The study was carried out in Îles-de-Boucherville National Park (Quebec, Canada), which covers an area of approximately 8.14 km² and consists of five islands in the Saint Lawrence River between the Island of Montreal and the municipality of Boucherville (Figure 1) (Laliberté et al., 2006). The area has experienced a long history of heavy agricultural use, beginning with the arrival of the first settlers towards the end of the 17th century (Giroux, 1986) and continuing into modern times with the presence of privately-owned agricultural areas on Île de la Commune and Île Grosbois.

Îles-de-Boucherville National Park was officially created in 1984, when the Quebec Ministry of Forests, Wildlife and Parks designated the islands a protected area, which led to the subsequent abandonment of multiple agricultural fields. Herbaceous and shrub vegetation soon became the dominant land cover type within the park. More than 450 plant species are present, across both terrestrial and aquatic ecosystems (Société des établissements de plein air du Québec, 2020). Commonly found species include Cornus sericea (red osier dogwood), Asclepias syriaca (common milkweed), Solidago gigantea (giant goldenrod), and Phalaris arundinacea (reed canary grass). Young forested areas are scattered throughout the park, mainly comprised of Populus deltoides (eastern cottonwood) and Fraxinus pennsylvanica (green ash). Approximately 18 ha of established mature forest (∼70–90 years old) is located on Île Grosbois, and is the only mature forest within the park (Ross, 1990; Laliberté et al., 2006).

Commonly found throughout the park, in both terrestrial and aquatic ecosystems, is the invasive common reed P. australis subsp. australis (Figure 2). Extensive stands of this invasive reed can be found in open fields, along hiking trails, in former agricultural ditches, along riverbanks and in river channels: one of the largest stands of Phragmites in the province of Quebec is located along the western edge of the park boundary in the Courant Channel (Hudon et al., 2005). Controlling Phragmites as part of the restoration of former agricultural fields back to natural habitat is a priority issue in the 2017–2022 Îles-de-Boucherville National Park Conservation Plan, set forth by the Société des établissements de plein air du Québec (SEPAQ) (Société des établissements de plein air du Québec, 2020).

Phragmites demonstrates an affinity for disturbed ecosystems such as former agricultural fields or wetland environments, particularly those that have been enriched with nutrients such as nitrogen due to agricultural or other anthropogenic sources (Belzile et al., 2010). Methods of spread include seed dispersion or by the transport of rhizome fragments to new areas, either via environmental mechanisms (e.g., wind and wave action) or via anthropogenic means. The plant is characterized by long, flat, green leaves that grow from a hollow stem. A bushy panicle, or seed head, will develop in late summer and will persist into winter after the seeds have been released. This seed head is distinct, and can vary in color depending on its maturity: common colors include brown/tan or deep purple (Figure 2). Phragmites is considered an indicator of ecosystem disturbance due to its aggressive nature and tendency to grow in dense monotypic stands up to 3 m tall that consist of an extensive network of stolons and rhizomes. These dense stands can prevent the growth of other surrounding vegetation, or even outcompete existing vegetation, effectively reducing plant biodiversity in the area (Saltonstall, 2002). This loss of species richness and the associated potential loss of biodiversity or extinction is one of the primary concerns regarding Phragmites (Chambers et al., 1999; Cronk and Fuller, 2001). These characteristics and tendencies of Phragmites make removal of established stands difficult, especially once it has grown over an extensive area. Efforts to manage and eradicate established populations of Phragmites can be difficult and typically require extensive physical labor and considerable investments of time and money (Martin and Blossey, 2013). There are several chemical options for controlling Phragmites, such as broad-spectrum herbicides that have documented effectiveness (Avers et al., 2007). However, these chemicals are non-selective and could damage native species, and are not always permitted in certain areas such as wetlands. Mechanical methods of control (e.g., hand cutting or mowing) can be effective in the short term, but are less likely to be successful in the long term without combining them with other control methods such as herbicides or controlled burning (Getsinger et al., 2006; Avers et al., 2007). Mechanical methods also risk transferring rhizome fragments or seeds, which could cause new invasions.

