Weed Detection in Grassland and Field Areas Employing RGB Imagery with a Deep Learning Algorithm Using <i>Rumex obtusifolius</i> Plants as a Case Study

Abstract

The bluntleaf dock/ broad-leaved dock (Rumex obtusifolius) is a fast growing, highly competitive and resistant weed. It is endemic to Austria and generally a very common weed in Europe. Rumex obtusifolius prefers nutrient-rich, moist soils. As a light germinator, it spreads easily in patchy plant stands. Its taproot can penetrate compacted, waterlogged and oxygen-poor soil layers to a depth of 2.60 m. It is considered a pest in agriculture, both in field and pasture, because of its rapid growth, ability to vegetatively propagate from leftover roots and its extensive taproot system. The most important management strategy is to prevent dock plants from establishing. If plants are already present in the field, the population must be assessed. If there are up to two dock plants per square meter, single-stock measures such as pricking out or tilling and reseeding are used. If there are more than two plants per square meter, uprooting will help. Furthermore, it will become necessary to adjust the crop rotation. The application of pesticides is possible; however, mechanical removal is preferred. The goal of this study is to develop a CNN (convolutional neural network) that is specially trained to identify dock plants and to capture location and position in the field/pasture. RGB photographs (n = 2500) were collected using an unmanned aerial vehicle and handheld cameras from March to August 2021. The obtained dataset contained photographs showcasing dock plants in all sizes and forms to include different phenotypes and age difference. The network was also trained to differentiate between whole plants and plant parts such as leaves.

Keywords

• AI
• neural network
• weed species detection
• deep learning
• RGB
• grassland