Development of Detection Method to Docks (Rumex spp.) in a Temperate Grass Meadow Using Aerial Image and a Convolutional Neural Network

Abstract

Grassland management often makes it easy for invasive weeds to grow without renovation of the grassland. The objective of this study is to develop a software that can automatically detect docks (Rumex spp.) —a major grassland weed— from aerial images captured by unmanned aerial vehicles, by using a convolutional neural network. Images of a grassland (8,240m²) were captured by unmanned aerial vehicles from an altitude of 10m. Patches of docks and patches of background were extracted from the image to establish training and validation data sets. A model for the identification of ducks was created in the convolutional neural network. The software was developed using the Python scripting language. To evaluate the detection accuracy of the developed software, each of the 12 aerial images taken during three different harvesting periods was analyzed. The detection accuracy for each period, compared to manual detection, was 85.4%, 78.6%, and 93.0%.