Using Neural Networks to Estimate Site-Specific Crop Evapotranspiration with Low-Cost Sensors

Abstract

Irrigation efficiency is facilitated by matching irrigation rates to crop water demand based on estimates of actual evapotranspiration (ET). In production settings, monitoring of water demand is typically accomplished by measuring reference ET rather than actual ET, which is then adjusted approximately using simplified crop coefficients based on calendars of crop maturation. Methods to determine actual ET are usually limited to use in research experiments for reasons of cost, labor and requisite user skill. To pair monitoring and research methods, we co-located eddy covariance sensors with on-farm weather stations over two different irrigated crops (vegetable beans and hazelnuts). Neural networks were used to train a neural network and utilize on-farm weather sensors to report actual ET as measured by the eddy covariance method. This approach was able to robustly estimate ET from as few as four sensor parameters (temperature, solar radiation, humidity and wind speed) with training time as brief as one week. An important limitation found with this machine learning method is that the trained network is only valid under environmental and crop conditions similar to the training period. The small number of required sensors and short training times demonstrate that this approach can estimate site-specific and crop specific ET. With additional field validation, this approach may offer a new method to monitor actual crop water demand for irrigation management.