Moving beyond visible flower counting: RGB image-based flower number and yield prediction in grapevine

Abstract

Accurate yield estimation is crucial for optimizing vineyard management and logistical organization. Traditional methods relying on manual and destructive flower or berry counts are labor-intensive and unsuitable for large-scale applications. To address these limitations, this study evaluated the potential of RGB image analysis for non-destructive estimation of flower number in grapevine inflorescences and its relationship with yield components. The objectives were: (i) to correlate the number of actual flowers with the number of pixels in inflorescence images considering the projected area and not visible flower count; (ii) to assess the model’s capacity for predicting yield components such as the number of berries and bunch weight.

The study was conducted during the 2023 vintage in a cv Catarratto/1103P vineyard located in Sicily, on 36 vines divided into two vigor classes (high and low). The vines were trained on a vertical shoot positioning trellis system with a bilateral cordon spur pruned. All the inflorescences, categorized into six length classes, were photographed using a smartphone and white cardboard under field conditions. Images were analyzed using FIJI/ImageJ© software, where inflorescences were segmented via Otsu’s method, and pixel counts were used to estimate flower numbers through linear regression. Flower counts were validated through manual counts of detached calyptra. Yield components were assessed by harvesting the vines, measuring bunch weight, and counting berries to calculate fruit set percentage. External model validation was performed on datasets from Catarratto, Chardonnay, and Vermentino cvs.

Results showed a strong correlation between inflorescence pixel count and flower number both on single inflorescences (R² = 0.85; MAPE 29%; n = 300) and in terms of total flowers per vine (R² = 0.95; MAPE 26%), with estimation accuracy unaffected by vine vigor class but varying slightly by inflorescence length class. External validation yielded a MAPE of 20% in Catarratto, 40% in Chardonnay and 34% in Vermentino. The model reliably estimated also the number of berries (R² = 0.94; MAPE 13%) and bunch weight (R² = 0.78; MAPE 20%) per vine. While environmental factors such as fruit set can affect yield, this study highlights the potential for early yield prediction with image analysis. The methodology holds promise for scalable application and integration into vineyard management technologies in vine and table grapes.