Automatic image analysis and spot classification for detection of fruit fly infestation in hyperspectral images of mangoes

Fruit fly infestation of mangos is a major concern for growers and exporters, leading to requirements for quarantine treatments such as vapor heat treatment or irradiation and subsequent reduction in quality and consumer acceptance. An on-line method for detection and removal of infested fruit would thus benefit producers and consumers. An algorithm has been developed to identify spots generated in hyperspectral images of mangoes infested with fruit fly larvae. The algorithm incorporates background removal, application of a Gaussian blur, thresholding, and particle count analysis to identify locations of infestations. Each of the four algorithm steps involves adjustable parameters which were iteratively tested to find the optimal combination for detection in terms of false positive and false negative results. For algorithm parameters selected to minimize false negative results, a false negative error rate of 1.0% was achieved with 11.1% false positive error and 6.0% overall error in heavily infested samples. For the same sample set, the lowest overall error rate achieved was 2.0%, with 1.0% false positive and 3.0% false negative. For samples with lower infestation rates, the error rates were much higher, the lowest overall error being 12.3%. This therefore demonstrates the feasibility of hyperspectral imaging for fruit fly detection while highlighting the need for technology with improved resolution and signal to noise ratio to allow detection of single larvae.