Greenhouse Climatic Sensing through Agricultural Robots and Recurrent Neural Networks

In today’s greenhouse farming the assumption of climatic uniformity is no more acceptable since it may lead to sub-optimal decisions; however, the deployment of a sensor for each point of interest is expensive and incompatible with field operations. Agricultural robots are quickly moving from the research field to real scenarios for weed removal, harvesting and other agronomic operations. In this paper, we show the possibility, enabled by Recurrent Neural Networks, to exploit successive samplings of the mobile robot in its wandering in the greenhouse to infer the value of the desired climatic variable in specific points of interest. More specifically, we use recurrent neural networks to model the dependency of the climate conditions in each point of interest as a function of the distance with respect to various points on robot’s trajectory. The proposed technique has been applied to a real dataset coming from a greenhouse located near Verona and compared with traditional approaches such as considering a single sensor located at the center of the greenhouse and compressed sensing.