Interpretable Framework of Physics‐Guided Neural Network With Attention Mechanism: Simulating Paddy Field Water Temperature Variations

Abstract

With the development of large‐scale rice cultivation management initiatives in East Asia, there is concern that a reduction in the number of human cultivators per unit area may lead to poor water management, which could result in decreased land productivity, owing to abnormally high‐ and low‐temperature damage to crops. Accurate simulation of paddy field water temperature is important for studying its impact on crops and providing timely information to aid in decision‐making for more efficient management under limited resources. We propose a neural‐network framework that considers the heat transfer by the vegetation canopy and applies physical‐theory constraints in its training. A novel tuning method is proposed to cope with the trade‐off between water temperature accuracy and physical consistency during training to ensure that the calculated water temperature variations in a paddy field enjoy high accuracy and physical consistency. In the experiments, the proposed framework outperforms physical process models and pure neural network models while maintaining high accuracy in the case of sparse data sets. Furthermore, an attention‐mechanism input layer is integrated into the model to rank feature importance, providing global interpretation to the proposed framework. We also perform sensitivity analysis on the physical process and propose models to compare their different strategies of feature ranking. The results show that the two methods have different sensitivities to different feature patterns, but they complement each other. In summary, the proposed model is credible and stable for practical applications and has the potential to guide more efficient paddy management.

A Deep Learning Approach Based on Physical Constraints for Predicting Soil Moisture in Unsaturated Zones

“…This combined implementation is not static and can be extended and generalized to other problems in a more flexible way. The underlying theory‐driven model for deep learning does not only need to be LSTM (Cho & Kim, 2022; Read et al., 2019; Xie et al., 2022; Xing et al., 2022) but can also be replaced with other machine learning or deep learning methods validated and applied in the field of hydrology under different demand scenarios. Examples are Gate Recurrent Unit (GRU) (Huang et al., 2022), Restricted Boltzmann Machine (RBM) (Xing et al., 2022), Convolutional Neural Network (CNN) (Mo et al., 2017, 2019), Multilayer Perceptron (MLP) (Vincent De Paul Adombi et al., 2022), and Random Forests (RF) (Zahura et al., 2020).…”

“…Physical inconsistency (PI) indicates that the prediction results of the model deviate from the physical meaning (Xie et al, 2022), and the degree of deviation can be expressed by the PI index, which is a dimensionless, nonnegative number, and the larger the value, the less it conforms to the physical meaning; when PI is 0, it means full compliance. The PI is calculated as follows:…”

“…Examples are Gate Recurrent Unit (GRU) (Huang et al, 2022), Restricted Boltzmann Machine (RBM) (Xing et al, 2022), Convolutional Neural Network (CNN) (Mo et al, 2017(Mo et al, , 2019, Multilayer Perceptron (MLP) (Vincent De Paul Adombi et al, 2022), and Random Forests (RF) (Zahura et al, 2020). In addition, the physical constraints used in combining the two different driving methods can be replaced according to the specific context of the different research problems and applied to other areas of hydrological research, such as replacing the core physical process from the water balance principle and Richard's equation in this study with similar elements in other problems, such as the Navier-Stokes equation，the heat transport equation, the basic differential equation for groundwater seepage, and the Saint-Venant equation for river flow, to solve other problems accordingly (Huang et al, 2022;Kamrava et al, 2021;Ma et al, 2022;Read et al, 2019;Vincent De Paul Adombi et al, 2022;Xie et al, 2022). All of these hybrid modeling implementations use a framework similar to the PIDL in this study to achieve the combination of physical constraints and deep learning and obtain good results, confirming the generalizability of this combined framework in the field of hydrological research.…”

“…For data sets with more complex laws and patterns (e.g., data sets with more complex lithological stratification and deeper observation points), the prediction accuracy of PIDL is slightly lower (Figures S5 and S6 in Supporting Information S1) because for simple patterns of SM change, the deep learning model can capture the corresponding laws more easily. In general, for data-driven models, high accuracy prediction becomes more difficult when the fluctuations in the data set are larger and the patterns are more complex, and PIDL is an improved version obtained from LSTM, so this limitation still exists as well (Xie et al, 2022).…”

Optimal Postprocessing Strategies With LSTM for Global Streamflow Prediction in Ungauged Basins

“…Hybrid models have been extensively used in many fields, such as climate simulation, temperature prediction, and streamflow forecasting (Read et al, 2019;Tsai et al, 2021;Xie et al, 2021;Y Zhu et al, 2022). W Xie, Kimura, et al (2022) proposed a neural-network framework that consider heat transfer by vegetation canopies and has the potential to guide more efficient paddy management. Wi and Steinschneider (2022) showed that streamflow projections under warming can be improved if LSTM incorporates additional inputs from reliable physical-based models.…”