Application of machine learning techniques for identifying soil-dwelling fungi and Chromista

Abstract The main goal of this paper is to construct automated system for accurate identification of soil microorganisms on a genera level based on microscopic images of the monocultural colonies. The microorganisms in question belong to one of the following genera: Fusarium, Trichoderma, Verticillium or Phytophthora. Proposed classification system is fully automated and works on raw microscopic images without extra preprocessing such as marking indicators or coloration applied. The implemented scheme comprises of the following steps: image preprocessing to enhance microorganisms characteristic traits, image segmentation to distinguish background from region of interest, features calculation focused on color and texture of image, feature selection and lastly classification using machine learning methods. Due to a differentiation of analyzed microorganism structures a novel method of segmentation, consisting of several steps and combining various machine graphics methods, that maximizes classification accuracy is developed. This work reports on research examining the impact of selected classifiers on final microorganisms’ identification. The highest noted mean classification accuracy renders 91.73% across 50runs of 10% cross-validation for Regularized Extreme Machine Learning classifier tested on a full set of features with parameter values set to 3900 neurons with Leaky Rectified Linear activation function in hidden layer and regularization constant C = 2.8.