Here we present an overview of the methodology conducted for this research, as shown in the flowchart in Figure 3. These steps are discussed at length in the following sections.

A total of 319 ground truth points were collected across the extent of Îles-de-Boucherville National Park to provide a high-accuracy dataset (<1 m spatial error) for training and validation of the target detection algorithm. A detailed description of the methodology used to collect the points, as well as the accuracy assessment and available data products are described in Elmer and Kalacska (2021). The ground truth points were randomly separated into 160 points for training the target detection algorithm and 159 points for validating the accuracy of the algorithm’s output. The training set consisted of 79 points with no Phragmites present and 81 points where Phragmites was present, while the validation set consisted of 83 points with no Phragmites present and 76 points where Phragmites was present. The ground truth points were split evenly in this way to avoid creating a bias in the training/validation sets. Additionally, each set of ground truth points represented a fairly equal geographic spread of points across the park extent. This was an important consideration because the park is not uniform in ground cover: the upper portion of the park is dominated by agricultural fields, while the lower portion of the park is characterized by natural fields and areas of shrubby vegetation.

The training set of 160 points was used as the basis for creating the samples of target spectra that served as input for the target detection algorithm. In order to increase the number of target spectra for better comparison during the target detection analysis, the 160 training points were expanded to include nearby pixels that fell within the appropriate land cover types. For the target spectra, samples were taken from measured Phragmites stands. Non-target spectra were also collected to provide examples of spectra to reject during target detection: this included water, agricultural fields characterized by dry soil and dried out vegetation, asphalt roads, and other vegetation that was not Phragmites.

Airborne hyperspectral imagery (HSI) was acquired on July 08, 2019 as part of the Canadian Airborne Biodiversity Observatory project (Arroyo-Mora et al., 2019) and consisted of five flight lines collected over the extent of Îles-de-Boucherville National Park (Table 1). The aircraft used as the sensor platform was a DeHaviland Twin Otter owned and operated by the National Research Council of Canada’s Flight Research Laboratory (NRC-FRL, Ottawa, ON, Canada). The hyperspectral sensor used to acquire the imagery was a Compact Airborne Spectrographic Imager 1500 (CASI-1500, hereafter CASI) (ITRES Ltd., Calgary, AB, Canada). The CASI collects up to 288 spectral bands between 375 and 1,054 nm, with a field of view of 39.9° and 1,498 across-track pixels.

The five individual CASI flight lines underwent pre-processing using software modules developed by the sensor’s manufacturer before they were used for analysis, as described by Soffer et al. (2019). First, a calibration was made to account for the effects of temperature and pressure shifts in the internal sensor alignment. Next, a spectroradiometric calibration was performed that removed the estimated signal offsets inherent in the recorded digital values from various sources (dark current, electronic offset, internal scattered light, second order scattered light, and frame shift smear) and converted the data to units of spectral radiance (uW cm⁻² sr⁻¹ nm⁻¹). The flight lines were then atmospherically corrected using ATCOR-4 (Richter and Schläpfer, 2011). Finally, the flight lines were geometrically corrected (2 m resampled pixel size) and mosaicked such that pixels with a viewing angle closest to nadir were prioritized and preserved in the areas of overlap between the flight lines (Figure 1). The reported positional accuracy of the HSI is an average error of 0.91 in easting and 0.67 m in northing (Elmer and Kalacska, 2021).

