Neural Networks and Low Cost Sensors to Estimate Site-Specific Evapotranspiration

Abstract. Efficient irrigation depends on matching application rates to the crop water requirement. Precision irrigation systems also require localized crop water data. One direct measure of crop water consumption is the actual evapotranspiration (ET). Growers have many resources to estimate ET and crop water use, but existing methods don‘t provide real-time, site-specific ET. Without local and real-time data, science based ET estimates lack the precision required for incremental allocation decisions.

Adaptive neural networks (ANNs) can be trained to convert data from low cost sensors into site-specific ET estimates. ANNs can also incorporate publicly available data from weather networks and satellite imagery. The resulting site-specific algorithm can allow growers to obtain reliable, site-specific ET estimates affordably.

Data from a related field study on crop water use was used to train an ANN and demonstrate the proof of concept, using eddy covariance for ANN training. Initial results show that once trained, ANNs can determine ET from a small array of sensors at least as well as the Penman-Monteith (P-M) equation under most conditions. Under drought stress conditions, the ANN could predict ET more accurately than P-M. ANNs were also used to compensate for broken and unreliable sensors. Ongoing field research in 2017 will explore the robustness of the ANN method for measuring ET in different climactic regions and over different crop covers. Research objectives include: minimizing the required training time; optimizing a list of required low cost sensors; determining ANN stability over seasonally varying growing conditions.