Classification of high-voltage power line structures in low density ALS data acquired over broad non-urban areas

Step 1: tower tracking

The first step consists of mapping the positions of the transmission towers. Starting from the map of the network in the form of spatial lines, we decompose the network into linear sections and then each line is buffered to encompass the power lines (Fig. 4A). This allows for working iteratively on linear sections with a known orientation and is helpful for reducing scene complexity. This way, the algorithm can focus on smaller numbers of points instead of screening the whole scene, which in our case usually covered an area greater than 1 km² and was free of any human structures.

In each region of interest we applied a point-cloud-based local maximum filter (LMF) to find the local highest points. The local maximum filter we used is analogous to that normally used to detect individual trees (Popescu & Wynne, 2004), but specifically uses narrow but very long rectangular search windows oriented parallel to the wires to reduce the number of false positives. By design, the LMF usually finds all the towers but tends to also produce several false positives corresponding to high vegetation or to sparse earth wire points that are locally higher than neighbouring points (Fig. 4B). In addition, the true positive towers are usually not very accurately positioned because the LMF finds the single highest point of each tower i.e. the “ears” of the towers instead of the centre (see waist type in Figs. 3 and 4B). The algorithm then fixes these two issues by removing locations that do not actually correspond to a tower position and centres the position of the tower top for the remaining true positives.

Because the success of steps 2, 3 and 4 is highly dependent on the accuracy of step 1, this correction task is critical. If a single tower is missing, or a single false positive tower remains, it will invalidate subsequent analyses. Consequently, a very robust trimming of false positives is key to the success of the segmentation. To achieve this, we extract the surrounding point cloud at the position of each candidate tower (Fig. 4B). The subset is then analysed to confirm whether or not it actually contains a tower. This is based on an analysis of the vertical and horizontal distribution of the points. Interested readers may consult the publicly available source code for further details of how this task is performed. Here, we opted to provide limited detail because this step is likely to be modified and improved in the future, which could in turn invalidate the description provided in this paper.

Using the subset of points surrounding the position of the tower, the realignment of its position is made by averaging the X and Y coordinates of the points located 5 to 10 m below the location of the highest point (Fig. 4B). Because towers are vertically elongated structures, this task is easy to implement and provides good results.

The xyz positions of each tower is then assigned as their centred tops. With the map of the network and the prior information about the tower types, we can also assign the general orientation of the power lines as well as the bounding box of the towers (Fig. 4C). In Fig. 4C the towers drawn in red were found twice in each linear section because the network is forking. These correspond to “deflection towers”, which have wires with different orientations on each side. In Fig. 4C, we can see that two deflection towers were detected, but one of them was a false positive. This is not an issue as will be made evident in the next section.

Step 2: topology reconstruction

Once transmission towers are accurately identified and positioned, step 2 consists of retrieving the network topology by determining how towers are interconnected. For this we use the known orientation of the power line and send 500-m beams in the direction of the previous and next towers with a range of angles corresponding approximately to the width of one transmission tower. The topology of the network is built progressively by identifying the beams that intercept the exact location of the next tower. The 500-m limit was used to avoid scenarios whereby implausibly distant towers are connected by mistake. Figure 4D provides an example of connections in which we retrieved three rows of power lines with one being deflected in a different direction. We can see that the false positive deflection tower presented in Fig. 4C was not an issue because it could not connect with any other tower in the direction of the deflection.

In this step we also needed to introduce the concept of “virtual towers”. Because two towers are needed to compute a connection, we necessarily had troubles at the edges of our datasets where the last towers could not be connected. Virtual towers are thus added when a tower cannot be matched with another tower. While this was always the case at the edges of the studied area, the concept was also useful in cases where towers were missed (false negatives). In such cases, two consecutive towers may be too far apart to be matched, thus creating a gap in the network. The addition of virtual towers helped make the topology reconstruction more robust. The exact location of missing towers was not known, but this was not our purpose. The objectives pursued with the creation of virtual towers were twofold: (1) to ensure the topological validity of networks, and (2) to reduce the effects of missing towers on the detection of wires. The latter will be made clearer after “Step 4: Wire Classification”. Although they are an approximation, virtual towers contribute to reducing the effects of errors made in step 1. There is no missing tower in Fig. 4D, but we can see the prolongation of the network in lighter colours after the last towers at the edges of the area.

Step 3: tower classification

At this stage we know the positions of the towers, their height, their orientation, their type and their interconnection in the network. However, the point cloud remains to be classified. As we know the tower type, we also know their xy dimensions. In this step, we assign all points located within the bounding boxes of the towers down to the ground level as belonging to a transmission tower (Figs. 5A and 5B).

