Implementation of deep learning algorithms

The research team took photographs of some 2.000 strawberries, both exterior and half-cut, from different improvement lines provided by the Planasa company and collected them in the 2018 campaign in Huelva (Spain). “Evaluating the shape of a given object, a strawberry in this case, from a photograph is not as simple as it seems. The classical linear descriptors - area, perimeter, height, width ... - have certain limitations that lead to the loss of relevant information by extremely simplifying the morphological characteristics. To evaluate the shape in a more detailed way, we complement these linear methods with multivariate and deep learning techniques ”, stated the first author of the article, Laura M. Zingaretti, who carried out this work as part of her doctoral thesis at CRAG.

By combining these methods, the research team was able to generate automated software that analyzes shape and color patterns extracted from strawberry images. The tool they developed is much more automated than its predecessors, as it requires minimal user intervention and limited computing time, providing a fast and inexpensive way to evaluate phenomics.

A tool to improve agricultural efficiency

“In addition to morphological analysis, our deep learning tool is capable of predicting the appearance of new fruit crosses. This contribution can be very valuable in the first step of plant improvement programs, since it would allow evaluating several crosses without the need to test them directly in the field, saving time and resources ”, stated Miguel Pérez-Enciso, ICREA researcher at CRAG and co-director of the study.

"This tool also has the potential to be adapted to measure the visual phenomic traits of fruits directly in the field, to analyze other characteristics of plant conformation (leaves, flowers, roots ...), or for the early evaluation of diseases," he stated Amparo Monfort, researcher at IRTA of the CRAG and co-director of the work.