Forecasting irrigation scheduling based on deep learning models using IoT (#965)

Authors

Jimenez Lopez, Fabian

Jimenez Lopez, Andres Fernando

Castellanos Patiño, Juan Sebastian

Abstract

Potatoes are one of the staple foods in Boyacá and play an important role in family nutrition and food security in Colombia. Therefore, timely and accurate information on the irrigation of this crop is relevant in agricultural decision-making, the sustainable development of its production and the reduction of unnecessary water consumption. This study estimated the irrigation prescription of a potato crop from crop meteorological information with the support of IoT technologies to solve the problem of inefficient water dosing in the crop. Two deep learning models were developed, a One-Dimensional Convolutional Neural Network (1D-CNN) and a short-term long-term memory neural network (LSTM). Training data was collected daily from a potato crop from 2018 to 2020 using two weather stations located in the UsoChicamocha irrigation district. To predict irrigation prescriptions, deep learning architectures were trained using Python® by selecting input climatic variables measured with a subsystem of sensors installed in the crop and an actuation subsystem with control of Latch-type solenoid valves, both remotely controlled wirelessly. The algorithms were validated by calculating precision metrics such as MSE and coefficient of determination. The results showed that the LSTM model surpassed the 1D-CNN model, obtaining training and validation errors less than 0.096 and presenting greater precision in the estimation of crop irrigation, giving a coefficient of determination R2 between 0.881 and 0.919. Irrigation prediction algorithms using deep learning techniques achieved promising results and serve as a decision support tool for farmers to automatically decide when and how much water to irrigate.