Estudo de algoritmos de otimização multivariados para a determinação de configuração amostral e tamanho amostral na análise da variabilidade espacial

Abstract

Precision agriculture can be defined as a set of techniques and technologies that can be implemented to improve the decision-making process in agricultural production, as it allows the precise application of fertilizers at each location. As agricultural areas are usually not homogeneous, one of the options to deal with the heterogeneity of the soil or the distribution of chemical and physical attributes is to define application zones. The application zones make it possible to reduce both the spatial and temporal variability of the crop yield under study as well as the environmental impacts. Therefore, the application zones can also represent indicators to guide future soil sampling, aiming at a possible reduction in the sample size. The objective of this work was to determine a better sample resizing (with traditional sampling – Article 1; and with optimization process – Article 2) for a commercial area of soybean cultivation, where an activity of localized application of agricultural inputs is developed, through zones of application generated from the evaluation of five clustering methods (Fuzzy C-means, Fanny, K-means, Mcquitty, and Ward). Soil chemical attributes obtained from an agricultural area located in the municipality of Cascavel, PR, Brazil, referring to four years of soybean harvest (2013-2014; 2014-2015; 2015-2016; and 2016-2017) were used. Initially, a descriptive and geostatistical analysis of the chemical attributes of the soil was carried out. Subsequently, the application zones were obtained through clustering methods considering the dissimilarity matrix that aggregates information about the Euclidean distance between the sample elements and the spatial dependence structure of the attributes. Subsequently, reduced sample configurations were obtained with 50 and 75% of the initial sample points in these application zones. Afterwards, the descriptive and geostatistical analyzes of the reduced sample configurations were performed again. Finally, the sample configurations (initial and reduced) were compared, by means of the measure of similarity Global Accuracy and the Kappa and Tau concordance indices, in order to determine which configuration provided a better estimation of the variable in unsampled locations. For the crop years under study, the K-means and Ward clustering methods were efficient in defining the application zones, dividing the study area into two or three application zones. Comparing all the reduced sample configurations with the initial one, it was observed that the configuration proportionally reduce and optimized by 25% (selecting 75% of the initial configuration points, which corresponds to 76 sample points) were the most effective in terms of accuracy indices (global accuracy, Kappa, Tau), indicating greater similarity between the thematic maps of these sample configurations. Thus, the reduced sample configurations could be used to generate the application zones, as well as reduce the costs with laboratory analyzes involved in the study.