Technical Yearbook 2024

Buffer [50 cm]

Polygon to raster

Raster to polyline

Create random points

Thin

FIGURE 1. Centreline extraction workflow.

FIGURE 2. MRS segmentation at a high zoom level.

FIGURE 3. Resulting classification (three classes).

FIGURE 4. Creation of centreline (in red).

FIGURE 5. Seeds (red) generated along the centrelines.

FIGURE 6. Seeds (crosses) grown.

FIGURE 7. Final region grow objects outlined in red to the extent of the vine row.

segmentation (MRS) and the decision trees (DT) classifier, also known as classification and regression trees or CART.

2022) using the vine row polygons as the bounding boxes and the generated vine plant location estimates as the seed pixels. The vine plant extraction process was iterated 50 times to ensure that the full extent of each vine was detected. Results Part 1: Vine row extraction The results from Part 1 of the process can be seen in Figures 2 and 3. Figure 2 depicts a zoomed-in view of the image objects generated in the segmentation process. Figure 3 shows the final row extraction results obtained through the DT classification. In Figure 3, green represents the vine, black represents the shadow and yellow represents the inter row area (IRA). The vine rows were extracted as polygons and used as input for Part 2 of the processing workflow. It should be noted that some post-processing was required to remove grainy misclassifications and to remove grass patches incorrectly classified as a vine.

Part 2: Individual vine delineation Centreline and seed extraction

The extracted vine rows were imported into ArcGIS Desktop 10.8.1 (Esri, 2020) for further processing. The workflow outlined in Figure 1 was followed to identify the centrelines of each vine row. An average vine length of 1.75 m, as determined by in-field GPS measurements, was used to generate seed points or pixels from these lines. Vine plant extraction The region grow algorithm was used to obtain the final vine plant delineation. Region grow is designed to take a seed pixel and expand it outwards to a specified bounding box threshold. The algorithm was implemented in the eCognition Developer 10.2 software (Trimble Geospatial,

11

TECHNICAL YEARBOOK 2024