Technical Yearbook 2024

JANUARY/FEBRUARY

Mapping individual vines using drone imagery By Matthew Wrensch & Kyle Loggenberg This study proposes a new method of creating precision input data for precision viticulture and autonomous farming applications using remote sensing. Abstract Precision viticulture (PV) seeks to optimise the health and yield of vineyards by providing farmers with site-specific management schemes. Through these practices, the environmental impact of farming can be reduced. Remote sensing (RS) developments have risen through the years, and new data availability has revolutionised the PV field. While previous RS studies have attempted individual vine delineation, no industry standard exists. This study aims to delineate individual vine plants using drone-captured RGB imagery (red, green and blue waveband). The computer vision algorithm, region grow, has been investigated in a geographic object-based image analysis (GEOBIA) environment to extract individual vine plants. The model achieved an intersection over union (IOU) score of 0.54 and an area under curve (AUC) score of 0.59, representing the algorithm’s predictive performance against the ground truth. While the results indicate only moderate performance, they provide ground for further research.

Background Precision viticulture (PV) and autonomous farming (AF) have emerged as a dynamic duo, reimagining grape production. These transformative practices can bring automation to existing techniques and inspire the creation of new ones, thereby making invaluable improvements to the quality and yield of grapes and promoting more sustainable farming techniques. PV approaches and AF are highly reliant on accurate crop data. Remote sensing (RS) technologies, such as unmanned aerial vehicles (UAVs), commonly referred to as drones, with their ability to capture detailed measurements from a distance, provide crucial data and insights driving automation in farming. The model proposed by this study utilises RGB imagery (red, green and blue waveband) captured by UAVs and computer vision to detect individual vine plants. Accurate delineation of individual vines could contribute to more precise spraying of pesticides by autonomous vehicles or provide more detailed information on vine plants’ biophysical characteristics. This research demonstrates the potential of the region grow algorithm and RGB-UAV/RGB drone imagery for individual vine extraction. Materials and methods The process of extracting individual vine plants was achieved through a two-step workflow, which consisted of 1) extracting vine rows, and 2) delineating individual vines. The two-step workflow is detailed in the following sections.

Part 1: Vine row extraction Site and data

The data used was provided by Stellenbosch University’s Department of Viticulture and Oenology and consisted of drone RGB imagery captured at a 2 cm resolution. The imagery was collected over Thelema Mountain Vineyards in Stellenbosch, Western Cape of South Africa, during the 2020/2021 growing season. Additionally, ground control points (GCPs) were collected on-site, which served as a reference to digitise the outline of 13 vine plants. The digitised vines were crucial in the accuracy assessment process, as they were used to compare the model predictions with actual ground-truth values. The GCPs also provided a measurement of the average vine plant spacing in this vine parcel. Data preparation Three vegetation indices were created from the RGB imagery to allow for better vine row detection. The indices selected were the green leaf index (GLI), excess green excess red (ExG-ExR) and colour of vegetation (CIV). Recommendations by De Castro and other authors (2018) informed the selection of the vegetation indices. Row extraction and classification The row extraction and classification process were completed using the eCognition Developer 10.2 software (Trimble Geospatial, 2022). The drone images were classified into three categories: Vine, shadow, and inter row area (IRA) using two algorithms, multiresolution

10

TECHNICAL YEARBOOK 2024