Technical Yearbook 2024

FIGURE 8. AUC curve.

Part 2: Individual vine delineation Figure 4 depicts the result of the penultimate phase of the workflow illustrated in Figure 1. This phase involved the creation of centrelines using the vine polygons from Figure 3. Figure 5 shows the final output of the Figure 1 workflow, showcasing the generation of seed points along the previously generated vine row centrelines. Each seed represents a prediction of a vine plant’s location. Figure 6 shows the seeds overlaying the generated individual vine objects. The final individual vine objects can be seen in Figure 7.

Accuracy assessment The intersection over union (IOU) accuracy assessment was used to validate the correctness of the detected vines. This approach measures the overlap between the predicted vine areas and the ground truth values. The model achieved an IOU score of 0.54, representing moderate predictive accuracy. An area under curve (AUC) score was also employed to assess predictive accuracy, with the model producing an AUC score of 0.59. Figure 8 shows that the AUC curve lies above the 0.5 line, indicating that the model does possess a level of predictive prowess.

Conclusion The results highlight the potential of drone imagery for the extraction of individual vines. The IOU and AUC scores of 0.54 and 0.59 show that the model is not ready for large-scale application. However, the research does provide grounds for future research. One of the major positives of the study’s results is the demonstrated utility of RGB imagery. While multispectral imagery is more commonly used than RGB imagery due to the usefulness of the near-infrared (NIR) band for vegetation indices, these multispectral sensors are more expensive than RGB sensors, making them less accessible, particularly for smaller farms. Using RGB drone imagery would provide affordable methods for improving farm management techniques, benefiting a broader spectrum of farmers. The developed methodology provides a point of departure for detecting accurate vine foliage areas and vine plant locations, allowing farmers to monitor the health of individual vine plants and detect missing or dead plants. Vine-level data could also assist with more accurate automated crop spraying. With continued research and enhancements, the model has the potential to achieve greater accuracy, robustness and automation. This progress could pave the way for widespread adoption within the viticulture industry. 

References https://www.wineland.co.za/mapping-individual-vines-using-drone-imagery/

For more information, contact Matthew Wrensch at mwrensch98@gmail.com.

12

TECHNICAL YEARBOOK 2024

Made with FlippingBook Ebook Creator