Pure and physics-guided deep learning approaches for spatio-temporal groundwater level prediction
Authors: Matteo Salis, Gabriele Sartor, Rosa Meo, Stefano Ferraris, Abdourrahmane M. Atto
Organizations: Computer Science Department, University of Turin, Turin, Italy · LISTIC Laboratory, Université Savoie Mont Blanc, Annecy-le-vieux, France · Interuniversity Department of Regional and Urban Studies and Planning, Politecnico di Torino and University of Turin, Turin, Italy
Groundwater represents a key element of the water cycle, yet it exhibits complex and context-dependent relationships that make its modeling challenging. Theory-based models have been the cornerstone of scientific understanding. However, their computational cost, simplifying assumptions, and calibration requirements limit their use. In recent years, data-driven models have emerged as powerful alternatives. In particular, deep learning has proven to be a promising approach for its design flexibility and ability to learn complex relationships directly from the data without requiring extensive domain information. We proposed an attention-based pure deep learning model, named STAINet, to predict weekly groundwater levels in Piedmont (Italy), leveraging both irregular groundwater time series and weather image sequences. To enhance the model's trustworthiness and generalization ability, we merged the theory and data-driven approaches by considering physics-guided strategies to inject the groundwater flow equation into the model. Firstly, we restructured the tail of the architecture to predict the three terms of the governing equation, named the autoregressive, diffusion, and residual components - we thus obtained the PSTAINet-IB. Then, we further injected physics priors by adding loss terms related to the estimated equation components, obtaining the PSTAINet-ILB model. Lastly, we developed the PSTAINet-ILRB by imposing a loss term specific to the residual component, which forces the groundwater recharge to occur within the groundwater body recharge zone, which is identified by domain experts. The models were evaluated both by feeding true lagged values as input and by iterating their own predictions (rollouts) over the whole test set. The PSTAINet-ILB model performed the best, achieving remarkable test performance, and generating equation components in line with domain experts' expectations.
Figures & tables
Figure 1 : On the left, the Piedmont region with the ROI box and the selected piezometers with a colormap representing the elevation; bold names represent Piedmont’s Provinces. On the right, a zoom over the ROI with an additional colormap representing the mean groundwater levels; bold names represent cities.
Figure 2 : Time series of the sensors in a) Cuneo, b) Racconigi, c) Scalenghe, and d) Vottignasco.
Figure 3 : The adopted modeling pipeline. Each of our models takes as input the weather and groundwater data from a window of length T+1 and T , respectively, from the target time t∗ . The predictions are compared with true measurement data, and through backpropagation, the weights of the model are updated to lower such differences. What differs between models is whether or not the physics knowledge is leveraged (dashed lines). This could be done by applying an inductive bias or a learning bias that acts on the model architecture and on the learning algorithm (backpropagation), respectively.
Figure 4 : a) Multi-Head Attention block adopted in the models. b) STSA Module c) STAI Module that takes as values the hidden representation of the input points’ value embeddings (ST-VEmb), as keys the input points’ spatio-temporal coordinates embeddings (ST-CEmb), and as queries the prediction points’ spatio-temporal coordinates embeddings (ST-CEmb).
Figure 5 : STAINet architecture.
Figure 6 : a) D Module in which Act & Norm stand for activation and normalization respectively; b) ΔGW Module.
Figure 7 : PSTAINet-IB architecture.
Model
αortho
αcoh
αdiff
α∣∣R∣∣1
α∣∣R∣∣2
αRCH
STAINet
10−3
-
-
-
-
-
PSTAINet-IB
10−3
-
-
-
-
-
PSTAINet-ILB
10−3
1
2.5⋅10−2
5⋅10−4
10−4
-
PSTAINet-ILRB
10−3
1
2.5⋅10−2
5⋅10−4
10−4
5⋅10−4
Table 1 : Loss terms weights.
STAINet
PSTAINet-IB
PSTAINet-ILB
PSTAINet-ILRB
True Data
Rollout
True Data
Rollout
True Data
Rollout
True Data
Rollout
NBIAS
0.1042
0.1713
0.0594
0.0862
-0.0093
0.0124
0.0177
0.0685
(0.2297)
(0.2297)
(0.1741)
(0.2028)
(0.1137)
(0.1351)
(0.2212)
(0.1636)
RMSE [m]
0.8351
1.0292
0.6296
0.7336
0.5058
0.5851
0.6521
0.6783
(0.8635)
(1.2064)
(0.8175)
(0.9495)
(0.3031)
(0.4485)
(0.5347)
(0.8296)
MAPE [%]
0.2834
0.2868
0.2319
0.2606
0.1479
0.1604
0.1900
0.1953
Table 2 : Median performance metrics on the test set feeding true data (True Data) and in the rollout setting (Rollout). Numbers in parentheses represent the Interquartile Range (IQR).
