An online learning method for embedded decision support in agriculture irrigation

Abstract

In view of the advantages of Wireless Sensor Networks (WSNs), acquisition devices, environmental user-interfaces and low-cost monitoring systems in the agricultural and farming domain, an innovative use of automatic learning decision support is proposed to manage the irrigation process. The aim of this work is to develop a low computational cost technique for a smart irrigation support system, which can be implemented into a simple microcontroller. The classical methods in automatic irrigation use basic on-off controller or complex models with a large number of variables and focus to replace the farmer in the control scheme. Conversely, this proposal uses the farmer experience as the center of the closed loop interpreting its irrigation rules and adapting to the crop changes depending on growth cycle, weather and soil sensors’ signals without involving any model. The three weeks dataset for the testing was constructed from a one-week experimental setup with soil-water potential, soil temperature and sunlight sensors by using approximation functions. Moreover, an online algorithm (AdaDelta) based on gradient descent was tested in an adaptive binary classification with a single layer neuron via MATLAB simulation. Preliminary results of this application have shown its potential with an accuracy of 97% and 6.5% mean square error over the reference method, which poses new possibilities to work in this approach and generate precision agriculture applications for low cost and common irrigation plants by using the new age technologies.