Identifying Priority Areas for Vegetation Management in the Context of Energy Distribution Networks Using PlanetScope Images

Abstract

In Brazil, approximately 30% of unscheduled interruptions of energy supply are caused by fires and vegetation interference in the energy distribution networks, resulting in great losses for companies of the electricity sector. To reduce the interruptions caused by these kinds of events, the energy distribution companies continually monitor and manage the vegetation in the vicinity of electric cables. However, due to the great extension and capillarity of the networks, it is not always possible to cover the entire network, and it is necessary to define priority segments to be managed. Taking into the account this context, the main objective of this study was to develop multi-criteria indicators to identify segments of the energy distribution networks with higher priority for management, based on vegetation attributes extracted from remote sensing images. For this purpose, we tested two artificial intelligence algorithms, support vector machine (SVM) and artificial neural networks (ANN), to automatically identify different classes of vegetation using PlanetScope images. Our results showed that the ANN algorithm presented better results for the vegetation classification when compared to the results obtained with the SVM algorithm. The application of the developed indicators showed adherent results, even in densely urbanized areas. We hope that the use of the developed indicators can help Brazilian energy distribution companies in optimizing vegetation management and consequently reducing unscheduled interruptions.

1. Introduction

The presence of vegetation close to aerial energy distribution networks can cause interruptions in energy supply, negatively impacting quality indicators and hence generating losses for energy distribution companies in Brazil. The interruptions of energy caused by vegetation are mainly due to its natural growth, its movement during meteorological events, and also due to the occurrence of fires during dry periods. In addition, poor planning of energy distribution networks, combined with technical and legal restrictions, worsen the problem, since in many cases, the network layout coincides with previously vegetated areas. To reduce the risk of interruption in the energy supply, companies perform the routine monitoring and management of the vegetation in the vicinity of electrical networks. However, due to the large extension and capillarity of the distribution networks, it is not always possible to monitor and manage their entirety using conventional methods with on-site inspections, which demands a large amount of human and financial resources.

Among the different technologies that have been tested and applied in order to support the monitoring and management of vegetation in the context of the electricity sector, one of them stands out—remote sensing. Images collected remotely by sensors aboard satellites and aircrafts have been used to map vegetation using automatic classification techniques since the 1990s [1,2]. The state-of-the-art nature of the subject shows us that the technological advances that have taken place in the last decade have provided a greater power for remotely discriminating vegetation, with a growing number of scientific articles being published in the last years (e.g., [3,4,5,6,7,8]). This growth is especially apparent with regard to the spatial and spectral resolutions of sensors, as well as with the development of new platforms and data acquisition technologies (e.g., LiDAR, hyperspectral cameras, and UAV).

Recently, images acquired by nanosats constellations have been gaining prominence in commercial projects, with potential applications in different areas, such as precision agriculture and vegetation management. Although this is not an entirely new image collection technology [9], its use has only recently come into evidence for commercial applications. This higher rate of use was largely influenced by the founding of Planet Labs in 2010 in California, which currently operates the largest constellation of nanosatellites in Earth orbit. Among the main advantages of using nanosatellites for data collection, is its high spatial resolution combined with a high frequency of daily data acquisition, which enables the generation of 100% cloud-free mosaics with high spatial resolution. On the other hand, as it is a technology with few applications and published works, it still presents great challenges in relation to data processing and information extraction, requiring further investigations regarding its application potential.

Taking into account the context described above, the main objective of this study was to develop multi-criteria indicators to identify segments of the energy distribution network, with higher priority for vegetation management. For this purpose, Nanosat Multispectral Images were tested to map the vegetation in the vicinity of energy distribution cables, and extract attributes of interest for the multi-criteria analysis.

2. Materials and Methods

2.1. Study Area

The study area encompasses four different sites located in Paraná State, Southern Brazil (Figure 1). The area was defined together with Compania Paranaense de Energia (COPEL), the largest electricity company of Paraná State, a partner in the development of this study, and a provider of valuable information and data for the analyses. The sites were selected, considering the following criteria:

• The existence of energy distribution networks;
• Different landscape complexities: Rural, urban, and dense urban;
• COPEL team interest;
• The existence of auxiliary data in public and COPEL’s databases; and
• The coverage and current costs of nanosatellites images.

Table 1. Summary information about the four sites selected in this study.

