High-throughput CNN-based phenotyping of grape berry heat stress as an objective indicator of sunburn sensitivity

Abstract

Ongoing climate change poses increasing challenges for viticulture, as higher temperatures, more frequent heat extremes, and intensified ultraviolet (UV) radiation expose grapevines to severe abiotic stress conditions. One critical consequence of these conditions is grape berry sunburn, an abiotic disorder resulting from complex interactions between environmental stressors and plant physiological responses. Sunburn leads to substantial yield losses and reduced berry quality, thereby compromising the economic sustainability of grape production. Reliable phenotyping approaches are therefore essential to objectively evaluate varieties for their risk of developing grape sunburn, enabling meaningful comparisons across genotypes and environments. However, the appearance of natural sunburn symptoms in the field is strongly influenced by environmental variability. To overcome this limitation, we established a controlled laboratory-based heat stress (HS)-treatment of young berries that induces varying degrees of brown lesions. The proportion of lesion area depends on the genotype and is correlated to the sensitivity to grape sunburn in the field. This method was further optimized through the development of a fully automated phenotyping pipeline to enable objective and high-throughput lesion quantification. The pipeline comprises automated acquisition of up to 1,800 high-resolution berry images per hour, followed by automated image analysis using convolutional neural networks (CNNs). Accordingly, a modified DenseNet architecture was trained using image patches extracted from 25 representative single-berry images selected to cover the full sunburn severity range (five images per severity class from low to high). The CNN classified image patches into ‘background’, ‘lesion’, and ‘non-lesion’ categories and achieved an overall classification accuracy of 97.41%. Model performance was evaluated in two independent validation steps: (i) analysis of an additional set of 25 berry images acquired in a different season, and (ii) analysis of berry images from 50 genetically diverse Vitis vinifera cultivars (one image per cultivar). Highly significant correlations (above, r > 0.94) were observed between CNN-based predictions and independent assessments by three expert evaluators across both validation datasets. Furthermore, CNN-based results revealed clear differences in HS sensitivity among varieties, indicating which genotypes are likely to exhibit a higher risk of grape sunburn under field conditions. Overall, the presented methodological framework provides a robust, objective, and scalable solution for grape sunburn phenotyping and generates high quality phenotypic datasets suitable for improved variety characterization, linkage-based QTL mapping, genome-wide association studies, and the training of genomic prediction models.