Sustainable fertilisation management via tensor multi-task learning using multi-dimensional agricultural data

Abstract

Precision fertilisation is crucial to agricultural system, which enables to balance soil nutrients, save fertiliser, reduce emissions, and increase crop yield productivity. Due to the low-level sensor and network technologies on most farms, it is difficult to acquire diverse and comprehensive agricultural data. Hence, the absence of agricultural data becomes a major obstacle to the applications of machine learning techniques in precision fertilisation. In this work, we investigate a newly acquired real-world agricultural dataset collected from four genuine winter wheat farms in the United Kingdom, which covers various sorts of agricultural information, such as climate, soil nutrients, and crop yield. To deal with the spatio-temporal characteristics of the agricultural dataset, we propose a novel multi-task learning (MTL) approach that utilises a tensor created from original data to efficiently predict the amount and timing of base fertiliser and topdressing. Specifically, the agricultural measurements (such as climatic data, soil nutrients, etc.) are encoded into a three-dimensional tensor, and tensor decomposition is utilised to extract a series of interpretable temporal and spatial latent factors from the raw data. The latent factor is then utilised as a multi-task relationship to train the spatio-temporal tensor prediction model. The temporal latent factor can be regarded as a temporal pattern shared by different farms on the fertilisation operation of the same crop, and the spatial latent factor can be regarded as the influence of different farm locations on the fertilisation operation of the same crop. Extensive experiments are carried out to evaluate our proposed method utilising the real-world agricultural dataset, in comparison to the standard regression models. Results show that our proposed method provide superior accuracy and stability in fertilisation prediction. Moreover, we have constructed a precision fertilisation system that integrates the proposed algorithm and multi-dimensional agricultural data to assist farms in achieving intelligent, precise and personalised farm management and fertilisation decisions.