Site	Characteristics	Area	Medium Voltage Distribution Network Length	Climatology	Energy Supply Indicators *
S1	Urban	61 km2	564 km	Annual precipitation: 1045 mm Mean wind speed: 0.94 m/s	DEC: 3.99 FEC: 3.24
S2	Rural—predominantly covered by vegetation	1.250 km2	937 km	Annual precipitation: 1498 mm Mean wind speed: 1.24 m/s	DEC: 22.87 FEC: 13.16
S3	Rural—silviculture	830 km2	1.457 km	Annual precipitation: 1449 mm Mean wind speed: 1.40 m/s	DEC: 6.04 FEC: 4.97
S4	Rural—mixture of vegetation and silviculture	850 km2	707 km	Annual precipitation: 1713 mm Mean wind speed: 1.18 m/s	DEC: 22.50 FEC: 11.68

* DEC: Equivalent interruption duration per consumer unit; FEC: Equivalent interruption frequency per consumer unit.

presents a summary of the characteristics of the four areas selected for the development of the project.

2.2. Planet^® PlanetsScope Images

Planet^® is a company based in San Francisco, CA, USA, and it is the owner of the largest constellations of nanosatellites in operation, providing daily multispectral images, with high spatial resolution, of the entire terrestrial globe, More information about Planet® company is available online: https://www.planet.com/ (accessed on 4 January 2022). It is currently the world’s largest operator of Earth observation satellites, operating different constellations of nanosatellites, with different capabilities and complementary characteristics for different applications. The PlanetScope constellation used in this study, has over 130 Dove nanosatellites, with an imaging capacity of 200 million km² per day. The product acquired was the Planet Ortho Scenes, which provides orthorectified images at the surface reflectance level, in four bands of the electromagnetic spectrum (blue, green, red, and near-infrared), and with a pixel size of 3 m. Detailed information about the PlanetScope constellation and the Planet Ortho Scenes product is available online: https://assets.planet.com/docs/Planet_Combined_Imagery_Product_Specs_letter_screen.pdf (accessed on 12 January 2022). The images used in this work were acquired between 26 May 2020 and 30 May 2020, and are all 100% cloud free.

2.3. Image Classification

Two algorithms were evaluated for the classification of PlanetScope images, the support vector machine (SVM) and the Artificial Neural Network (ANN). The SVM is a supervised training machine-learning algorithm that is widely used in satellite image classification problems [10]. As it is a non-parametric algorithm, SVM has the advantage of not being sensitive to the statistical distribution of the input data. In its most basic form, the SVM algorithm can be understood as a binary classifier that is capable of identifying a unique limit between the classes of interest, through the determination of the optimal hyperplane. The ANN is a supervised training deep-learning algorithm that uses a logical structure that is inspired by human brains. The neuron is the processing unit of the ANN algorithm, and it can be connected to the input data or to other neurons through communication channels that are associated with specific weights. Although ANN is not a new algorithm, its use for classifying remote sensing images has become popular after 2000 [11].

Training, Validation, and Classification Procedure

The workflow defined for the training and validation of the classification algorithms consisted of five tasks: (1) Image preprocessing; (2) Image interpretation and classes definition; (3) Sample collection; (4) Training and validation; and (5) Classification and post-processing. The details of the tasks are presented below.

The preprocessing was the initial task necessary for the preparation of images that would be consumed during the training and validation procedure. In the case of the PlanetScope images, the only preprocessing consisted of the operation of mosaicking the images in order to obtain a single image for each area of interest.

The definition of the classes of interest to test the classification algorithms considered the main uses and land cover in the areas selected for the project. Based on the visual interpretation of the PlanetScope images, it was possible to define nine classes of interest, with three classes related to vegetation types: Arboreal vegetation, silviculture, and grass or cultivated areas. The other classes are related to other uses and coverages, such as bare soil, water surface, paved roads, built-up areas, and shadows. Figure 2 shows some PlanetScope image samples of the nine classes considered in this study.

The samples collection for training and validation of the classification algorithms was made using PlanetScope images, with the aid of images of better spatial resolution available from Google Earth^®. The samples were manually collected by an experienced interpreter, using a Geographic Information System (GIS), via visual inspection of the images and the acquisition of georeferenced and labeled points for the nine classes of interest.

