Avaliação e recomendação de espectrômetro-Vis-NIR portátil para o monitoramento da qualidade fisiológica de sementes florestais nativas

Abstract

The use of visible and near-infrared (Vis-NIR) spectroscopy to assess seed quality allows an initial analysis of seed characteristics to be obtained before and during seed collection, avoiding the destruction of the material as occurs in the standard method. Vis-NIR, combined with machine learning (ML) techniques, contributes to the development of models that allow the assessment of seed quality of native forest species of the Caatinga biome. This study aimed to develop accurate calibration models with machine learning algorithms to assess the seed quality of two native forest legumes of the Caatinga biome, using a portable Vis-NIR spectrometer. Sublots of seeds with different water contents (WC) were formed from the original lots, adjusting the content through controlled hydration or dehydration. The seeds were placed on a screen in plastic boxes type "gerbox" with 40 mL of water (hydration) or silica gel (dehydration) for periods of 24 or 72 hours, generating sublots with different WC. The spectral evaluation of the WC of the seeds (WCNIR) was performed by individual reading with the portable Vis-NIR spectrometer. After collecting the spectral data, the reference WC (WCREF) was measured with a traditional method adapted for individual seeds. The original spectral data were used, and a sampling filter was applied to adjust the models. In the second study, the spectral readings of the seeds were followed by the germination test in both species, with 25 seeds and four replicates per lot. After germination, the seeds were classified according to their germination capacity and seedling formation (normal, abnormal or without formation). To determine the vigor, the readings were also performed individually, with the seeds incubated in B.O.D. at 30 °C, with vigor classified as high, low or null. Four discriminative algorithms were evaluated: Multilayer Perceptron (MP), Random Forest (RF), J48 and Support Vector Machine. The Gaussian Processes algorithm was used only in the first study, while J48 was applied exclusively in the second study. MP stood out in predicting the WC of Anadenanthera colubrina and Cenostigma pyramidale seeds with the original data. On the other hand, RF obtained better results after applying the sampling filter. In the second study, for A. colubrina and C. pyramidale, the Random Forest (RF) and J48 algorithms presented the best results in the classification of germination, seedlings and vigor. These methods stood out in classifying germinated seeds, those that did not produce seedlings and those with high vigor, with the highest accuracy rates in both species. The combination of Vis-NIR with machine learning proved to be effective in discriminating between germination and vigor classes, although it is necessary to include more data to improve the accuracy of the models.