Rede neural artificial aplicado ao manejo de irrigação

Abstract

Irrigation is an agricultural practice that leads to high crop production, however the success of this practice depends largely on correct computation of the timing of the application to avoid excessive or deficit application. Thus, the use of moisture sensors, such as capacitive sensors for determining soil moisture combined with artificial neural networks (ANNs) to calculate irrigation time can be a promising tool for automation of irrigation systems. The objective of this work was to develop an ANN that estimate the irrigation time and to contrast the results with the management based on a volume balance method on a watermelon field. Multilayer perceptron types of ANNs were tested. For ANNs training, data obtained in previous harvest were used. The watermelon field was located in Baixo Acaraú Irrigation District in Ceará State – Brazil, where soil moisture was determined using capacitive sensors developed at the Universidade Federal do Ceará (UFC). Networks were tested for two growing stages. The first stage spanning from 0 to the 30th day after seeding (DAS) and the second stage from the 31st to 60th DAS at harvesting. Networks were tested with 2 and 4 inputs, with 5, 10 and 20 neurons in the intermediate layer (NIL) and 1,000, 5,000 and 10,000 epochs. Upon training, the artificial neural networks were field-tested for validation by comparing their responses to the volumetric water balance method (VWBM) for the second stage an succeeding crop cycle. It was found that networks with four entries presented the largest mean square error, converging rapidly to values close to zero, compared the networks with two entries. For the NIL, it was not found significant difference in the mean square error between all 3 tested architectures, therefore it was not necessary to test networks larger than 5 NIL for this application. For the number of training epochs, the one with the best fit values were networks with 10,000 epochs for the first stage of the crop cycle, and 5,000 epochs for the second stage of the crop cycle. It was found no statistical difference in watermelon yield between the two irrigation timing strategies tested (ANN and VWBM). Therefore the artificial neural network was efficient in irrigation management in the field even though the network was presented to some values not occurring during the training process. Thus, one can conclude that the ANN for best performance was a 4-5-1 with learning rate 0.9, and 10000 and 5000 training epochs, respectively in the first and second crop stage. In addition, it was found that the network successfully scheduled the irrigation during the validation process.