Spatio-temporal Tensor Multi-Task Learning for Precision fertilization

Precision fertilization refers to the targeted variable fertilization technology adopted according to the variation of soil fertility on a certain area. Precise fertilization can balance soil nutrients, save fertilizer, reduce environmental pollution and increase food production. In this article, we propose a new multi-task learning (MTL) method, which is based on a tensor composed of farm data to effectively predict both amount and time of base fertilizer and topdressing. Specifically, we encode farm measurement data (e.g., climate data, soil information, etc.) into a three-dimensional tensor, and extract a set of interpretable temporal and spatial latent factors from the original data through tensor decomposition, and utilise latent factors as predictors to train a set of spatio-temporal prediction models. We conduct extensive experiments using data from SoilHealthDB. The experimental results show that the proposed method has advanced fertilization prediction accuracy and stability in terms of root mean square error.