APLICAÇÃO DE ANÁLISE MULTIVARIADA E VISÃO COMPUTACIONAL A PÓLEN DE Apis mellifera E ABELHAS SEM FERRÃO

Abstract

Bee pollen has stood out due to its nutritional and functional characteristics, being considered an excellent dietary supplement with possible commercial use as an ingredient in food formulations, cosmetics, and biomedical applications. However, its utilization on an industrial level is still limited due to variability in its botanical, geographical, and physicochemical composition. Therefore, it is essential to understand aspects related to its collection in order to enhance its application as an ingredient or simply to offer consumers a product with less variable composition. Hence, it becomes essential to identify factors affecting the botanical composition, such as foraging behavior of different bee species and seasonality, which can influence the coloration and chemical composition of bee pollen to ensure its nutritional quality. In this thesis, multivariate analysis and computer vision techniques were applied to data on pollen analysis and the physicochemical composition of Apis mellifera and stingless bee pollen. Network analysis demonstrated that stingless bees Melipona quadrifasciata, Melipona marginata, and Scaptotrigona bipunctata, due to seasonal variations and abiotic factors, collected monofloral pollen samples from plants of the Myrtaceae, Euphorbiaceae, and Fabaceae families. Bees Apis mellifera and Tetragona clavipes exhibited a more generalized pollen foraging behavior, collecting heterofloral samples in most seasons of the year. Results from the Generalized Linear Models showed that the 'bee species' and 'season' factors affect the quantity and types of pollen collected by bees. To assess how the foraging behavior of different bee species and seasonality can affect the coloration of bee pollen, a nested factors ANOVA was applied to instrumental color data from samples collected by Apis mellifera and stingless bees throughout four seasons of the year. ANOVA results demonstrated that the coloration of bee pollen is primarily influenced by the bee species, but it can also be influenced by seasonality and the characteristic flora of the producing region. Furthermore, significant correlations were observed between the L*, a*, and b* color coordinates, which could be associated with the predominant contribution of specific plant species due to the availability of flora in the producing region and foraging habits of bee species. This study also explored the potential of combining digital image processing with machine learning for the development of a rapid, cost-effective, and environmentally friendly analytical methodology for determining crude protein contents in bee pollen. Digital images of bee pollen samples were obtained using a smartphone camera with controlled lighting, and RGB channels and color histograms were extracted using specific open-source software. Crude protein content was determined using the standard Kjeldahl method, and in combination with digital image informations, it was used to generate a predictive model through the Random Forest algorithm, which exhibited good performance and predictive ability (R2 = 80.93 %; RMSE (root mean squared error) = 1.49 %; MAE (mean absolut error) = 1.26%). The developed analytical methodology for crude protein analysis can be considered environmentally friendly, serving as an excellent alternative to conventional analysis methods as it avoids the use of toxic reagents and solvents, is energy-efficient, employs low-cost instrumentation, and is robust and precise. The results achieved in this thesis demonstrate the importance of recognizing the characteristics of the production region (climatic conditions, available plants for bees, type and quantity of flowers) and the pollen foraging behavior of different bee species. These informations can be used to establish best bee management practices for obtaining monofloral and heterofloral bee pollen samples with more uniform coloration, contributing to the development of national beekeeping and meliponiculture. Moreover, the development and validation of the methodology for crude protein analysis combining digital image processing and machine learning can be easily implemented in routine analyses in quality control laboratories for products derived from bees, ensuring food quality and consequently supporting future commercial applications of bee pollen.