Rapid tomato DUS trait analysis using an optimized mobile-based coarse-to-fine instance segmentation algorithm

As climate change continues to impact agriculture, there is a growing demand for the discovery of new crop varieties in order to address key goals such as accelerated production, disease resistance, and overall improved quality. One of the necessary procedures before a crop variety is accepted for production is distinctness, uniformity, and stability (DUS) testing. However, the current practice of DUS testing relies primarily on manual examination with limited technological assistance. This work aims to provide a solution to this challenge by developing an algorithm for rapid tomato DUS trait analysis using a mobile application. An image dataset comprised of tray and individual tomato images was compiled using multiple mobile devices. A coarseto-fine instance segmentation algorithm was developed to analyze the tray images by detecting individual tomato images and detecting tomato and peduncle scar contours. In order to accommodate different mobile devices and achieve finer measurements, a conditional upscaling approach was applied on each individual tomato image, with the support of super-resolution. Android ARCore was utilized to obtain distances of each tomato from the mobile device camera, enabling fast morphological measurements without using reference scales. The proposed algorithm has a precision of 0.99 in detecting each tomato from each tray image, while having IoUseg values of 0.97 and 0.83 in segmenting tomato and peduncle scars, respectively. Manual vs. automated trait analysis results also show that the mobile application was able to measure traits with an error from 1.66% to 7.19%. From the best of our knowledge, this work presents one of the first mobile phone applications for rapid tomato DUS trait analysis.