Redes neurais artificiais aplicadas à modelagem da variabilidade espacial de atributos físico-químicos de solos do cerrado

Abstract

Agriculture, like other activities is becoming global and therefore, it suffers competition and influence from what happens around the world. This behaviour is forcing the use of better management in order to minimize costs and increase productivity. The precision agriculture provides the necessary technology to achieve this level of management. Additionally, it enables the acquisition of wide range data, proper processing of this data and subsequent use of the information for effective improvement. Consequently, it optimizes material, natural and human resources and therefore allows higher productivity. One of the factors that influence the production is the spatial and temporal variability of soil properties. By knowing this variation, one can use management techniques to adjust the soil to the needs of each crop. By using the geostatistics inserted in precision agriculture, it is possible with a certain number of sample points to detect and model this variability, inferring the understanding of the area with the necessary detail. The Artificial Neural Networks (ANN), a branch of artificial intelligence, seeks to emulate human reasoning and is able to perform data inference and it is considered a universal approximator with ability to learn. One of the ANN features is the capability to establish multidimensional characteristic functions to identify presented patterns or object classes. Thus, it has the capacity of becoming a reasonable alternative to successfully implement modeling of spatial variability. Were applided the ANN for modeling the spatial variability of soil attributes and to accomplish this goal the following tasks have taken place: data collection, descriptive statistics and geostatistics analysis. The definition, training of different ANN and consequent choice of networks that had lower mean error have finished with the comparison between estimated versus measured results and the calculation of the mean relative error ending with comparing the estimates values made by ordinary Kriging of the atributes that presented spatial dependence. Consequently, it was possible to conclude that ANN has the potential to accomplish the modeling of spatial variability of physical and chemical soil properties.