Table 2. Samples dataset used to train and validate the classification algorithms.

Class	CAL/VAL Samples Available for Each Site
	S1	S2	S3	S4
Arboreal vegetation	189	670	132	99
Silviculture	-	465	139	104
Grass or cultivated areas	113	409	117	130
Bare land	85	256	114	123
Water	68	91	127	40
Paved roads and streets	108	69	97	-
Built-up areas, Type 1	126	127	112	-
Built-up areas, Type 2	141	-	-	-
Shadow	135	80	134	75
TOTAL	965	2167	972	571

presents a summary of the dataset collected for the training and validation of classification algorithms.

The training and classification of PlanetScope images was performed using the ENVI^® software (https://www.l3harrisgeospatial.com/Software-Technology/ENVI (accessed on 3 January 2022)), which provides several tools for image processing, and algorithms for classification. For each image and project area, several training and classification tests were performed using different hyperparameter configurations, as shown below:

• Activation function: Sigmoid;
• Training threshold: Ranging from 0.1 to 0.4;
• Learning rate: Ranging from 0.7 to 0.9;
• Momentum: Ranging from 0.5 to 0.9;
• Hidden layers: Ranging from 1 to 2;
• Interactions: Ranging from 10 k to 15 k.

The validation step consisted of calculating metrics in order to comparatively evaluate the results of the classifications obtained with the different classifiers and hyperparameter configurations. Among the most frequent metrics used for image classification validation [12], the Kappa index (K) and the overall accuracy (OA) were selected for the present study. For each site, the classification result with the best performance in validation metrics was selected to develop the priority indicator.

2.4. Priority Indicator for Vegetation Management

The priority indicator for vegetation management (PIVM) was developed using the multi-criteria decision analysis technique (MCDA), which is widely used for decision making in critical events, and operations planning in different areas of knowledge [13,14,15,16,17]. In this specific study, MCDA was used to identify the distribution network units (e.g., the network segment, feeders or electrical sets) that have a higher priority for vegetation management, with the aim to improve the performance of energy supply indicators (DEC and FEC).

2.4.1. Criteria, Sub-Criteria and State Variables

The initial proposition of the criteria and state variables to compose the indicator was based on the analysis of the literature [18,19,20], and were initially considered as criteria that were related to vegetation and the environment. Subsequently, in a joint meeting with experts from the COPEL team, the proposed criteria were evaluated, and adjustments were suggested, such as the inclusion of operational criteria, as well as a differentiation between their application in urban and rural areas. The initial proposal was refined, with two indicators being developed, one focused on prioritizing tree pruning in urban areas, (

Table 3. Criteria, sub-criteria, and state variables defined to prioritize tree pruning in urban areas.

Criteria	Sub-Criteria	State Variable
Vegetation	Quantity	Proportional area of arboreal vegetation
	Proximity	Proportional length of network intersecting with arboreal vegetation
	Condition	Mean NDVI of arboreal vegetation
Environment	Climate	Historical mean wind speed
Network	Impact	Number of consumer units
		Proportional length of trunk in the electricity network
		Proportional length of cable in the electricity network
	Performance	DEC-vegetation
		FEC-vegetation
		Accidental interruptions shorter than 3 min

) and another focusing on mowing in rural areas (

Table 4. Criteria, sub-criteria, and state variables defined to prioritize vegetation mowing in rural areas.

Criteria	Sub-Criteria	State Variable
Vegetation	Quantity	Total area of vegetation (arboreal, silviculture, and grass)
		Proportional area of silviculture
	Proximity	Length of network intersecting with arboreal vegetation
		Proportional length of network intersecting with silviculture
	Condition	Mean NDVI of arboreal vegetation
		Mean NDVI of silviculture
		Mean NDVI of grass or cultivated area
Environment	Climate	Historical mean wind speed (m/s)
Network	Accessibility	Distance of accesses
	Impact	Number of consumer units
		Proportional length of trunk in the electricity network
		Proportional length of bare cable in the electricity network
		Network voltage
	Performance	DEC-vegetation
		FEC-vegetation
		Accidental interruptions shorter than 3 min

).

