Gumière, Silvio José; Camporese, Matteo; Botto, Anna; Lafond, Jonathan A.; Paniconi, Claudio; Gallichand, Jacques; Rousseau, Alain N.
(2020).
Machine Learning vs. Physics-Based Modeling for Real-Time Irrigation Management.
Frontiers in Water
, vol. 2
.
DOI: 10.3389/frwa.2020.00008.
Résumé
Real-time monitoring of soil matric potential has now become a common practice for
precision irrigation management. Some crops, such as cranberries, are susceptible
to both water and anoxic stresses. Excessive variations in soil matric potential in
the root zone may reduce plant transpiration, due to either saturated or dry soil
conditions, thereby reducing productivity. A timely supply of the right amount of water is,
therefore, fundamental for efficient irrigation management. In this paper, we compare
the capabilities of a machine learning-based model and a physics-based model to
predict soil matric potential in the root zone. The machine learning model is a random
forest algorithm, while the physics-based model is a two-dimensional solver of Richards
equation (HYDRUS 2D). After training and calibration on a dataset collected in a cranberry
field located in Québec (Canada), the performance of the two models is evaluated for
30 different time frames of 72-h soil matric potential forecasts. The results highlight
that both models can accurately forecast the soil matric potential in the root zone. The
machine learning-based model can achieve better performance when compared to the
physics-based model, but forecasting accuracy decreases rapidly toward the end of the
72-h lead time, while the error for the Richards equation-based model does not increase
with time and remain small compared to the typical measurement error.
Actions (Identification requise)
 |
Modifier la notice |