(579a) A Physics-Informed Deep Learning Approach to Predict Soil Water Content for Agricultural Decision-Making

Global warming has imposed severe climatic patterns, causing extensive droughts and water shortages and adversely affecting ecological processes [1]. Soil water content is one of the critical agricultural parameters susceptible to the adverse impacts of global warming. It directly affects the soil’s physical and chemical properties, plant growth, and microbial activity, influencing agricultural productivity and environmental quality. Unprecedented fluctuations in soil water content implicitly impact irrigation, crop growth, and agricultural productivity. The accurate measurement and modeling of soil water content are thus fundamental for water resource management, climate change adaptation strategies, and agrarian decision-making [2-4]. Combining soil physics and transport phenomena has resulted in the development of Richards’s equation for predicting soil water content. However, the complexity of Richards’s equation requires simplifying assumptions to compute soil water content values, leading to a Markov chain-based model. Nevertheless, this model fails to capture all soil water content patterns.

This work proposes a physics-informed neural network (PINN) approach for precise soil water content predictions that benefit agricultural decision-making [5]. This approach integrates soil physics principles with the temporal pattern recognition capabilities of recurrent neural networks (RNNs), enhancing soil water content predictions compared to existing analytical methods. The uncertainty-quantifying PINN model reveals subtle soil water content patterns that analytical Markov chain-based models cannot demonstrate. The proposed PINNs learn deep soil water content temporal patterns and are aware of soil physics. Further evaluations reveal that PINNs outperform data-driven and soil-physics-based models by incorporating soil physics domain knowledge into deep learning algorithms. The uncertainty associated with soil water content predictions by the Markov model is quantified by training RNNs, such as long short-term memory (LSTM) networks, on real-time soil water content data available from the Kansas mesoscale network (Mesonet) [6,7]. Furthermore, multiple Mesonet time series have been used to evaluate the influence of physics-based knowledge on PINN predictions. The results indicate significant improvements in soil water content prediction using PINNs compared to the Markov-based model.

References:
[1] J. Hansen, M. Sato, and R. Ruedy. Perception of climate change. Proceedings of the National Academy of Sciences, 109(37):E2415–E2423, 2012.

[2] G.C. Topp, G.W. Parkin, T.P.A. Ferre, M.R. Carter, and E.G. Gregorich. Soil water content, volume 2. Soil sampling and methods of analysis, 2008.

[3] C. Gardner, D. Robinson, K. Blyth, and J.D. Cooper. Soil and environmental analysis, Soil water content, pages 13–76. CRC Press, 2000.

[4] H. Yan, C. M. DeChant, and H. Moradkhani. Improving soil moisture profile prediction with the particle filter-Markov chain Monte Carlo method. IEEE Transactions on Geoscience and Remote Sensing, 53(11):6134–6147, 2015.

[5] M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.

[6] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436-444, 2015.

[7] A. Patrignani, M. Knapp, C. Redmond, and E. Santos. Technical overview of the Kansas mesonet. Technical report, Journal of Atmospheric and Oceanic Technology, 2020.