Abstract
The estimation of the ripening state in orchards helps improve post-harvest processes. Picking fruits based on their stage of maturity can reduce the cost of storage and increase market outcomes. Moreover, aerial images and the estimated ripeness can be used as indicators for detecting water stress and determining the water applied during irrigation. Additionally, they can also be related to the crop coefficient (Kc) of seasonal water needs. The purpose of this research is to develop a new computer vision algorithm to detect the existing fruits in aerial images of an apple cultivar (of Red Delicious variety) and estimate their ripeness stage among four possible classes: unripe, half-ripe, ripe, and overripe. The proposed method is based on a combination of the most effective color features and a classifier based on artificial neural networks optimized with genetic algorithms. The obtained results indicate an average classification accuracy of 97.88%, over a dataset of 8390 images and 27,687 apples, and values of the area under the ROC (receiver operating characteristic) curve near or above 0.99 for all classes. We believe this is a remarkable performance that allows a proper non-intrusive estimation of ripening that will help to improve harvesting strategies.
Highlights
Non-destructive estimation of fruit ripeness in orchards is one of the most important challenges for agronomical engineering researchers worldwide, since many management decisions can be directly related with the maturity state of fruits
A set of color features is extracted from each segmented object, based on a prior study of the most discriminant features
The experiments to assess the effectiveness of the proposed method are divided into the different steps of the process
Summary
Non-destructive estimation of fruit ripeness in orchards is one of the most important challenges for agronomical engineering researchers worldwide, since many management decisions can be directly related with the maturity state of fruits. It can be observed that 98.51% of all samples belonging to the apple class and 98.31% of the background objects were detected correctly, producing a total accuracy of 98.43% in the segmentation process This is a very remarkable result, allowing a potential use of the algorithm in practice under real environment conditions. These results are slightly worse than those reported in [25], with a 99.12% accuracy, they are still better than other state-of-the-art methods, as applied in [34,35], which achieved 92.5% and 95.72%, respectively, in segmentation of different agricultural products. Most Effective Color Features Using the Hybrid ANN-SA Method
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call