Precision Fertilization via Spatio-temporal Tensor Multi-task Learning and One-Shot Learning

Precision fertilization is essential in agricultural systems for balancing soil nutrients, conserving fertilizer, decreasing emissions, and increasing crop yields. Access to comprehensive and diverse agricultural data is problematic due to the lack of sophisticated sensor and network technologies on the majority of farms, and available agricultural data are generally unstructured and difficult to mine. The absence of agricultural data is, consequently, a significant impediment to the utilization of machine learning approaches for precision fertilization. In this research, we investigate newly gathered genuine agricultural dataset from nine real winter wheat farms in the United Kingdom, which encompass an extensive variety of agricultural variables, including climate, soil nutrients, and farming data. To deal with the spatio-temporal characteristics of agricultural dataset and to address the problem of scarcity in agricultural data, we propose a novel machine learning approach integrating multi-task learning and one-shot learning, which utilizes a multi-dimensional tensor constructed from original data combined with fertilization temporal patterns extracted by contrasting with environmental information from existing real farms to accurately predict the quantity and timing of base and top dressing fertilization. Specifically, agricultural data are converted into a 3-D tensor and tensor decomposition technique is utilized to derive a set of comprehensible spatio-temporal latent factors from the original data. The latent factors are subsequently utilized to construct the spatio-temporal tensor prediction model as multi-task relationships. The proposed one-shot learning approach utilizes the Mahalanobis distance to evaluate the similarity of environmental information between the target farm and existing real-world farms as a determinant of whether to transfer the fertilization temporal pattern of existing farm to the target farm. Comprehensive experiments are conducted to compare the proposed approach with standard regression models utilizing the real-world agricultural dataset. The experimental results demonstrate that our proposed approach presents superior accuracy and stability for fertilization prediction.