Step 4: wire classification

In this step, we use a well-known phenomenon from physics and geometry to classify points belonging to the conducting wires. An idealized hanging chain or cable supported only at its ends assumes, under its own weight, a type of curve called a catenary. The general equation describing the catenary between two points is (Hatibovic, 2014):

(1) $$y(x)=2c\left(A(x)-B(x)\right)+h_1$$

With

(2) $$A(x)={\sinh }^2\left({\displaystyle \frac{1}{2c}\left(x-{\displaystyle \frac{S}{2}+c\times \mathrm{a}\mathrm{r}\mathrm{c}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{h}\left({\displaystyle \frac{h_2-h_1}{2c\times \sinh ({\displaystyle \frac{S}{2c})}}}\right)}\right)}\right)$$

(3) $$B(x)={\sinh }^2\left({\displaystyle \frac{1}{2}\left({\displaystyle \frac{S}{2c}-\mathrm{a}\mathrm{r}\mathrm{c}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{h}\left({\displaystyle \frac{h_2-h_1}{2c\times \sinh ({\displaystyle \frac{S}{2c})}}}\right)}\right)}\right)$$

Where (x₁, h₁) and (x₂, h₂) are the position and elevation of the cable ends, c a constant that (roughly) corresponds to the tension in the wire, and S the span length i.e. |x₂ − x₁|.

Despite the apparent complexity of this equation, it has a single parameter c, which is part of the description of the tower, assuming that for a given tower type the tension is always the same. We can thus reconstruct the wire curve from the top of the towers (Fig. 5C). This curve is positioned above the wire, but since we know the tower type we also know the typical distance between the top of the tower and the wires (Fig. 3). We then move this curve below the wires (Fig. 5D) and classify every point above the curve that is not already classified as a tower, as a wire. This includes the earth wires that, in practice, look like noise or outliers but are actually wires (Fig. 5E).

The final classification for the dataset used to describe the method is showed in Fig. 6. Figure 2 was also the result of using this method.

Datasets

To validate the algorithm, we selected 130 1-km² tiles containing wires among a 30,000-km² dataset located in the Côte-Nord region of eastern Quebec, Canada (Fig. 7). Although the tile selection was blind (i.e. made without looking at the point cloud), it was not fully randomised. Tiles were grouped into 28 regions of interest, which were selected manually to include difficult cases, including deflections, forks, gaps, single rows, multiple rows, multiple tower types, etc. A fully randomised selection would have very likely produced only linear sections with waist-type transmission towers, which are the most common in the network. Figure 2 shows some small linear subsets from some of the chosen regions of interest. Our validation dataset consisted of 28 regions of interest, with conductor lengths ranging from 2 to 12 km in areas ranging from 4 to 8 km². Each section contains an electrical network in single (Fig. 2A) or multiple rows (Fig. 2B). To estimate the network length, we multiplied the length of the power lines by the number of rows. For example in Fig. 2B there are three times 900 m of wire conductors i.e 2.7 km.

The point density averaged 7 points/m² over the full dataset and ranged locally from 2.6 to 13.8 points/m². The full 30,000-km² dataset was not sampled in a single contract was acquired over several years. Thus, the densities tended to change among regions. Because of the multiple contractors, the acquisition devices also varied. For these reasons, the sampling design specifications are not reported here. Our proposed method should be seen as working independently from the acquisition device and point density.

Validation

To evaluate the accuracy of the method, we manually assessed the results obtained in our 28 selected regions of interest. In the absence of reference data on the positions of the transmission towers, we manually assessed the tower segmentation and counted the number of true positives, false negatives and false positives. This relatively trivial task, was performed on a total of 415 towers.

Assessing the accuracy of the wire segmentation was more challenging. In the absence of a perfect segmentation upstream of the method, we could not directly and automatically estimate the percentage of points that were correctly or incorrectly classified. The key to the successful application and development of our method was to identify situations in which points tend to be misclassified. To this end, counting the percentage of correctly classified points did not bear much relevance. For example, we could miss all points belonging to conductor wires in dataset #15 (for which a subset is presented in Fig. 2D) without significantly affecting the overall percentage of correct classification because this dataset is so sparsely populated that it actually accounts for a very small proportion of the points included in our analysis. We thus chose the length of true positive, false positive and false negative sections of power supply facilities as a quantitative measure of segmentation accuracy. This way, the same weight was attributed to each dataset independently of their sampling density. By design, the misclassification errors output by our algorithm are grouped into continuous clusters. Results were thus free of isolated and randomly-located misclassified points, which ensures the relevance of the chosen metric. This approach also allowed us to derive some key statistics to assess model performance, such as precision, recall and F-score, which were interpreted in kilometres of power lines instead of percentage of points.

