Mapping Micro-Climate in a Greenhouse Through a Context-Aware Recurrent Neural Network

Existing greenhouse climatic models usually rely on the assumption that climate conditions inside a greenhouse are uniform. This allows to maintain the model as simple as possible at the expense of less accuracy in the final predictions. However, in real-world applications this uniformity assumption is not satisfactory from the agronomic perspective, since it may lead to wrong decisions and irreversible damages to crops. Conversely, more sophisticated techniques need to collect data from a great number of sensors. In this paper, we try to overcome this situation by proposing the concept of virtual sensor whose behaviour is modeled by a context-aware recurrent neural network trained by the relationship between temporary sensors placed in specific point of interests, and a set of sensors placed in permanent positions in the greenhouse. More specifically, we try to model the dependency of climate conditions as time series where the variability does not depend only on temporal aspects, but also on richer contextual dimensions, like space locations and distance relationships with a predefined and small set of permanent sensors. The proposed technique has been applied to a real-world dataset coming from a greenhouse located in Verona in which one permanent sensor has been placed at the center of the greenhouse and seven temporary sensors have been maintained for a limited period of time.