Single shoot detection algorithms represent a promising tool for real-time application of deep learning models. YOLO (You Only Look Once) is a single shoot object detection algorithm that combines fast classification and good accuracy. Over the last few years, several versions of YOLO have been developed, improving its performance. In this study, the last version of YOLO (version 4) was evaluated in its full-size model and in the tiny version, in order to assess grape yield spatial variability. The tiny and full models were previously trained and tested on almost 3000 images collected during several growing stages in different vineyards and varieties. YOLO models were used to classify 24 georeferenced RGB images acquired before the harvesting on an 8-hectare experimental vineyard, where Vitis vinifera cv. Glera vines were trained to Sylvoz and characterized by spatially structured variability. The models were used to detect the number of bunches, based on different resolution images (from 320 up to 1280 pixels) and different confidence thresholds (from 0.25 up to 0.35). The detected number of bunches was then compared with the actual one as well as with the relative final weight harvested from the vines used as target for the collected images. According to preliminary results, the number of bunches detected in high-resolution images exhibited a higher correlation with the number of bunches visible in the images rather than with the final weight. On the other hand, the number of bunches detected in low-resolution images gave evidence of a higher correlation with the total weight of grapes harvested from the target vines. Although high-resolution images allowed the model to detect almost all bunches not covered by leaves, in low-resolution images YOLO models were weakly affected by small bunches, which were rarely detected, thus increasing the correlation with the vines yield. The best linear regression model for vines yield was obtained with 416 pixels images, which showed a coefficient of determination (R2) of 0.59, indicating YOLO as a suitable tool for detecting yield spatial variability. The models used in this work represent non-destructive methodologies for grape yield spatial variability assessment, and they may be easily implemented as on-the-go tools.

Grape yield spatial variability assessment using YOLOv4 object detection algorithm

Sozzi, M.;Cantalamessa, S.;Cogato, A.;Kayad, A.;Marinello, F.
2021

Abstract

Single shoot detection algorithms represent a promising tool for real-time application of deep learning models. YOLO (You Only Look Once) is a single shoot object detection algorithm that combines fast classification and good accuracy. Over the last few years, several versions of YOLO have been developed, improving its performance. In this study, the last version of YOLO (version 4) was evaluated in its full-size model and in the tiny version, in order to assess grape yield spatial variability. The tiny and full models were previously trained and tested on almost 3000 images collected during several growing stages in different vineyards and varieties. YOLO models were used to classify 24 georeferenced RGB images acquired before the harvesting on an 8-hectare experimental vineyard, where Vitis vinifera cv. Glera vines were trained to Sylvoz and characterized by spatially structured variability. The models were used to detect the number of bunches, based on different resolution images (from 320 up to 1280 pixels) and different confidence thresholds (from 0.25 up to 0.35). The detected number of bunches was then compared with the actual one as well as with the relative final weight harvested from the vines used as target for the collected images. According to preliminary results, the number of bunches detected in high-resolution images exhibited a higher correlation with the number of bunches visible in the images rather than with the final weight. On the other hand, the number of bunches detected in low-resolution images gave evidence of a higher correlation with the total weight of grapes harvested from the target vines. Although high-resolution images allowed the model to detect almost all bunches not covered by leaves, in low-resolution images YOLO models were weakly affected by small bunches, which were rarely detected, thus increasing the correlation with the vines yield. The best linear regression model for vines yield was obtained with 416 pixels images, which showed a coefficient of determination (R2) of 0.59, indicating YOLO as a suitable tool for detecting yield spatial variability. The models used in this work represent non-destructive methodologies for grape yield spatial variability assessment, and they may be easily implemented as on-the-go tools.
2021
Precision Agriculture '21
978-90-8686-916-9
978-90-8686-363-1
File in questo prodotto:
File Dimensione Formato  
Paper.pdf

non disponibili

Descrizione: Articolo
Tipologia: Published (publisher's version)
Licenza: Accesso privato - non pubblico
Dimensione 1.89 MB
Formato Adobe PDF
1.89 MB Adobe PDF Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11577/3398210
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? 9
social impact