Accuracy

Table 1 shows, for the 28 evaluation datasets, the dataset IDs (so we can refer to them in the text), the number of power line rows, the number and types of transmission towers supporting the wires, as well as the number of true positives, false positives and false negatives. Regarding the wires, the table also presents their total length (i.e. the length of the network multiplied by the number of wire rows) as well as the total lengths of true positive, false positive and false negative sections. Overall, 21 of the 28 trials were perfectly classified with with 0 false positives and 0 false negatives. Figure 6 provides an example of such perfect classification obtained in dataset #13. The seven remaining datasets contained various degrees and types of classification errors. The classification errors of datasets #1 #15 and #17 are presented in further detail in the next sections to highlight some limitations of the method. The NA values associated with datasets #10 and #22 correspond to areas where we considered the algorithm had failed. In both cases, largely incorrect results were produced whereby false positives and false negatives were no longer clustered in continuous sections.

For power lines supported by waist-type transmission towers, a total of 1.7 km of structures were misclassified (mostly false negatives), which represents 1.4% of the total studied length. If all power supply structures had been equally densely sampled, this would correspond to an accuracy of 98.6%, a precision of 99.7%, a recall of 98.9%, and an F-score of 0.993 at the point level. For power lines supported by double circuit transmission towers, a total of 740 m of structures were misclassified (false negatives only), which represents 1.4% of the total studied length and corresponds to an accuracy of 98.6%, a precision of 100%, a recall of 98.6%, and an F-score of 0.993. For power lines supported by small waist-type towers, however, the results were largely unusable with several occurrences of false positives at the tower detection stage. Tall trees at random locations were often classified as towers in such situations, which subsequently led to vegetation being classified as wires and the actual wires being missed. Overall, ≈75% of the studied length was unusable in such situations.

It was not possible provide an overall estimate of accuracy for the application of the algorithm to the entire network because the relative proportions of these three tower types was unknown. However, it was evident from our samples that waist-type towers are dominant in the studied region, with double-circuit towers also being relatively frequent and small waist-type towers relatively rare. One key factor associated with the poor results obtained for small waist-type towers was their height being sometimes lower than some of the neighbouring trees.

Imperfect classification

Figure 8A shows a subset of region #1 in which three towers are missing where the conductors split and joined back. Here, the “split towers” are of a different type to the others on the line, which explains why they were trimmed during the candidate tower identification step. We can observe that where the two side-by-side towers were missed, the section was automatically removed. This left a continuous cluster of unclassified points but none were actually misclassified. This is because the span of the connected towers that were correctly identified on each side was so large that the theoretical wire profile would be passing below ground according to the tension normally applied to this type of tower. Facing this impossible scenario, the algorithm automatically removed the entire section. Where only one of the two side-by-side towers is missing, as seen on the right hand side of Fig. 8A, the topology remains valid. In this case, only one part of the split was correctly classified while the other remained unclassified, once again without any false positive classification.

Figure 8B shows a subset of the region #15 that focuses on the missing tower. Consequently, one tower could be not matched to any other and the topology of network was interrupted. Despite this, large sections of the wires were still correctly segmented, again avoiding false positives. This was attributable to the addition of a virtual tower, as explained previously, which left only a small portion of the wires unclassified with no false positives.

Figure 8C shows the entire region #17 in which the segmentation missed a lot of features. This scene corresponds to an extreme case where facilities are so sparsely sampled that we can only see the insulators of the transmission tower in the point cloud. Consequently, several towers were not found, and thus all subsequent processes failed. Because of the missing towers, large spans were computed, and consequently catenary curves went very low. This explains why some parts of the vegetation were classified as wire. Despite this extreme case being very hard to segment, approximately 75% of the scene was still perfectly segmented. Figure 2D shows another part of this evaluation dataset that was perfectly segmented.

These three cases are representative of the errors typically found in our validation exercise. The latter example can be used to highlight the importance of reporting results in terms of the proportion of correctly classified section lengths. Indeed, due to the the lower accuracy of the method in areas with low point densities, a point-counting estimate of the overall accuracy of the algorithm would be artificially high. In this study, a point-counting estimate would have produced an accuracy value above 99.5% compared to 98.9% for a length-based assessment.

