Irrigation volumes detection through ensemble physical modelling

Agriculture is the major freshwater user worldwide, averaging 70% of the water resource consumption. Despite this heavy incidence, irrigation represents the most uncertain water flux, difficult to predict because of both anthropogenic and natural factors. In this work, a methodology to detect the irrigation signal is presented, integrating ground and satellite data within energy-water balance modelling. The proposed algorithm adopts a Montecarlo approach, simulating the impacts of hundreds of thousands of possible irrigation schedules over selected hydrological variables and extracting the most likely event series by comparing the model results with different kinds of references. This approach guarantees the physical soundness of the irrigation detection procedure by adding hydrological robustness to the different observed signals affected by irrigation. In increasing steps of uncertainty, model results are compared with those from a benchmark simulation (fed with observed irrigation data), with in-situ measurements and satellite observations. This provides a complete framework to the algorithm reliability, and the inclusion of satellite imagery allows to export the procedure to data-poor areas. In this work, numerous variables were tested to identify the fittest for the analysis, specifically: surface soil moisture (SSM), deep soil moisture (SM2), evapotranspiration (ET) and land surface temperature (LST). Of these, SSM qualified as the most suited to the algorithm, as differences in irrigation timings caused little spread in the other variables ensembles.

The used hydrological model is the FEST-EWB, an energy-water balance model where the two equations are coupled and solved jointly looking for the land surface temperature that closes the system.

The procedure was tested over a variety of Italian field case studies where eddy covariance stations are available, ranging from semi-arid to wet climates, from on-demand to turn irrigation, from homogeneous to heterogeneous agricultural landscapes, and including low-lying, high-stemmed and arboreal crops. The results indicated three main conclusions: (1) the algorithm works best over fields with fewer irrigation events in a season (<10), as very frequent events (>2-3 per week) crowd the signal and can make the procedure redundant; (2) high-ET periods (e.g., summer and/or high-vegetation density periods) within the agricultural seasons increase the  ensemble spread and improve the efficacy of the procedure, allowing to better distinguish between different irrigation schedules; (3) uncertainty in satellite retrievals of SSM, specifically over heterogeneous agricultural landscapes, negatively influences the accuracy of the algorithm by muddling the signal coming from the target field.