The HSI mosaic was spectrally subset to remove bands with low signal level and high levels of noise due to residual atmospheric absorption features: these consisted of wavelengths below 450 and greater than 800 nm, respectively. The mosaic then underwent spectral polishing to remove high frequency noise using the Empirical Flat Field Optimal Reflectance Transformation (EFFORT) polishing tool (Boardman, 1998, December). One of the challenges of HSI is its large dimensionality and high correlation between bands, which can make target detection more time consuming and the results less accurate if all bands are used (Ready and Wintz, 1973; Landgrebe, 2002). In order to reduce the dimensionality of the data and isolate the signal from noise, the spectrally subset and polished mosaic underwent a Minimum Noise Fraction (MNF) transformation (Green et al., 1988). The output eigenvalues indicated the approximate number of bands that contained meaningful information, and visual assessment of these bands for spatial coherence determined that the first 13 MNF bands contained the majority of the relevant signal. The output MNF image was subset to these first 13 bands, which became the input image for target detection.

Many remote sensing applications assume that the contribution to a pixel’s signal is uniform across the given pixel size for the data product (i.e., a pixel with a 2 m × 2 m spatial resolution contains only the signal(s) from the materials that fall within the pixel’s footprint). In reality, this assumption is not true and the contribution of materials to each pixel is spatially non-uniform according to its net point spread function (PSF) (C. Huang et al., 2002). For example, the imagery acquired using the CASI for this study has a net PSF that extends approximately 1.88 m in the along-track direction and 1.85 m in the cross-track direction, with a true pixel size of 1.64 m in the across track and 1.63 m in the along track and only 76% of the pixel’s signal originates from materials within the given pixel extent, while the remaining 24% is contributed by materials in neighbouring pixels (Inamdar et al., 2020). Consideration of the specific characteristics of a sensor’s geometry is therefore critical to understanding the actual contributions to a pixel’s spectrum, which has important implications for the analysis and interpretation of data products derived from such sources. This characteristic of the HSI was accounted for during analysis using the methods described in Target Detection Analysis below.

Open access LiDAR-MCH (mean canopy height) data were used to investigate the height profiles of established Phragmites stands within the study area, and to mask out materials taller than a specified height so that they would not be analyzed as part of the target detection. The LiDAR data was acquired by the Quebec Ministry of Environment, and Fight Against Climate Change, as part of the province-wide collection of open-access LiDAR data (available at https://www.foretouverte.gouv.qc.ca/). The two tiles that covered the area of interest (31H11SO and 31H11NO) were acquired in 2018 with a point density of 4 points/m². The data are provided as rasterized tiles with a pixel size of 1 m. The LiDAR tiles were mosaicked, clipped to the area of interest, and resampled to a 2 m pixel size (Figure 4).

Tall, forested areas of the park could be disregarded from the target detection analysis, as any Phragmites covered by tree canopy would not be visible in the HSI and therefore not reliably detectable. However, Phragmites commonly grows up to 6 m in height (Goodrich and Neese, 1986) so in order to avoid inadvertently masking out Phragmites stands on the taller end of the growth profile, the canopy height of 158 Phragmites ground truth points were extracted from the LiDAR-MCH data. The resulting distribution of Phragmites’ heights are shown in Figure 4. As there were no sampled areas of Phragmites taller than 6 m, the threshold for the height mask was set to >6 m.

The 13 band MNF image was the input for target detection based on the Spectral Angle Mapper (SAM) (Boardman, 1993; Kruse et al., 1993). The Phragmites ground truth points (training subset) were separately provided as sample target spectra. Additionally, the sample spectra of non-target materials (water, fields, roads, and non-Phragmites vegetation) were specified as targets to reject. SAM treats both the reference target (r) and the unknown spectra (t) as vectors in order to calculate the angle in radians between the spectra as a method of determining spectral similarity (Boardman, 1993) (Eq. 1). c o s − 1 ( ∑ i = 1 n b t i r i ( ∑ i = 1 n b t i 2 ) 1 / 2 ( ∑ i = 1 n b r i 2 ) 1 / 2 ) (1) where (nb) is the number of bands. The more similar an unknown spectrum is to the reference spectrum, the smaller its angle: if the calculated angle falls within the user-defined angle threshold, the associated spectrum is labeled as the target. The SAM algorithm is insensitive to differences in illumination and albedo as it only uses the vector direction and not the vector length to determine spectral similarity (Kruse et al., 1993). Therefore, it can be particularly useful for scenes that are not acquired under uniform illumination conditions, such as the HSI used here: over the period of acquisition, the solar azimuth angle changed between 129.95 and 144.83° and the solar elevation angle changed between 59.07 and 63.34° (Table 1). These changing illumination conditions can influence the magnitude of the spectral reflectance for vegetation, as reflectance will typically increase with increasing solar zenith angle due to the effects of the bidirectional reflectance distribution function that is unique to each material (Ranson et al., 1986; Gross et al., 1988).

