Image analysis as a potential tool for marker-assisted selection

Abstract

The recent rapid changes in climate due to global warming have increased the frequency of severe natural disasters. Such disasters easily damage the fruits with large weight and volume, inflicting a great loss on the farms. Although the central and regional governments have persuaded farm owners to obtain the Crop Yield and Revenue Insurance, the lack of objective criteria of damage analysis have often prevented an actual compensation. It has also been difficult to attain an accurate statistics of the national fruit production in South Korea, as the data collection through interviews and sample analyses was based on the manpower at the city and county agricultural technology centers. This study developed a deep learning model of citrus fruit detection to be used in the yield estimation and damage analysis for the Crop Yield and Revenue Insurance. The model was based on the YOLOv5 algorithm, which allows the fruit number to be estimated using images. The model showed an outstanding detection performance at AP₅₀ 0.817. This image-based deep learning model can also be widely applied in breeding. Notably, in breeding programs focused on increasing the production, the image-based high-throughput phenotyping could readily determine the fruit production per line. In the future, models for the detection of other fruit crops, including apples, and a smartphone application for the Crop Yield and Revenue Insurance and fruit production estimation will be developed.