Multi-task neural networks for multi-step soil moisture forecasting in vineyards using Internet-of-Things sensors

Promoting an efficient management of water resources is one of the most crucial challenges in smart farming for the coming years. In this context, developing accurate soil moisture forecasting methods is fundamental in order to optimize irrigation and avoid waste. In this paper, we present a deep learning approach based on the multi-task paradigm, which is exploited to jointly forecast soil moisture at multiple time steps in the future, using a multivariate time-series as input features. Experiments are conducted on a real data set collected via data fusion techniques from Internet-of-Things (IoT) sensors located in a vineyard in Montalcino (Tuscany), showing the advantages of joint multi-step forecasting for prediction horizons that range from 24 to 48 hours ahead.