The vegetation criterion was subdivided into three sub-criteria related to the amount of vegetation in the analyzed area, the proximity of the vegetation in relation to the network segments, and the condition of the vegetation (health). The environmental criterion considered only the historical climatic conditions in the vicinity of the analyzed network segment. Finally, the network criterion was subdivided into two sub-criteria for analyses in urban areas and three sub-criteria for rural areas. For urban areas, the sub-criteria related to the impact and the performance of the network in relation to the DEC and FEC indicators were considered. For rural areas, in addition to the sub-criteria considered for urban areas, the accessibility of the network through highways or other types of access was also considered. Each sub-criterion was related to one or more state variables that could be calculated for the prioritization analysis.

2.4.2. Vegetation and Auxiliary Data Processing

As presented in Section 2.4.1, the indicator developed to identify areas with high priority for vegetation management, PIVM, considered criteria that were related to the vegetation itself, as well as criteria related to the environment and the network. All indicators related to the vegetation criterion were obtained from the results of PlanetScope image classification, in this case, the classification results with the best performance were used for each of the analyzed sites. For this, we only considered the pixels classified as vegetation (arboreal vegetation, silviculture, and grass or cultivated) being the pixels classified as shadows masked out prior to the analysis.

The definition of the area of interest for the analysis of the vegetation in the vicinity of the networks was carried out from the processing of the georeferenced network containing the segments of low and medium voltage available in the Geographical Database (BDGD) provided by COPEL. Considering the information obtained in the meeting held with the COPEL team, buffers were defined for the analysis of the vegetation around the networks. The definition of the buffer zones took into account the distances around the network that must be managed, where vegetation can interact directly with the cables of the electrical network. Specifically in the case of Paraná State, State law No. 20.081 [21] establishes parameters for the planting and management of vegetation around energy distribution networks. It is worth saying that the suppression of native vegetation, even within the buffer zone, depends on authorizations from the competent environmental agency, in accordance with current legislation.

• Urban areas (S1):

  o

  Medium- and high-voltage networks: A 2 m buffer for each side of the network segment;

  o

  Low-voltage network: A 1 m buffer for each side of the network segment.

• Rural areas (S2, S3 and S4):

  o

  Medium- and high-voltage networks: A 5 m buffer for each side of the network segment.

Historical wind intensity data were obtained from the ERA5-Land product provided by the European Center for Medium-Range Weather Forecasts (ECMWF). The ERA5-Land product is publicly available for the period 1950–current, providing regular grids (9 × 9 km²) with hourly values of hydrometeorological variables on the Earth’s surface. More information about the ERA5-Land product can be found at the ECMWF website, available online: https://www.ecmwf.int/en/era5-land (accessed on 4 January 2022). The historical series of hourly data were processed using the Google Earth Engine platform, available online: https://earthengine.google.com/ (accessed on 14 January 2022), generating numerical grids with maximum and median values of hourly wind intensity (m/s) between 1980 and 2021 for the four areas of interest. The resulting wind intensity in each network segment was obtained by crossing the generated grid and the vector containing the segments of the low- and medium-voltage distribution system.

Most of the operational variables were obtained using data available in the BDGD, and provided by the COPEL team. All of the data provided were organized using a GIS environment, and the state variable values were calculated for each segment of the low- and medium-voltage distribution system.

Rural accesses were vectorized manually with the aid of a GIS, and by using PlanetScope images as data source. For the three rural sites considered in the pilot application (S2, S3, and S4), all accesses possible for visualization in the images (paved and unpaved) were vectorized to compose the distance analysis database. The distance between the access segments and the network segments was also performed with the aid of a GIS, with three ranges of distances of interest being defined for analysis: (1) Closer than 50 m; (2) Between 50 and 250 m; and (3) Farther than 250 m.

Figure 3 shows an example of environment and network data processed at the segment level, in this case, for Site S3.

2.4.3. Standardization, Weighting, and Aggregation

Standardization was performed to bring the state variables into a numerical scale of priority, ranging from the lowest priority to the highest priority. Standardization enables the comparison and combination of state variables for the unit analyzed.

Continuous variables were standardized based on maximum or minimum values, depending on the relationship between the variable’s value and the level of prioritization.

• Variables with values that are directly proportional to the degree of prioritization:

$$Valu{e}_{std}=\left(\frac{Value}{Valu{e}_{m\mathrm{ax}}}\right)\times 100$$

• Variables with values that are inversely proportional to the degree of prioritization:

$$Valu{e}_{std}=\left(\frac{Value}{Valu{e}_{m\mathrm{in}}}\right)\times 100$$

For both cases, values close to 0 indicate a smaller contribution of a given variable in the PIVM computation, and values close to 100 indicate a greater contribution of a given variable in the PIVM computation.

