Inteligência computacional aplicada ao estudo da divergência e fenotipagem em cultivares de soja

Abstract

The artificial neural networks are computational models of the human brain that recognize patterns and regularities of the data and represent an alternative as a universal approximation of complex functions. They may outperform conventional statistical models, with the advantage of being non-parametric, not requiring detailed information on the physical processes of the system being modeled, and tolerating data loss. The computational intelligence has great potential, it is already widely consolidated in the computational areas and becomes an interesting approach to be used in the field of plant breeding. Thus, in the first article, the objective was to study the genetic dissimilarity between soybean cultivars, using SOM type neural networks, and to test the efficiency of the methodology used, through Anderson's discriminant analysis. In order to select the best network topology, different network architectures were tested for the highest average hit rate, Anderson discriminant analysis, and the lowest number of empty clusters, with 1000 trainings performed. After that, 1000 new net trainings were performed and, later, a matrix of dissimilarity was constructed, observing the frequency with which the genotypes were considered as being of different groups. The consistency of the clustering was verified through Anderson's discriminant analysis. Six groups were formed by the UPGMA method, indicating the existence of variability for the evaluated characteristics. Through Anderson's discriminant analysis, 100% of hits were correctly classified. Therefore, the SOM networks are efficient for the grouping of soybean strains, showing the existence of genetic dissimilarity and the formation of six clusters among the studied strains. The second article had as objective to obtain methodologies that allow estimating the leaf area in soybean, through the use of artificial neural networks, considering different leaf formats. Thirty-six cultivars were evaluated, which were divided into three groups, according the leaf format. Multilayer perceptrons were developed, using 300 sheets per group, with 70% for training and 30% for validation. The most important morphological measures were also tested in these studies. The RNAs were efficient to estimate the leaf area in soybean, with determination coefficients close to 0.90. Only two foliar measurements of leaflets are sufficient for estimation of leaf area. The network 4, trained with leaves of all groups, was more general and, consequently, is the most indicated for the prediction of the leaf area in soybean.