Accuracy and scope of the study

The application of our algorithm provided excellent results with an accuracy of 98.6% (F-score = 0.993). However, the high accuracy of our method should be taken in the specific context of this study as they are not directly comparable to other methods. First, unlike previous methods cited in the introduction, the application of our method is dependent on the availability of a priori information. Second, our method applies to a totally different type of data than is analysed using most of the other existing methods. As highlighted in the introduction, line-shape-based methods would have likely failed at segmenting our data, at least where the point density is low. Conversely, our method would also fail to segment several of the cases reported in the literature. The accuracy of our results should therefore be interpreted within scope of our study i.e. the percentage of high voltage facilities that can reliably and accurately be removed from low density point clouds to facilitate the subsequent study of the neighbouring vegetation. This method was never intended to be used for monitoring the state of the electrical distribution network. We found McLaughlin (2006) and Zhu & Hyyppä (2014) to be the two closest comparable studies to this one. Zhu & Hyyppä (2014) studied power line segmentation in a forestry context and achieved an average correct classification rate of 93% with 55 pts/m² on approximately 2.5 km of power lines. McLaughlin (2006) worked with a low density point cloud (2.5 points/m²) in a forested area testing their method on a relatively larger network (14 km) and obtained an accuracy of 82%.

Further improvements could be implemented relatively easily by refining the tower detection method in step 1, which represents a very small fraction of the source code (50 lines among more than 1,000) for the two most frequent tower types in our study area (waist-type and double-circuit). The tower detection step is indeed the most critical in the process. When tower detection reaches an accuracy of 100%, the overall segmentation will also generally reach an accuracy of 100%. However, if a single tower is missing, or worse a false positive tower is detected, the algorithm will start failing.

An important feature of the method is that, by design, the errors are grouped in contiguous clusters. This is important and highly advantageous for safely cleaning the point cloud. When false positives or false negatives can be randomly spread over the study area, even a 99% accuracy can lead to a high percentage of pixels in which the underlying vegetation is statically hidden by the remaining outliers, as explained in the introduction. Our method ensures that the vast majority of pixels will be fully cleaned and that only a minority of contiguous pixels may contain measurement artefacts.

Imperfect classification and unsupported cases

Most of the regions tested were perfectly segmented according to our manual inspection; however, some errors were still occurring when towers were badly detected. Because our method relies fully on the location of the towers to find the wires, a single error in the tower tracking step may have cascading effects. We observed four cases:

1. Figure 8B shows a very simple case where a single tower that apparently looks easy to find was missing for an unknown reason. This case will probably be better handled as a result of the continuous improvement of the tower candidate correction step, but we think it highlights that some single towers might be missed. In such cases, the introduction of a virtual tower can limit the effect of a single missing tower on the wire detection. Our example shows that wires were mostly correctly segmented and only a small proportion were missing. Virtual towers can be seen as safeguards providing greater robustness to the method.

2. Figure 8A shows a case that was simply not handled by our method. We did not anticipate the existence of such topologies and we discovered this case at the validation stage. This case will be harder to process correctly but it remains a rare occurrence over the full network. Limiting errors in such cases will require in-depth redesign of the topology reconstruction step, as well as an improvement of the tower candidate correction step.

3. Figure 8C shows a very complex scene where the electrical facilities are only sparsely sampled. Towers are often reduced to 10 to 30 points at their very top and we actually see only the suspension insulators. Despite the limitations of the method in very low density point clouds we found a smaller proportion of false classification than expected and the segmentation accuracy still reached 75%. In the portions that were incorrectly classified, the human eye could not even distinguish the suspension insulators of the towers.

4. Regions #10 #23 and #24 presented cases that were badly handled where small transmission towers were situated adjacent to tall trees, which were wrongly interpreted as towers. Improving this result will require further refinement of the candidate tower correction step, which represents a very small portion of the code that drives the entire method.

In case 4, two situations may occur: (1) a tall tree is close to a tower and the local maximum filter detects the tree instead of the tower, leading to one false negative and one false positive tower; (2) several tall trees are detected as towers and even if the correction step is robust, a few false positive towers may remain because trees and towers are more similar than expected. While this has resulted in very poor outcomes, our results suggest that producing a more robust tower detection method in step 1 would be sufficient to improve the segmentation accuracy. No modifications would likely be required for the other three steps.