Figure 8 : Violin plots representing MAPE values for the proposed models on the test set. Q1 , Q2 , Q3 represent the first, second (median), and third quartiles, respectively. The shapes around the box plots are the KDE density estimates of the MAPE values over the sensors.
Figure 11
Figure 9 : STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB predictions on the test set feeding true data in a) , c) and in the rollout setting b) , d) in Racconigi and Scalenghe sensors.
Figure 10 : STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB groundwater level prediction maps at 1.5 km/pixel for 2022-07-17.
Figure 11 : Predicted equation components by PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB for the weeks starting on 2022-07-17 ( a ), and 2022-11-13 ( b ). The first column reports Δ^GWt∗ , the second R^t∗ , and the last D^ . Green lines identify the recharge zones. For Δ^GWt∗ , a blue pixel means inflow while a red one means outflow; for R^t∗ , a blue pixel means recharge while red represents a loss.
NBIAS
RMSE [m]
MAPE [%]
NSE
KGE
PSTAINet-ILB
-0.0278
0.4490
0.1404
0.6314
0.7547
Table 3 : PSTAINet-ILB time series reconstruction median Metrics, in the rollout setting with a 26 time step forecast horizon.
Figure 12 : PSTAINet-ILB time series reconstruction in the rollout setting with a 26 time step forecast horizon for a) Racconigi and b) Scalenghe piezometers.
Figure 13 : Loss plots a) training L , b) test Ldata , c) training Lcoh , d) training Ldiff , e) training LRCH .
Figure 14 : Correlation plots between MAPE and missing values percentage for STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB; in all plots, the red line is the bisector, r is Pearson’s correlation coefficient, and pvalue is the corresponding p-value.
Figure 15 : Monthly average MAPE over all sensors in the rollout setting on the test set. Vertical lines represent the corresponding monthly standard deviations.
Figure 16 : Correlation plots between MAPE and Trend Strength ST for STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB; in all plots, the red line is the bisector, r is Pearson’s correlation coefficient, and pvalue is the corresponding p-value.
Figure 17 : Correlation plots between KGE and Trend Strength ST for STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB; in all plots, the red line is the bisector, r is Pearson’s correlation coefficient, and pvalue is the corresponding p-value.
Figure 18 : Correlation plots between MAPE and Seasonal Strength SS for STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB; in all plots, the red line is the bisector, r is Pearson’s correlation coefficient, and pvalue is the corresponding p-value.
Figure 19 : Correlation plots between KGE and Seasonal Strength SS for STAINet, PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB; in all plots, the red line is the bisector, r is Pearson’s correlation coefficient, and pvalue is the corresponding p-value.
Figure 20 : NBIAS for all sensors in the rollout setting on the test set.
Figure 21 : MAPE for all sensors in the rollout setting on the test set.
Figure 22 : Box plots representing predictions for Δ^GWt∗ ( a) ) and R^t∗ ( b ) generated by PSTAINet-IB, PSTAINet-ILB, and PSTAINet-ILRB. The star in each box plot represents the mean, while the horizontal line inside the box is the median.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Bar plot representing the percentage of missing data for each sensor considered in the study. In the x-axis, sensors are labeled using both the municipality where they are placed and the identification code (sensor ID).
Figure S2 : Bar plot representing the number of observations for each sensor in the training, validation, and test sets (also reported by the vertical numbers for each bar, respectively). In the x-axis, sensors are labeled using both the municipality where they are placed and the identification code (sensor ID).
Figure 29Figure 30
Figure S3 : Groundwater level time series.
Figure 32Figure 33
Figure S4 : Test set predictions in the rollout setting.
Figure 35Figure 36Figure 37
Figure S5 : Test set predictions in the rollout setting.
