Using AI to Manage Powerline Vegetation
Challenges in Powerline Vegetation Management
Powerlines are essential to delivering electricity to homes and businesses [1]. Vegetation falling across powerlines is one of the largest causes of power outage and bushfires [2,3]. Therefore, vegetation needs to be managed and cleared to reduce risk and improve safety.
Vegetation management has been costly for energy providers. Traditional vegetation encroachment monitoring methods have relied on manual inspection. This is labour and cost-intensive, especially in regional and remote areas due to limited access and large distances to cover [2,3,4]. Ergon Energy, one of Australia’s energy giants, has been spending $80 million annually to inspect and manage vegetation encroachments on powerlines [2].
Satellite Imagery and Machine Learning Solutions
Recently, remote sensing data has been utilised to enable offsite vegetation encroachment monitoring at lower cost than online inspection [2]. These data can be collected using various techniques including satellite observation, Light Detection and Ranging (LiDAR), airborne laser scanning systems, Unmanned Aerial Vehicles (UAV) [4], etc.
Satellite imagery provides a cost effective option to offsite monitoring of vegetation encroachment. In comparison with UAV and airborne drones, satellite imagery can be obtained at lower cost for large areas and also has wider coverage [4]. Satellite imagery providers, such as MAXAR or Planet, offer the data of various ages and locations at different costs.
The advance of artificial intelligence (AI) provides capabilities to turn satellite imagery data into valuable insights and better facilitate offsite monitoring of vegetation encroachment. With appropriate labelled data, machine learning (ML) models can be taught to locate current encroachment, detecting changes and growth rates of vegetation, estimating grow-in and fall-in risks [5], etc.
Case Study: Vegetation Height Segmentation on Satellite Imagery with ML
The first and foremost step towards monitoring vegetation encroachment is to locate and segment the vegetation by height. Knowing the boundaries and heights of vegetation enables vegetation management experts to quickly locate current or imminent vegetation encroachment and come up with timely actions. In the long run, this information can be utilised to estimate horizontal and vertical growth rates of vegetation by observing the changes for a period of time. The risk of fall-in or grow-in can then be evaluated and inform the vegetation clearance schedules.
In this case study, we outline the process of training a ML model to segment vegetation by height on a satellite imagery sample. Our goal was to produce a ML model that took a GeoTIFF satellite image as input and output the height classification for pixels in the image.
Each pixel could be one the 4 classes:
(1) background,
(2) vegetation shorter than 1 metre,
(3) vegetation from 1 to 5 metres, or
(4) vegetation taller than 5 metres.
We used a 8-band pan-sharpened WV3 imagery sample with 6758 x 9313 pixels and 30cm resolution at Malak, Northern Territory as our training data (see Figure 1). The height classification labels (ground truth) were provided by ArborCarbon. Remote sensing techniques, such as LiDAR, provide point cloud classification that could also be used as the ground truth. We reserved a hold-out sample of 20% data for validation and trained the ML model on the other 80% (see Figure 2).
Figure 1. Training data: a pan-sharpened (8-band) WV3 imagery sample at Malak, Northern Territory with corresponding vegetation height classification labels (ground truth) by ArborCarbon overlayed.
Figure 2. A 80% / 20% split for training / validation. Red indicates training and blue indicates validation. Patches (size 256 each) containing more than 90% background class are excluded from consideration, thus creating holes and irregular boundaries.
For the ML model, we employed UNET – a convolutional neural network model initially built for biomedical Image segmentation [6]. Figure 3 demonstrates the learning process of our ML model, which automatically adjusts itself and performs better after each epoch (scanning the whole training data). The process stops after the validation performance becomes saturated.
Figure 3. The learning process of our ML model. Ground Truth (left) and the Model Prediction after training (right).
We then applied our trained ML model to segment vegetation by height in our area of interest (AOI). We use a 4-band pan-sharpened HD imagery sample from Maxar with 16384 x 16384 pixels and 20cm resolution at Goulburn, New South Wales for AOI. To visualise the segmentation results, we used an extracted sample of 2048 x 2048 pixels near Crookwell Road as in the left image of Figure 4.
The right image of Figure 4 presents the vegetation segmented based on height by our ML model, together with powerlines and poles (provided as shapefiles by Essential Energy) on the satellite imagery. Because vegetation lower than 1 metre is unlikely to physically encroach powerlines, its segmentation results were omitted from the visualisation.
Figure 4. A Maxar World View 3 Satellite imagery sample at Crookwell Road, Goulburn, New South Wales (left), and vegetation height segmentation by our ML model and Essential Energy powerlines (right).
Figure 5. Potential vegetation encroachment areas identified by the ML model.
The segmentation results show potential vegetation encroachment areas where the vegetation might overlap with powerlines (Figure 5). These results provide valuable insights to assist experts in making informed vegetation management decisions.
Our ML model can be significantly improved if it is trained on larger imagery data that covers similar regions, and has similar resolution and number of bands, as the AOI where the model is applied.
Acknowledgements
The segmentation model used in this blog was developed as part of a research project funded by Hort Innovation (ST18000) and in collaboration with AARSC. Maps of powerlines and associated spatial data were provided by Essential Energy.
References
- Department for Energy and Mining (2017), Vegetation near powerlines. Online Article.
- Li, Z., Walker, R., Hayward, R., & Mejias, L. (2010). Advances in vegetation management for powerline corridor monitoring using aerial remote sensing techniques. In 2010 1st International Conference on Applied Robotics for the Power Industry (pp. 1-6). IEEE.
- The Australian Business Review (2021). Power outages decline with AI and high-res weather forecasting. Online Article.
- Haroun, F. M. E., Deros, S. N. M., & Din, N. M. (2020). A Review of Vegetation Encroachment Detection in Power Transmission Lines using Optical Sensing Satellite Imagery. arXiv preprint arXiv:2010.01757.
- Schouten, A. (2020). Rethinking Vegetation Management on Electric Utility Corridors — combining AI & Satellite Data. Medium Article.
- Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234-241). Springer, Cham.