Towards Smart Farming: Fog-enabled intelligent irrigation system using deep neural networks

Abstract

The most amount of withdrawn freshwater in the world is used for agriculture activities to extract essential products for human survival. Smart Farming can manage and optimize the water amount given to crop fields in various crop development stages, weather, and soil condition. Sensors installed in several monitoring spots can gather soil moisture of the crop field to indicate the level of water retained in it. However, due to connectivity or sensor failure problems, the smart farming system can not receive the soil moisture data given by the irrigation management. Deep learning techniques can predict soil moisture data using other irrigation management types as data from weather, crop, and irrigation systems. The Fog Computing paradigm also tackles the connectivity problem in the farms since it extends the traditional cloud computing architecture to the edge of the network, providing edge nodes with computational resources to process and analyze sensor requests. In this context, we propose different deep neural network architectures to build prediction models of soil moisture. We also handle the problem of missing data for the dataset features. For this purpose, we use KNN data imputation, which requires filling the values of unknown (or missing) features with values that ensure a desired degree of reliability. Finally, we also embedded the prediction models on a small single-board computer, often used as a fog node, to evaluate the performance of the prediction models according to CPU and RAM usage. This evaluation showed a maximum consumption increase of 10% CPU and 1% RAM. Therefore, the models are viable to fog architectures in the Internet of Things context. Our results indicate that the predictive models achieve a satisfactory efficiency to improve irrigation water saving.