Mushroom spawn quality classification with machine learning

Abstract

During spawn incubation in mushroom cultivation, spawns can be contaminated by a variety of pests and pathogenic molds, which cause virulent crop damage. Contaminated spawns must be classified and discarded before they are delivered to the fruiting stage. In most mushroom farms, humans visually classify spawn, which is labor-intensive and susceptible to human errors. To solve this problem, we designed a machine learning algorithm to classify oyster mushroom (Pleurotus ostreatus) spawns. Spawn samples were collected from a farm in Thailand. Sample regions of interest of spawns, in polypropylene polypropylene bags, were extracted and filtered to reduce noise. Trivariate histograms of these regions were used as a feature. We analyzed the effects of two techniques, including feature scaling and feature compression, using principal components analysis (PCA) in a pre-processing step. We measured performance of five machine learning classifiers: support vector machines (SVMs), nearest centroid classifier (NCC), k-nearest neighbor (KNN), deep neural network (DNN) and decision trees. Parameters of the methods were optimized and overall performances were compared. Although the number samples obtained was limited and unbalanced, a 4-fold cross validation showed that the DNN classifier had the highest accuracy at 98.8%, with residual variance of 2.5%. Thus, our algorithm can be effectively used to create a model for further application in an embedded system for mushroom spawn classification.