The standardization of categorical variables was performed using the rules described below:

• Distance of accesses:

• If distance < 50 m, Value_std = 0;
• If distance between 50 and 250 m, Value_std = 50;
• If distance > 250 m, Value_std = 100;

• Network voltage:

• If voltage = 34.5 kV, Value_std = 100;
• If voltage ≠ 34.5, Value_std = 0.

The weighting step consisted of assigning weights to each standardized state variable, sub-criteria, and criteria according to their degree of importance. In this case, we used the analytical hierarchy process (AHP) [22], which consisted of a paired comparison of the state variables, sub-criteria, and criteria. The definition of the weights considered the literature review and the experience of the COPEL team, which has already developed and operates a similar analysis for prioritizing vegetation management.

Table 5. Weights used to compute PIVM for urban areas.

Criteria	Criteria Weight	Sub-Criteria	Sub-Criteria Weight	State Variables	State Variables Weight
Vegetation	35	Quantity	10	Proportional area of arboreal vegetation	10
		Proximity	20	Proportional length of network intersecting with arboreal vegetation	20
		Condition	5	Mean NDVI of arboreal vegetation	5
Environment	5	Climate	5	Historical mean wind speed	5
Operational	60	Impact	22	Number of consumer units	12
				Proportional length of trunk	3
				Proportional length of bare cable	7
		Performance	38	DEC-vegetation	15
				FEC-vegetation	20
				Accidental interruptions less than 3 min	3

and

Table 6. Weights used to compute PIVM for rural areas.

Criteria	Criteria Weight	Sub-Criteria	Sub-Criteria Weight	State Variables	State Variables Weight
Vegetation	35	Quantity	10	Total area of vegetation (arboreal, silviculture, and grass)	8
				Proportional area of silviculture	2
		Proximity	20	Length of network intersecting with arboreal vegetation	18
				Proportional length of network intersecting with silviculture	2
		Condition	5	Mean NDVI of arboreal vegetation	1.67
				Mean NDVI of silviculture	1.67
				Mean NDVI of grass or cultivated area	1.67
Environment	5	Climate	5	Historical mean wind speed (m/s)	5
Operational	60	Accessibility	5	Distance of accesses	5
		Impact	25	Number of consumer units	10
				Proportional length of trunk	4
				Proportional length of bare cable	3
				Network voltage	3
		Performance	35	DEC-vegetation	16
				FEC-vegetation	16
				Accidental interruptions less than 3 min	3

present the weights defined for each level of the multi-criteria analysis.

The PIVM was obtained using a weighted linear combination aggregation of criteria. It consisted of multiplying the weights from the weighting of each standardized factor and then summing them.

$$\mathrm{PIVM}=\frac{\left(35\times Vegetation\right)+\left(5\times Environment\right)+\left(60\times Operational\right)}{100}$$

The scale level defined to calculate most of the variables was the network segment, which was later aggregated to the feeder level, and which corresponds to a set of segments.

3. Results and Discussion

3.1. Image Classification and Vegetaion Mapping

Considering the four pilot sites, the two algorithms tested (ANN and SVM), and the hyperparameter configurations, a total of 103 calibration and validation tests were performed. In general, the ANN algorithm performed better than the SVM algorithm for the classification of PlanetScope images in all sites selected for the project (Table 7). Only for S2 did the SVM algorithm present a better performance than the ANN. For S1, the best classification performance was obtained using the ANN algorithm with a training threshold of 0.9, a learning rate of 0.2, momentum 0.7, 1 hidden layer, and 10,000 interactions (Figure 4). For S2, the classification with the best performance was obtained with an SVM algorithm, using a radial basis function as the kernel and setting the penalty parameter to 0.7. For S3, the best result was obtained using the ANN algorithm with a training threshold of 0.7, a learning rate of 0.2, momentum 0.7, 1 hidden layer, and 10,000 interactions. Finally, for S4, the best classification was obtained using the ANN algorithm with a training threshold of 0.7, learning rate of 0.3, momentum 0.9, 2 hidden layers, and 15,000 interactions. Figure 5 shows an example of classification, with best result being obtained for the rural sites selected for the pilot application (S3).

