Automated Debris Mass Estimation for Citrus Mechanical Harvesting Systems Using Machine Vision

Abstract

An algorithm for automatically estimating mass of debris in a citrus canopy shake and catch harvester using machine vision was investigated. Debris material consists of non-citrus objects such as leaves, twigs, and branches which are mechanically harvested along with citrus fruit. While harvesting, the amount of debris is surprisingly large even in the cleanest loads. It is mandatory for the juice extraction plants to separate this debris from fruit and this process has a huge economic impact. Current situation with catastrophic diseases such as citrus greening, a bacterial plant disease called Huanglongbing (HLB), and citrus canker infestation in Florida requires citrus growers to dispose of leaves and twigs very carefully in the grove. Although not mandated, the first step to treat such diseased vegetation is to determine how much debris material is generated during harvesting. So an efficient method to estimate debris mass is required to assist debris separation mechanism. A hardware/software system to automatically estimate debris mass was developed for a citrus canopy shake and catch harvester. The idea was to estimate the debris mass from the total pixel area of the debris detected using image processing. An experimental test bench was set-up at a machine shop in the University of Florida, Gainesville, Florida, to train as well as validate the image processing algorithm. This algorithm included steps such as image rectification, overlapped area removal, morphological operations, and removal of undesired debris on the ground using a novel Parse and Add algorithm. The results showed that the coefficient of determination (R2) between the pixel area and debris mass for calibration was 0.946 and R2 between actual and estimated mass for validation set from the test bench was 0.815 with an RMSE of 1.88 kg. For the field experiment, over 27,000 images were acquired from a commercial citrus grove during harvesting in Ft. Basinger, Florida, in 2008. These images were post-processed with the algorithm and the debris objects were identified from a representative set of 180 images. The R2 between the actual and estimated debris mass for individual images was 0.78, and the RMSE was 0.02 kg. The error between total actual and estimated mass for the field experiment was 25.3%. The debris mass estimates were also correlated with GPS data to create a geo-referenced map of the debris gathered. The developed debris mass estimation system could play a crucial role in solving the problem of safe and economical disposal of diseased leaves and twigs.