SEPARATE: A tightly coupled, seamless IoT infrastructure for deploying AI algorithms in smart agriculture environments

Abstract

Precision agriculture generates large datasets from IoT infrastructures deployed for continuous crop monitoring. This data requires analysis to usefully transform this data deluge into insights that can deliver value-generating services to farmers in a timely manner. This paper introduces SEPARATE; a dynamic interoperable and decentralized infrastructure for executing both, training and inference stages of deep learning (DL) algorithms in smart agriculture scenarios. The presented infrastructure allows the execution of the inference stage at the edge, achieving a highly efficient and responsive local temperature prediction service to take actions based on the predictions generated. Moreover, the training stage is offloaded to the cloud along with the generated historical data, allowing the trained model to be periodically updated at the edge. On the one hand, our results show that the Convolutional Neural Network model together with the Long Short-Term Memory technique (CNNLSTM) obtains the best results in both prediction accuracy and computational time. On the other hand, an analysis has been carried out to determine how often the model must be retrained, obtaining results that indicate that from day 9-10, it would be necessary to retrain the model, although, until day 20, the precision is not greatly reduced. Moreover, the SEPARATE infrastructure enables the execution of real-time inference from sensor-generated data and seamless model retraining in an operational greenhouse for temperature forecast with satisfactory performance.