Combining computer vision and deep learning to classify varieties of Prunus dulcis for the nursery plant industry

Abstract

Varietal control to avoid unwanted varietal mixtures is an important objective for the nursery plant industry. In this study, we have developed and analyzed the capabilities of a computer vision system based on deep learning for the control of plant varieties in the nursery plant industry and for evaluating its capabilities. For this purpose, three datasets of nursery plant images were compared. The datasets came from two varieties of almond trees (Prunus dulcis) named Soleta and Pentacebas. Each dataset contained images with three different scales: whole plant, leaf, and venation. The Gradient-weighted Class Activation Mapping (Grad-CAM) technique was used to unveil the most important features to discriminate between both varieties. The three datasets provided classification accuracies above 97% in the test set, being the leaf dataset, with a 98.8% accuracy, the one providing the best results. Concerning the most important features of the plants, the Grad-CAM showed that they are located in the center of the leaf, that is, the venation. In conclusion, we have shown that computer vision is a promising technique for the control of plant varietal mixtures.