USO DA MINERAÇÃO DE DADOS PARA EXTRAÇÃO DE CONHECIMENTO AGRONÔMICO ENVOLVENDO O USO DE GESSO AGRÍCOLA

Abstract

The subsoil acidity is harmful towards the growing of plants roots and, consequently, affects the agricultural productivity. In handled areas in the no-till cropping system (SPD), the toxic effects caused by high levels of Al and Mn, due to the soil acidity, are corrected by the superficial liming. This technique improves the acidity of the superficial layers, but it presents no great efficiency in the acidity correction of deeper layers of the soil. The agricultural gypsum (CaSO4.2H2O) is an input that might help in the improvement of the subsoil’s chemical conditions, because besides of being a Ca and S source, it is also able to transport cationic nutrients to the sub superficial layers and reduce the Al activity. It happens, however, that there are questions about which situations may be expected beneficial effects regarding the agricultural gypsum use, and as for the amount that should be applied to reach such effects. A possible form to assist the comprehension of these questions is with the Data Mining (MD) utilization. However, the agronomic databases usually involve a limited number of registers, which difficult the MD use. As a result, this study addresses, beyond the MD utilization, a new research area involving MD in small databases. Therewith, the goals of this work were:(i) obtaining a better comprehension of the gypsum application effects in the chemical attributes of handled soil in SPD, (ii) identifying the chemical attributes of the soil that present narrower correlations with the estimation of the need of gypsum using selection techniques at the pre-processing stage, and (iii) defining models to the estimation of the need of agricultural gypsum in soils under SPD. The database used in this study was obtained from three distinct areas of the region of Campos Gerais do Parana, containing chemical attributes of the soil in different epochs coming from SPD experiments, which received increasing doses of agricultural gypsum on their surfaces. It was used Principal Component Analysis techniques based on B2 and B4 criteria, and also the Supervised ACP technique. Still regarding the pre-processing techniques, it was implemented a covariance matrix that assumes the marginal independence between the base attributes in their calculus and utilizes the B2 and B4 criteria for the attributes selection. For the databases expansion, besides the SMOTE technique, it was implemented the megatrend-diffusion (MTD) method. The M5Rules algorithm was utilized to find models of estimation of the need of agricultural gypsum. The results showed that the elapsed time after the gypsum application (epoch), the saturation through Ca and the saturation through Mg in the capacity of effective cations exchange (CTCe) of soil were the attributes which presented the narrower correlations with the dose estimation of gypsum. The work identified four possible models for the estimation of the need of agricultural gypsum, showing that the M5Rules algorithm was efficient for such prediction. The MTD method presented positive results because increased the correlation coefficient and reduced the average absolute error.