Time series forecasting has seen signicant advancements with the emergence of new deep learning models. However, forecasting time series in applications involving physical processes remains a major challenge. Despite the apparition of Physics Informed Neural Networks (PINN), recent models do not estimate unobservable intermediate physical variables, which are important for domain experts to understand the target behavior. To this end, we propose a Physics Informed Recurrent Neural Network (PIRNN) which predicts, along the target, unobservable variables on both historic data and forecast target. This approach enhances the model robustness and results interpretation using domain knowledge. Our method is easily adaptable to any physical model using several equations, each having its own set of unobservable variables, to describe it-self. As a case study, we incorporate physical equations used for groundwater levels predictions by the physical model called Gardenia. This model uses transfers equations between reservoirs, optimized with data assimilation, to simulate the evolution of groundwater levels. Evaluation includes several well known neural network models and the Gardenia model compared on twelve real world datasets. In addition, we study the impact of each component through an ablation study. Our model outperforms other models on ve out of the twelve datasets and our ablation study underlines the importance of having a physical background in our time series forecasting task. Finally, the coherence of the physical variables predicted by our neural network is assessed by a domain expert.
Etienne Lehembre, Pascal Audigane, Vincent Nguyen +2
CA, LIFO · Université d’Orléans, INSA CVL, LIFO, UR 4022, Orléans, France · BRGM +3
Accurate and scalable flood mapping remains challenging due to limited ground observations, heterogeneous terrain conditions, and the difficulty of enforcing hydrodynamic consistency within data-driven models. This work introduces a physics-guided deep learning framework that integrates multi-modal remote sensing (Sentinel-1 SAR, Sentinel-2 optical imagery, and DEM-derived terrain features) with constraints from the depth-averaged shallow water equations (SWE). The proposed hybrid architecture combines a UNet to capture fine-scale spatial details with a Fourier Neural Operator (FNO) to model basin-scale hydraulic interactions, while physics-informed residual losses ensure mass and momentum consistency. Evaluated across diverse floodplain settings, the hybrid model achieves an Intersection over Union of 0.82 and an F1 score of 0.90 for flood extent prediction, outperforming UNet-only and FNO-only baselines. Using hydrodynamic simulations as reference data, the model achieves an RMSE of 0.21 m for water depth and 0.15 m/s for flow velocity. Physics consistency is maintained, with low residuals and mass imbalance below 2.1%. Ablation studies confirm that removing physicsbased regularization significantly degrades performance, underscoring the value of physical constraints for stability and generalization. These results demonstrate that embedding hydrodynamic principles into deep learning yields more accurate, reliable, and physically coherent flood predictions, offering strong potential for operational monitoring and large-scale deployment.
Built Environment Department, College of Science and Technology, North Carolina A&T State University, Greensboro, NC 27405, USA · United Nations University Institute for Water, Environment and Health, Richmond Hill, ON, Canada
The research explores the pioneering integration of Physics-Informed Neural Networks (PINNs) into the domain of Ground-Penetrating Radar (GPR) data prediction. This research presents a detailed development framework for a specialized PINN model, proficient at interpreting and forecasting GPR data, much like how medical imaging models predict tumor behavior. By harnessing the synergy between deep learning algorithms and the physical laws governing subsurface structures or in medical terms, human tissues the model effectively embeds the physics of electromagnetic wave propagation into its architecture. This ensures that predictions not only align with fundamental physical principles but also mirror the precision needed in medical diagnostics for detecting and monitoring tumors. The suggested deep learning structure comprises three components: a CNN, a spatial feature channel attention (SFCA) mechanism, and ConvLSTM, along with temporal feature frame attention (TFFA) modules. The attention mechanism computes channel attention and temporal attention weights using self-adaptation, thereby fine tuning the visual and temporal feature responses to extract the most pertinent and significant visual and temporal features. By integrating physics directly into the neural network, our model has shown enhanced accuracy in forecasting GPR data. This improvement is vital for conducting effective assessments of bridge deck conditions and other evaluations related to civil infrastructure. The use of Physics Informed Neural Networks (PINNs) has demonstrated the potential to transform the field of Non-Destructive Evaluation (NDE) by enhancing the precision of infrastructure deterioration predictions. Moreover, it offers a deeper insight into the fundamental mechanisms of deterioration, viewed through the prism of physics-based models.
Mehrdad Shafiei Dizaji, Hoda Azari
Post-Doctoral Research Fellow, Federal Highway Administration, Turner-Fairbank Highway Research Center, McLean, VA 22101 · Nondestructive Evaluation Program and Laboratory Manager, Federal Highway Administration, Turner-Fairbank Highway Research Center, McLean, VA 22101