Smart and accurate: A new tool to identify stressed soybean seeds based on multispectral images and machine learning models

• Multispectral images, captured by different sensors and submitted to machine learning, are used to map and segment physical and physiological properties of stressed soybean seeds. • Chlorophyll a fluorescence images are a strong marker of seed stress. • A new method has been discovered to identify seeds produced under environmental stress. Extreme environmental conditions have been recurrent during the last few years and have impacted crop seed quality worldwide, mainly but not limited to, soybeans ( Glycine max (L) Merrill). To overcome this, seed companies often demand innovative tools to address seed quality factors. Machine learning models based on multispectral imaging are a novel seed quality analysis approach. Thus, we hypothesize that segmenting stressed (those produced under conditions that are not favorable to the mother-plant) and non-stressed (produced under conditions favorable to the mother-plant) soybean seeds would be possible with this technology, opening a new opportunity for seed quality management and elucidating quality factors. Soybean seeds (cultivar BR/MG 46-Conquista) were produced under water deficit and heat during maturation (from R5.5 onwards). Multispectral images were acquired from stressed and non-stressed seeds, and the reflectance, autofluorescence, physical properties, and chlorophyll parameters were extracted from the images. In parallel, we determined seed vigor. We designed machine learning models using multispectral imaging data based on three algorithms: neural network, support vector machine, and random forest. Our results demonstrated that the stressed seeds have spectral markers that enable their recognition. Concomitantly, these markers had a direct relationship with seed vigor. The machine learning models developed based on neural network algorithm showed the highest performance in segmenting stressed seeds (≥90% of accuracy, precision, recall, specificity and F1 score) in contrast to random forest- and support vector machine algorithm (≥88% of accuracy, precision, recall, specificity and F1 score). Here, we report a new approach for multispectral imaging with the potential to identify soybean seeds of lower vigor as a result of unfavorable environmental conditions during seed maturation.