Table 7. Summary of validation metrics for the classifications with best performance.

Site	Classifier	Kappa	Overall Accuracy	Conditional Kappa Grass and Cultivated Areas Class	Conditional Kappa Vegetation Class
S1	ANN	0.883	0.940	0.715	0.918
S2	SVM	0.838	0.913	0.521	1.000
S3	ANN	0.925	0.959	0.929	0.908
S4	ANN	0.829	0.892	0.572	0.915

Table 7. Summary of validation metrics for the classifications with best performance.

Site	Classifier	Kappa	Overall Accuracy	Conditional Kappa Grass and Cultivated Areas Class	Conditional Kappa Vegetation Class
S1	ANN	0.883	0.940	0.715	0.918
S2	SVM	0.838	0.913	0.521	1.000
S3	ANN	0.925	0.959	0.929	0.908
S4	ANN	0.829	0.892	0.572	0.915

The comparison between the final results of the classifications obtained in this study and those presented in recent works published in the literature shows that the performance of the classifications using the PlanetScope images and the ANN was similar or superior to those obtained with medium spatial resolution images (e.g., MSI Sentinel-2 and OLI Landsat-8) [23,24,25]. In addition, the comparison with the results obtained using images with very high spatial resolution, for example, using WorldView-2 images [26], indicate a very similar performance by PlanetScope images for mapping vegetation in densely urbanized areas, such as S1, even when using the pixel-by-pixel classification approach.

Vegetation Analysis within the Buffer Zone

The data processing for the four sites selected for the pilot application resulted in a total of 3665 km of network for analysis (54,581 segments), with 63% of the length being intersected by vegetation (arboreal and silviculture). Considering the rules adopted for the delineation of the buffer zone, the area analyzed comprised 33.27 km², with 62% showing some type of vegetation detected using the PlanetScope images. Among the four sites analyzed, S3 presented the largest area of vegetation mapped within the buffer zone (8.85 km²) and the longest network intersected by vegetation (213 km). In turn, S1 showed the smallest area of vegetation within the buffer zone, only 0.41 km². In relative terms, S4 presented the highest proportion of the buffer zone with vegetation (74%), and S2 presented the highest proportion of the network length intersected by vegetation (18%). On the other hand, S1 was the site that presented the smallest proportion of vegetation within the buffer zone (18%) and the smallest proportion of the network intersected by vegetation (9%).

Detailed analysis showed that at S1, 66% of the vegetation mapped within the buffer zone was classified as arboreal vegetation class, and 34% as grass class. Of the total of 19,332 network segments in S1, 67% did not present any vegetation mapped within the buffer zone, and 74% did not present any intersection with the mapped vegetation. It is worth noting that 50% of the vegetation mapped in this site is concentrated around less than 1% of the existing network segments in this site (1325 segments).

At site S2, 58% of the vegetation mapped within the buffer zone was classified as grass or cultivated, 40% as arboreal vegetation, and 2% as silviculture. Of the total of 9036 network segments analyzed at S2, only 9% did not present any vegetation mapped within the buffer zone, and 48% did not present any intersection with the mapped vegetation.

At S3, 42% of the vegetation mapped within the buffer zone was classified as grass or cultivated areas, 49% as arboreal vegetation, and 9% as silviculture. Of the total of 20,238 network segments in this site, only 13% did not present any vegetation mapped within the buffer zone, and 54% did not present any intersection with the mapped vegetation.

Finally, at S4, 63% of the vegetation mapped within the buffer zone was classified as grass or cultivated area, 28% as arboreal vegetation, and only 9% as silviculture. Of the total of 5975 network segments in this site, only 2.4% did not present any vegetation mapped within the buffer zone, and 39% did not present any intersection with the mapped vegetation.

3.2. Priority Analysis and Critical Areas

For the urban site (S1), 556 km of network were analyzed, resulting in a buffer zone area of 2.2 km². Of the total length of the existing network at S1, approximately 9% intersects with vegetation, and the mapped vegetation area was 0.3 km². The PIVM was calculated and grouped at the feeder level, resulting in 93 feeder units for analysis (Figure 6a).