Case number 3 is to put in the perspective of case 4. The first step of our method could be made more robust by the addition of tests aiming to determine if the distribution of points matched the expected morphology of a transmission tower rather than another object, such as a tree. Such tests were not included in our method because of situations where the point density is so low that even the human eye is unable to recognize the features of transmission towers in the data. The reality of our sampling implied an important trade-off between the capacity of our algorithm to robustly make the distinction between a tree and a tower, and the ability to robustly recognize towers that are so sparsely sampled that their morphological features are unrecognizable. Our method was made relatively robust to sparse sampling but this comes at the expense of the detectability of small towers. The threshold density at which power supplies are no longer recognizable is hard to estimate. Dataset #17 was sampled with a nominal density of 2.6 points/m², but what matters is the sampling density of the target structures, which may also be affected by the direction of the moving sensor relative to the structure. In this case it was both a low-density point cloud and a sensor moving perpendicular to the power lines.

There are obviously some other limitations to our algorithm for which consequences were not observed in our dataset. Our inability to fully anticipate all possible features of high-tension power lines implies that we cannot provide an exhaustive list, but the following limitations are worth mentioning. Our method does not apply to networks supported by different types of transmission towers along the same wires. A typical case can be found in Chen et al. (2018) figures 10 and 11. Similarly, it does not apply to networks with very close side-by-side power supply facilities of different types e.g. one row of waist-type towers adjacent to another row of double-circuit towers. Also, any overly-complex scenes, such as networks located in the vicinity of generators containing several wires and towers, would likely make the method inapplicable.

Priors and limits of application

Our method is assisted by prior knowledge of the network including wire orientation and tower types. This is an obvious limitation of the method if such prior information is not available. However, high voltage power lines are usually mapped. In more complex contexts, such as in urban regions where lower voltage conductors are present, having this kind of prior knowledge may not be possible. The segmentation of urban electrical facilities requires different methods and point clouds sampled with higher densities. Our method would not be applicable in such contexts.

Nevertheless, our method still uses a limited set of priors. For example, it does not require prior knowledge of the number of power line rows. Yet, we observed that it performs well on one to four rows (four rows not shown) and there is no hard upper limit. Our method also does not require any prior knowledge of the position of the towers along the network, since it is capable of detecting them. In practice, power supply companies typically maintain maps of high-voltage transmission towers, which may lead to even more accurate results. The method also includes safeguards to prevent against irrelevant false positive classifications. This is what we showed in Fig. 8A where the ground was initially classified as wire but was automatically corrected, and in Fig. 8B where a virtual tower prevented against missing too many wire points. Also, unlike previous methods presented in the literature, we proved that our method works not only on linear sections, but also supports branching and deflections of networks. It works with partial objects in low-density point clouds and does not need any special flying pattern (e.g. following the power line). We believe it is thus applicable to any natural landscape in which broad ALS coverage is available.

Implementation and reproducible work

Broad ALS coverage is normally split into multiple files. We presented the method on single blocks of data loaded in memory, but in practice it would not be possible to load thousands of tiles at once. Unlike methods presented in the literature that require pre-processing to fit each segment into a single file (e.g. Guo et al., 2019), our method works independently of the tiling pattern. Each tile is loaded with a buffer, thus ensuring that the virtual towers that are added at its edges are actually out of the processed region and do not affect the classification. This is made possible through the file collection processing engine provided with the lidR package (Roussel & Auty, 2021; Roussel et al., 2020). The method is therefore not only workable on small samples but it is also fully integrated into a framework dedicated to processing large areas seamlessly.

The text presented in this paper would not allow the methods to be fully reproduced. We have presented the overall concept, but for the sake of concision several details were omitted about candidate tower cleaning, virtual towers and deflection towers, for example. To ensure reproducibility, the source code has been made publicly available, so that any interested user can implement the method and analyse it in detail. The implementation is made of four functions corresponding to the four steps described in the materials and methods section. Any step can thus be replaced easily by another method to match users’ needs without affecting the other steps.

The method is available within the R language (R Core Team, 2020) and is implemented in the lidRplugins package available at https://github.com/Jean-Romain/lidRplugins. lidRplugins is an extension of the lidR package (Roussel & Auty, 2021; Roussel et al., 2020) and contains experimental methods. Depending on its future improvements and success when applied to other datasets, this algorithm might at some stage be made directly available in the lidR package.

The package does not record every existing transmission tower type in a database. Instead, users have the capability to define and use their own tower types, which makes the method reproducible in jurisdictions where power lines are supported by tower types that are different to those presented here.