Comparison of different computer vision methods for vineyard canopy detection using UAV multispectral images

• Published in: Computers and electronics in agriculture 2024-10, Vol.225, p.109277, Article 109277
• Main Authors: Ferro, Massimo Vincenzo, Sørensen, Claus Grøn, Catania, Pietro
• Format: Article
• Language: English
• Subjects: Deep learning; NDVI; OBIA; Unmanned aerial vehicles; Vitis vinifera L
• ISSN: 0168-1699
• DOI: 10.1016/j.compag.2024.109277

•Performance evaluation of vine canopy detection methods using multispectral images.•Deep learning methods achieved the best performance in vineyard canopy detection.•U-Net and Mask R-CNN provide more accurate monitoring to assess vineyard vigor.•Canopy NDVI varies as a function of image analysis techniques applied. In viticulture, the rapid and accurate acquisition of canopy spectral information through ultra-high spatial resolution imagery is increasingly demanded for decision support. The prevalent practice involves creating vigor maps using spectral data obtained from pure vine canopy pixels. Object-Based Image Analysis (OBIA) among conventional methods exhibits a reasonable efficiency in canopy classification due to its feature extraction capabilities. In recent years, deep learning (DL) techniques have demonstrated significant potential in orchard monitoring, leveraging their ability to automatically learn image features. This study assessed the performance of different methodologies, including Mask R-CNN, U-Net, OBIA and unsupervised methods, in identifying pure canopy pixels. The effectiveness of shadow and background detection methods and the impact of misclassified pixels on NDVI were compared. Results were compared with agronomic surveys conducted during the 2021 and 2022 growing seasons, focusing on two distinct phenological stages (BBCH65-BBCH85). Mask R-CNN and U-Net exhibited superior performance in terms of Overall Accuracy (OA), F1-score, and Intersection Over Union (IoU). Among OBIA methods, the Gaussian Mixture Model (GMM) proved to be the most effective classifier for canopy segmentation, and Support Vector Machine (SVM) also demonstrated reasonable stability. Conversely, Random Forest (RF) and K-Means yielded lower accuracy and higher error rates. As a result of the limited accuracy, it is noted for vineyard rows with low vigor canopies that NDVI was overestimated, while for high vigor canopies NDVI was underestimated. Significantly improved determination coefficients were observed for the comparison between Total Leaf Area (TLA) and NDVI data derived from Mask R-CNN and U-Net. Positive correlations were also found with NDVI data from GMM and SVM algorithms. Regarding leaf chlorophyll (Chl) and NDVI correlations, Mask R-CNN and U-Net methods showed superior performance. Additionally, the relationship between TLA and projected canopy area (PCA) was significantly better represented by U-Net and Mask R-CNN, while PCA was not recommended f