Redução da dimensionalidade para estimativa de teores de nutrientes em folhas e grãos de soja com espectroscopia no infravermelho

Abstract

The high dimensionality in databases is a problem that can occur in several fields, including the plants nutrients state analysis. These analyses are currently based on methodologies that spend time and reagents. (NIR-NearInfrared) and (MIR-MiddleInfrared) spectroscopy have been shown to be a faster and clean alternative to simultaneous quantification of compounds. Since reading occurs at wavelengths generating hundreds attributes for the NIR and thousands to the MIR the data obtained by such equipment have a high dimensionality. One of the difficulties is to identify which attributes are more relevant for the nutrient analysis. This work aimed to verify the correlation gain obtained with the use of dimensionality reduction techniques with data obtained by NIR and MIR spectroscopy. The goal is to estimated levels of 11 nutrients in grains and leaves of soybean: Nitrogen (N), Phosphorus (P), Potassium (K), Calcium (Ca), Magnesium (Mg), Sulfur (S), Copper (Cu), Manganese (Mn), Iron (Fe), Zinc (Zn) and Boron (B). For that, 231 soybean leaves and 285 soybeans samples were analysed by spectroscopy in the mid-infrared and nearinfrared region. The regression models were generated by machine learning algorithms: SMOReg which implements the support vector machine for regression; M5Rules that is based on decision trees with regression functions; and LinearRegression algorithm for linear regression. The results were evaluated by correlation coefficient (r) and the quadratic error (RRSE). Estimating leaf nutrients was satisfactory for both NIR and MIR spectroscopy, where correlations of 0.80 above were obtained for P, K, Mg, S, Mn, Cu, Fe and Zn. There were no correlations for B and Ca in soybean leaves. Estimating nutrient was also satisfactory for soybeans, but only in NIR spectroscopy data, where correlations above 0.7 were obtained for N, P, K, Ca, and S. Using dimensionality reduction techniques provided the high values for correlation of P, K, and S in soybean leaves, making use of the LinearRegression algorithm. For soybeans, the dimensionality reduction was essential in obtaining satisfactory correlations, except for N, always using the LinearRegression algorithm. When reducing the dimensionality was not used, satisfactory results were obtained by the SMOREg algorithm from foliar data to N, Mg, Cu, Mn, Fe, and Zn. Reducing dimensionality associated to the use of LinearRegression algorithm resulted in better correlations for three nutrients in leaves and satisfactory rates of grain. The observed results demonstrate a greater efficiency in the use of the NIR for foliar analysis than for grain analysis. SMOReg computational techniques and LinearRegression algorithm presented the best results, being the SMOReg indicated for large quantities of attributes and Linear- Regression for smaller quantities