Humidity forecasting in a potato plantation using time-series neural models

•Machine Learning models (using supervised learning) are proposed and validated under the frame of precision irrigation.•Time-series neural networks are combined with interpolation methods to precisely forecast the humidity level in soil.•The proposal is validated when applied to real-life data from a potato plantation, recorded with 3 probes installed at different depth.•There is not a combination of interpolation and neural models that clearly outperforms the alternate ones for all probes. It is widely acknowledged that, under the frame of sustainable farming, using the minimum water resources is a relevant requirement. In order to do that, precision irrigation aims at identifying the irrigation needs of plantations and irrigate accordingly. Artificial intelligence is a promising solution in this field as intelligent models are able to learn the soil moisture dynamics in the soil-plant-atmosphere system and then generating appropriate irrigation scheduling. This is a complex task as the phenology of plants and its water demand vary with soil properties and weather conditions. The present research contributes to this challenging task by proposing the application of neural networks in order to learn the time-series evolution of irrigation needs associated to a potato plantation. Several of such models are thoroughly compared, together with different interpolation methods, in order to find the best combination for accurately forecasting water needs. In order to predict the soil water content in a potato field crop, in which soil humidity probes were installed at 15, 30, and 45 cm depth during the whole cycle of a potato crop. This innovative study and its promising results provide with significant contributions to address the problem of predicting and managing groundwater for agricultural use in a sustainable way.