What is ground truthing and when is it undertaken?

This ensures that what is observed or measured in the imagery is aligned with the features at the ground level. Collecting ground truth information facilitates the calibration of imagery and is important for the imagery labelling process required for developing training datasets for artificial intelligence models.

Ground truthing for weed detection is initially carried out at the time of imagery capture, since this will ensure that all plant species and landscape features within the imagery are most accurately represented by what was observed on the ground in the region of interest at that time.

Ground truthing also occurs after machine and deep learning models have been trained by a portion of the data. The models will identify plants when provided with a test imagery dataset. To assess the accuracy of these detections, a ground truth check of the plants at the same location on the ground will be required to verify the accuracy of the models in detecting the species in reality. This is called the ground truth accuracy. It should be carried out as soon as possible after the imagery was captured to ensure there are minimal differences in plant growth that make the imagery look different to when the data was captured. If plants look considerably different, this can lead to a mismatch in the proportion of target plants present.

This process can  be summarised as follows:

1. Input all the data into the identification algorithm for the desired target plant.
2. Identify the plant desired by visual observation in the field with additional photographs of the habit and habitat for key identification factors through different life cycle growth phases.
3. Mark and identify the plant with a visual indicator  such as a quadrat and visual arrow to identify the plant at different levels of imagery height and resolution.
4. Take the imagery using the drone with and without the indicators in all growth and life cycle phases.
5. Fine tune the process using developmental algorithms for all the plant data and identifying in the  learning process to differentiate between look-alikes, similar species and multiple habitats of occurrence.
6. Proof the imagery by search and select drone imagery at another site.
7. Analyse the imagery and the algorithm interpretation.
8. Then verify the analysis of the algorithm by ground truthing its findings.
9. Continually refine the processes and the machine learning until it can find the desired species in multiple habits and habitats with a high % degree of acceptable accuracy.