Machine-learned actual Evapotranspiration for an Irrigated Pecan Orchard in Northwest Mexico

Agriculture in semi-arid regions like Northwest Mexico, is typically characterized by heavily irrigated fields surrounded by a desert environment. The strong contrast in surface conditions increases the non-linear and non-local character of the evapotranspiration dynamics at the irrigated fields, in particular through the oasis effect: strong local evaporative cooling, associated with evapotranspiration enhanced by advection of warm and dry air from the surroundings. To estimate evapotranspiration for individual fields, the agricultural practice relies on traditional empirical models (e.g. Makkink, Priestley-Taylor, FAO-Penman-Monteith) that only make use of standard weather station data. The aforementioned empirical models typically rely on arbitrary, manual tuning (e.g. adjusted constants or the application of a locally determined crop factor) to work reliably.

The goal of this study is to explore whether a physics-informed machine learning approach can be used to improve the estimated evapotranspiration for irrigated fields located in a desert environment, without arbitrary tuning after training and only using regionally available data as input. To this end we will focus on a typical irrigated pecan orchard in Northwest Mexico. At this orchard we have obtained a rich multi-year dataset that encompasses eddy-covariance measurements, irrigation data, soil moisture measurements, and meteorological station data (e.g. air temperature, specific humidity, wind speed and direction) at a half-hourly time scale. In addition, we obtained complementary vegetation indices at the scale of the pecan orchard (~100m-1km) from operationally available remote sensing products.

Using this dataset, we first identify and visualize the main non-linear physical processes (including amongst others the oasis effect) that drive the actual evapotranspiration at the irrigated pecan orchard, both on seasonal and daily time scales. Subsequently, we explore to what extent the effect of the previously identified non-linear processes on the actual evapotranspiration, can be captured with two different machine learning techniques (i.e. gradient boosting decision trees and multi-layer perceptrons) that only receive input variables from a regional meteorological station network and the aforementioned remote sensing products. We trained and tested the machine learning techniques on the evapotranspiration flux measured by an eddy-covariance station located at the orchard, where the estimates provided by the physics-inspired FAO-PM method were used as a starting point for the machine learning models. We find that the machine learning techniques primarily show promise in improving the representation of the seasonal dynamics.