Estimating Greenhouse Climate Through Context-Aware Recurrent Neural Networks Over an Embedded System

The assumption of climate homogeneity is no longer acceptable in greenhouse farming since it can result in less-than-ideal decisions. At the same time, installing a sensor in each area of interest is costly and unsuitable for field operations. In this paper, we address this problem by putting forth the idea of virtual sensors; their behavior is modeled by a context-aware recurrent neural network trained through the contextual relationships between a small set of permanent monitoring stations and a set of temporary sensors placed in specific points of interest for a short period. More precisely, we consider not only space location but also temporal features and distance with respect to the permanent sensors. This paper shows the complete pipeline to configure the recurrent neural network, perform training, and deploy the resulting model into an embedded system for on-site application execution.