Gerenciamento da irrigação na cultura da uva aplicando-se redes neurais artificiais

Abstract

The correct use of water in irrigation is an important strategy to increment the production and productivity of agricultural cultures, because by being applied in a continuous and sustainable way, it can contribute significantly for the increase of the property profitability, and the environment preservation. The correct irrigation brings a higher efficiency in water, energy and raw material usage, being even able to reduce the risks linked to the activity. Based on the studies we can notice that the neural artificial nets constitute a powerful and viable computing tool in the process of information, and they can be very useful to determine the water volume to be used in agricultural cultures considering the measurement of the productivity of the plots and of the volume of precipitation registered in the meteorological installations in the area. So, these models must supply subsides that permit to control the irrigation system in terms of water economy and environment protection. With that, the objective of the work is to manage the irrigation in the grape culture, due to the monthly precipitation, using artificial neural nets. This study was performed in Fazenda Melina (Melina Farm), located in the county of Nova Mutum, at the North of Mato Grosso State. Initially, monthly precipitation data were collected for four years, as well as the irrigation volume applied to the plots between the months of June and September, and the correspondent grape production. It was reduced the number of input variables to avoid complexity in the architecture of the nets and the monthly precipitation data were summed up to one single value. The variety was represented by three binary independent variables (“dummy”), with values of 0 or 1 for the fields of Isabel, Violeta and Niágara. The use of artificial neural nets was with the multiple layers structure. At first, it was made statistics based on 55 observations. After that, different neural net architectures having up to two intermediate or hidden layers and each having at most sixteen neurons were defined and verified, totaling 80 tested combinations. But only 12 presented satisfying numbers, due to the small difference among selection parameters and the test of the program, the average error, the coefficient of determination of the square and the root of the average squared error. Then, it was opted for verifying analytically the normality of the efficiency of the nets applying the tests Kolmogorov-Smirnov, Lilliefors and Shapiro-Wilk, as well as the test t (Student) between the real grape productivity observed values and the estimated by the nets ones, in which the nets 7 and 11 consisting of two hidden layers demonstrated to be as the most appropriated for predicting the grape productivity, due to the variables that characterize the irrigation in the drier months of the year, the total annual precipitation and the variety. Those nets were considered the one which were the closest to the real value collected, bringing trust for their use in future years, making it possible, the programming of rational use of water.