Combining Convolutional and Recurrent Neural Networks to Improve Greenhouse Microclimate Mapping

Recurrent Neural Networks (RNNs) and Convo-lutional Neural Networks (CNNs) are two machine learning models successfully used in literature for time series analysis. In particular, while RNNs can correctly capture the long-term dependencies in sequential data, CNNs can be used to learn local patterns inside data properly. In this paper, we propose a model based on the combination of an RNN and a CNN to correctly map the micro-climate inside a greenhouse. The general idea is that climatic conditions registered at a particular moment can depend not only on the value of such conditions measured some time before but also on some contextual factors happening at the given moment. In fact, several contextual factors can affect a climatic variable in different and changing ways. We explore some possible combinations of an RNN and a CNN for contextual time series analysis, and we experiment with their effectiveness in a real-world scenario regarding a greenhouse located in Verona, a city in northeastern Italy.