Machine Learning Enabled Smart Sensor Node for Evapotranspiration Prediction in Diverse Soil and Environmental Conditions

The smart sensor nodes have been built and deployed in an agricultural farm to collect real-time soil characteristics with essential environmental parameters crucial to creating an adaptable prediction model for evapotranspiration (ET) using advanced machine learning techniques. The study particularly demonstrated the implementation of three prominent machine learning models, support vector machine (SVM), extreme learning machine (ELM), and M5P regression tree, validated them with the standard theoretical models: Penman-Monteith and soil water balance (SWB). The developed models are further compared with climate ET data from the Australian Bureau of Meteorology (BOM), addressing the challenges of water accountability in traditional ET estimation methods. Integrating smart sensing with advanced machine learning offers excellent potential to improve spatial and temporal resolution in ET predictions and soil health. In the model development, key input features are measured in real time, such as air temperature (°C), barometric pressure (hPa), wind speed (km/h), relative humidity (RH) (%), rainfall (mm), solar exposure (MJ/m2), soil temperature (°C), and change in soil moisture ( $\Delta $ SM). The study findings suggest that the SVM demonstrated better accuracy ( ${R}^{{2}}=0.97$ ) with minimal prediction error [root mean squared error (RMSE) =0.19 mm/d and mean absolute error (MAE) =0.14 mm/d] in estimating ET, followed by ELM ( ${R}^{{2}}=0.91$ ) and M5P ( ${R}^{{2}}=0.88$ ), respectively. The proposed predictive model effectively captures complex relationships among soil and environmental parameters, providing insights into how different soil types retain or lose water tailored to specific climatic conditions. This allows for enhanced accuracy compared to the findings reported in the existing literature.