ET (mm/day) estimation. The ET under limiting condition is defined by: ET_{a}=K_{s}\cdot ET_{x}, (1) with K_{s} being a stress coefficient, resulting in a reduction of the ET. In an earlier work, the FAO described the relationship between water use and the relative yield loss ( Doorenbos and Kassam 1979 ) , stating that