Estimation of pore-water electrical conductivity in soilless tomatoes cultivation using an interpretable machine learning model

Soilless culture is widely adopted for improving produce quality and yield and increasing input efficiency. Most of the benefits potentially achievable in soilless systems are possible through precise and continuous management and adjustment of plant nutrition. Under operational conditions, the electrical conductivity (EC) is the main driving parameter leading fertigation strategies, but its measure in the drainage water can be not completely representative of the root zone in the growing medium. Nowadays low-cost sensors can be adopted to measure bulk EC (ECb) in the substrate. The Hilhorst equation is commonly used to convert the ECb into pore-water EC (ECw). This equation is widely calibrated for soil cultivation, but unable to perform properly for soilless substrate with high moisture content and water permittivity. In this work, two cultivation cycles of cherry tomato, managed in a closed-loop soilless system, were used to calibrate and validate two alternative models to the above equation (i.e., generalized additive model - GAM, and extreme gradient boost model - XGBoost). The models predicted ECw from the ECb recorded by substrate sensors. Plants were grown in rockwool using two different strategies for nutrient solution refill achieving different ECw trends during the cultivation. The Hilhorst equation confirmed its unsuitability for ECw prediction in soilless systems. ECw prediction through GAM was not satisfying at low and high ECw values. XGBoost was the most suitable model for ECw estimation, particularly at extreme EC values.