Format: ORAL

Maria A. R. Ferreira, Raquel A. Gomes, Jamille C. Silva, Jailton J. Silva, Williane A. S. Oliveira, Marcos V. C. Freitas, Srgio T. Freitas, Brbara F. Dantas

Embrapa Semiárido, Petrolina, Brazil

Seed-based conservation plays a critical role in the conservation, enrichment and restoration of degraded environments. Climate change and agricultural pressure has enhanced the vulnerability to desertification of tropical dry forests. Plant biodiversity conservation is urgent, mainly in populations located in climate change threatening hotspots. Traditional stored seed quality assessment relies on destructive tests to evaluate viability, germination and moisture content. These usually take more time and seeds than is convenient for restoration actions. Quick and accurate non-destructive methods allow efficient decision making in all native seed conservation activities, from harvest to storage, and also in seed-based restoration actions. Spectrometry allied to machine learning can be an effective method for evaluating the quality of native forest seeds, while maintaining the usually small bulk of stored seeds. Thus, we aimed to develop an easy and non-destructive method for evaluating stored seeds physiological quality and moisture content using a portable Vis-NIR spectrometer and a machine learning algorithm model. Stored and fresh seeds of different populations of two dry forest Fabaceae were individually evaluated for germination, moisture content and spectral data using a portable spectrometer. External and internal validation of the model were carried out by dividing the data into training (70%) and testing (30%) stages, using 10x cross-validation. The spectral data were processed using Weka 3.8.6 software. The discriminative algorithms applied were Support Vector Machine, Multilayer Perceptron, Random Forest and J48. For germination, the seeds were classified as YES (germinated) and NO (not germinated). Multilayer Perceptron algorithm obtained the best results for water content evaluation, with calibration and prediction correlation coefficients above 0.70. Random Forest and J48 algorithms showed best performances for seed germination classification. Although the algorithms were efficient in seed quality classification, increasing the number of sampled seeds may improve their prediction accuracy and aid seed-based conservation activities.