Espectroscopia NIR e algorítmos de seleção de variáveis para determinação de teor de água e classificação de sementes de algodão

Abstract

The differentiation between genetically modified seeds (GMOs) and conventional ones is essential for agro-environmental viability in the different existing crops, as it can avoid contamination of conventional seeds or extinction of native species, elimination of plants and insects, as well as their resistance within the production environment itself. The methods used for the identification of GMOs are based on DNA analysis, which despite being accurate and reliable, are generally expensive, time-consuming and not readily available. On the other hand, spectroscopic techniques, widely used in the agricultural area, for determining quality requirements in seeds, demonstrate better performance for this purpose. The use of algorithms for the selection of variables generates a subset containing the variables that best relate to the properties of interest and contribute to the development of simple and robust models. In recent decades, studies using bioinspired algorithms have been explored in several areas of study, especially in chemistry. Bioinspiration consists of seeking to understand the mechanisms based on the behavior of animal or plant species. Given the above, this work studied the performance of some bioinspired algorithms for selection of variables applied to NIR data for the distinction between conventional and transgenic cotton seeds. Some works have already demonstrated the success of pattern recognition (RP) models built using full NIR spectra to solve this analytical problem. The objective of this work was the application of algorithms to evaluate the predictive capacity of least squares regression models (PLS) built with the whole spectrum and with spectral intervals (iSPA-PLS and i-PLS) for the quantification of water content in cotton seeds. Furthermore, discriminant analysis models constructed with NIR spectral variables selected by variable selection algorithms were constructed and validated. For this, the deterministic algorithm of successive projections (SPA) and also bioinspired algorithms such as ant colony (AOC) and genetic (GA) were tested. The set of samples for the quantification of moisture content was composed of 30 seed genotypes. The data set for the PR methods was composed of 1195 transgenic seeds and 1195 conventional seeds. All samples were provided by EMBRAPA cotton. The predictive capacity of the quantification models was evaluated based on the values of figures of merit such as REP and RMSE. PR models considered sensitivity, selectivity and correct classification rate. The models built with the selected variables obtained comparable results to the models with full spectrum both for the quantification models and for the classification models. However, for the quantification of water content in seeds, iPLS proved to be a more robust model, presenting RMSEP = 0.20% and REP = 2.17%. Regarding the LDA models built with the selected variables, all achieved correct classification rates greater than 95%. The AOC-LDA bioinspired algorithm provided the best performance with 97% correct classification. Given the above, it is concluded that the use of variable selection algorithms combined with multivariate methods of quantification or classification, when applied to NIR data, are efficient. Being able to quantify water content or classify conventional and transgenic cotton seeds quickly and non-destructively.