Study of a Method for Recognizing Field Covering Situation by Applying Semantic Segmentation to RGB Images in Synecoculture Environment

Abstract

Synecoculture™ is a method of farming that produces useful plants while making multifaceted use of the self-organizing ability of the ecosystem by growing a wide variety of plants densely mixed in the same farmland. As a technology to support Synecoculture, robotics are being developed to automate major management tasks. Still, the complexity of recognition and operation is imposing a heavy burden against automation compared with conventional farming that is based on a uniform operation of a single plant. In Synecoculture, it is essential to cover the topsoil with vegetation. If the topsoil is exposed, it is necessary to introduce seeds and seedlings to fill the gap with vegetation. In this study, we aim to recognize the area of the bare soil surface with pixel-wise precision. In the proposed method, each pixel segments into two classes: “vegetation” or “no vegetation.” by applying semantic segmentation to RGB images with the Focal Loss function. By comparing accuracy with different values of parameters for the semantic segmentation, our approach showed that this method could achieve high accuracy with a relatively small number of images for training.