Spotibot: Rapid scoring of B otrytis lesions on rose petals using deep learning and mobile computing

Roses are renowned for their ornamental value and are available in a wide range of colors and shapes due to extensive breeding and ease of hybridization. During post-harvest, roses are highly susceptible to fungal decay by the grey mould fungus Botrytis cinerea. No complete resistance to Botrytis is known, and several studies indicate a quantitative nature of resistance. This implies that multiple genes are involved, and that each contribution may only have a slight effect on resistance. Accurate, fast, and objective phenotyping discriminating between minor effects would be essential for breeding selections and discovering novel resistance- or susceptibility genes against Botrytis. Spotibot, a phenotyping software available both as a web application and mobile application, utilizes deep learning and mobile computing for automatically detecting Botrytis lesions on rose petals making it highly applicable for breeding selection. The algorithm can measure petal area (mm²), lesion area (mm²), lesion diameter (mm) and lesion to petal ratio. The deep learning-based algorithm features a coarse-to-fine segmentation approach using two instance segmentation models. The first model (F ₁-score ​= ​0.99) detects and segments each petal, while the second model (F ₁-score ​= ​0.96) detects and segments Botrytis lesions on each petal. Spearman Rank correlation analysis showed a high near-monotonic relationship between human-assessed subjective scores and the objective data generated using Spotibot. An analysis of variance indicated that objective variables reveal more and stronger differences between rose genotypes than using subjective data alone. This is the first work on developing a fast and user-friendly application for image analysis of rose petals to screen Botrytis resistance and susceptibility.