Machine vision system for early yield estimation of citrus in a site-specific manner

Abstract. Detecting immature green citrus at an early stage for yield forecasting can help growers expect how many fruit they can harvest at the end of the year. Also, it can provide an in-field spatial variability of fruit that can be used for providing site-specific management of citrus trees to increase yield and profit. Yield forecasting using the machine vision technology has been a promising but challenging research area. In this research, an algorithm using 3D geometric features was developed to identify immature green citrus far in advance of harvesting. The specific objectives were to develop a robust machine vision algorithm for images that had more complicated backgrounds than previous studies and to build an algorithm to have stable performances in varying illumination conditions. The machine vision algorithm was fully automatic and consisted of three steps, 1) illumination enhancement of the RGB image, 2) potential area finding using depth information and 3) classification of fruit and background. A total of 93 images were acquired and used to develop and validate the algorithm. The final result was analyzed using the correctly identified fruit and false positives. Average correct identification rate and false positive rate were 72.1% and 23.2%, respectively.