Minute-wise frost prediction: An approach of recurrent neural networks

Abstract

Frost events incur substantial economic losses to farmers. These events could induce damage to plants and crops by damaging the cells. In this article, a recurrent neural network-based method, automating the frost prediction process, is proposed. The recurrent neural network-based models leveraged in this article include the standard recurrent neural network, long short-term memory, and gated recurrent unit. The proposed method aims to increase the prediction frequency from once per 12–24 h for the next day or night events to minute-wise predictions for the next hour events. To achieve this goal, datasets from NSW and ACT of Australia are obtained. The experiments are designed considering the scene of deploying the model to the Internet of Things systems. Factors such as model processing speed, long-term error and data availability are reviewed. After model construction, there are three experiments. The first experiment tests the errors between different model types. The second and third experiments test the effect of sequence length on error and performance for recurrent neural network-based models. All tests introduce artificial neural network models as the baseline. Also, all tests for model error are conducted in two rounds with testing datasets from the current year (2016) and next year (2017). As a result, recurrent neural network-based models are more suitable for short-term deployment with a smaller sequence length. In contrast, artificial neural network models demonstrate a lower error over the long term with faster processing time. With the results presented, the limitations of the proposed method are discussed.

Keywords

• Frost prediction
• Internet of Things
• Machine learning
• Recurrent neural network
• Temporal prediction