The PIVM value ranged between 14.53 and 55.99, with 19 feeders classified as high management priority, 27 as medium management priority, and 47 as low management priority. Taking into account the proposed classification for the PIVM, 170 km of the network analyzed at S1 were classified as high priority, which was 31% of the length analyzed at S1. The total area of the vegetation mapped around the feeders that was classified as high priority was 0.12 km², with 29 km of network intersecting with the vegetation. In turn, feeders with PIVM classified as medium priority accounted for 201 km of network, which was 36% of the total length analyzed at S1. The total area of vegetation mapped around the feeders that was classified as medium priority was 0.1 km², with 23 km of network intersecting with the vegetation. Finally, feeders with PIVM that were classified as low priority accounted for 185 km of the network, which was 33% of the total length analyzed. The total area of the vegetation mapped around the feeders that was classified as low priority was only 0.06 km².

For the rural sites (S2, S3, and S4) 3100 km of network were analyzed, resulting in a buffer zone area of 31 km². Of the total length of the existing network in these sites, approximately 15% intersects with vegetation, and the mapped vegetation area was 20 km², 65% of the buffer zone. As for the application at S1, the PIVM for rural areas was calculated at the feeder level, resulting in 26 units of analysis (Figure 6b).

The PIVM value ranged between 9.345 and 59.268. Of the total number of feeders analyzed, three were classified as being high priority for management, 13 as medium priority, and 10 as low priority. Considering the proposed classification for the PIVM, 2296 km of rural network were classified as high priority, 74% of the total analyzed. The total vegetation area mapped around the feeders that was classified as high priority was 15 km², with 355 km of network intersecting with the vegetation. In turn, feeders with PIVM that were classified as medium priority accounted for 709 km of the network, 23% of the total analyzed. The vegetation area mapped around the feeders classified as medium priority was 4 km², with 63 km of network intersecting with the vegetation. Finally, feeders with PIVM that were classified as low priority accounted for 95 km of the network, or 3% of the total. The area of vegetation mapped around the feeders that was classified as low priority was 1 km², with 16 km of network intersecting with the vegetation.

Figure 7 shows the spatial distribution of PIVM classified according to priority degree for the four sites evaluated in this pilot application.

4. Conclusions

This work aimed to develop indicators to identify stretches of electricity distribution networks with high priority for vegetation management, in order to improve the operational performance of the network and to reduce the frequency and duration of interruptions in the energy supply. The indicators were developed using attributes extracted from vegetation mapped in the area of interest. For this purpose, we tested two artificial intelligence algorithms, support vector machine (SVM) and artificial neural networks (ANN), to automatically identify different classes of vegetation, using PlanetScope images as input.

The results showed that, for our case, the ANN algorithm presented better results for the vegetation classification using PlanetScope images when compared to the results obtained with the SVM algorithm. In addition, PlanetScope images showed good performance and have a great potential for the automatic detection of different types of vegetation in landscapes with different complexities.

The comparison with other results published in the literature showed that the classification of vegetation using PlanetScope images and the ANN algorithm presented similar and, in some cases, higher performances compared to other approaches and algorithms that have been already used, even when compared with very high resolution images (with sub-meter resolution).

The indicators developed for the prioritization of tree pruning in urban areas and mowing in rural areas have a great potential to optimize efforts related to vegetation management in the context of energy distribution networks with high capillarity. Through the use of geographic intelligence, the PIVM is able to identify feeders of higher criticalness, detailing the spatial distribution of vegetation. It is worth mentioning that in this study, aspects related to vegetation functions and other ecosystem relationships were not considered for the purpose of prioritizing management areas. These aspects can be investigated and included in the prioritization analysis in future studies, using multiple sources of data and images to characterize the ecosystem aspects of vegetation within the buffer zone and its surroundings.

The characteristics of the images collected by nanosatellites, such as high spatial resolution allied high frequency of acquisition, allow for the scalability of the methodology to large areas (e.g., the entire COPEL’s distribution network). Moreover, the methodology developed has great potential for becoming an automatic application, with vegetation and other state variables being updated on a regular basis. Considering COPEL’s specific needs, the results of this R&D project are expected to contribute to the optimization of the use of human and financial resources, through the aggregation of geographic intelligence in decision making.

Finally, the PIVM can be adapted to prioritize vegetation management in another context, such as energy transmission systems and beside roads, green areas, and squares in urban areas.