On-device machine learning for forecasting reference evapotranspiration

Abstract

The current climate-world scenario demands the use of natural resources in a balanced way, especially water. The prediction of reference evapotranspiration (ETo) can assist in the management of water resources, being a determining parameter in the early assessment of water for applications such as irrigation scheduling and hydrological studies. The literature shows that Machine Learning (ML) models have been widely applied in the prediction of ETo and have shown promising results, even with the use of reduced parameters. Previous works also presented automated irrigation systems that connect to Web services to predict the parameters needed to make irrigation scheduling. However, a cloud-based system depends on an internet connection infrastructure; such infrastructure is not always available in rural areas. In addition, a device that relies on an internet connection will generally consume more power. Therefore, this research primarily aims to develop and assess a TinyML model designed to forecast ETo, with the goal of establishing an energy-efficient machine learning model suitable for integration into an embedded system. We evaluated three models considering air temperature, relative humidity, and wind speed as inputs. The models were embedded in an ESP32 microcontroller and compared to its cloud versions. The results show that the model that estimates ETo considering only air temperature and humidity is the best option considering the trade-off between cost and precision. Compared to its cloud-based version, this embedded model consumes 37.97% less memory space, uses approximately 99.98% less energy and runs 99.97% faster.