For the Phragmites class, the SAM threshold was set to 0.65 radians. This value was determined by examining the results and selecting the threshold value that resulted in the best coverage of known target areas (i.e., known extensive, homogeneous stands of Phragmites) while minimizing the extent of erroneously identified pixels.

Due to the positional error of the HSI following geocorrection, and the contribution of materials located in neighbouring pixels determined by the net PSF [as described in Airborne Hyperspectral Imagery (HSI)], a buffer with a distance of 2.25 m was determined in order to account for areas of uncertainty. This uncertainty buffer therefore determined the minimum mapping unit (MMU) for the Phragmites map and polygons with an area smaller than 16 m² (four pixels) were removed. Using the uncertainty buffer, the sections of the target polygons that were within 2.25 m of trees (determined using the LiDAR-MCH > 6 m mask) were removed due to the potential contribution of the spectra of tree components such as leaves, branches, and tree shadows to the neighbouring pixels. The 2.25 m buffer was also applied to each of the target polygons. The Phragmites polygons and their buffers represent the final map of identified Phragmites within Îles-de-Boucherville National Park, as shown in Figure 5.

To account for the associated positional error of the validation ground truth points (Elmer and Kalacska, 2021), a 1.75 m error buffer was calculated for each of the 159 points. To carry out the validation, the points were analyzed to determine if their respective error buffers overlapped with the uncertainty buffers associated with the Phragmites polygons.

As a way of comparing the accuracy of the algorithm-based target detection methods to conventional visual interpretation conducted by visual interpreters, an online tool was used in order to determine the ability of the visual interpreters to corrected identify Phragmites from aerial imagery. The “human accuracy” using the same set of validation points was compared to the calculated accuracy of the SAM target detection methodology. Fifteen-meter buffers were drawn around each of the 159 validation ground truth points and displayed on subsets of Satellite Streetview imagery (DMTI Spatial ULC, Richmond Hill, ON, Canada). This imagery consists of pansharpened (60 cm) satellite imagery acquired in 2012. Similar to the Picture Pile tool (https://geo-wiki.org/games/picturepile/) implemented by Danylo et al. (2021), the Satellite Streetview subsets were arranged and displayed online in Survey Legend (Malmö, Sweden) (Figure 6). A total of 10 visual interpreters examined the 159 ground truth points and determined if Phragmites was present within the buffers or not. Each interpreter was first shown several example images of Phragmites and non-Phragmites images and then asked to examine all 159 images, which allowed for direct cross-comparison of each individual image.

It is important to note that the high-resolution imagery was collected in 2012, while the validation data was collected in 2019. Therefore, there may be some inconsistencies in the appearance of land cover types between the high-resolution Satellite Streetview images and the HSI used for target detection. For example, one validation point was collected in a grassy field in 2019, while in the 2012 imagery it appears as a cultivated agricultural field. Each high-resolution image was examined to ensure that there was no misrepresentation of the validation points to the visual interpreters (i.e., all ground truth points indicating no Phragmites appeared with no Phragmites in the 2012 imagery, and vice versa for the Phragmites points).

Investigation of the input spectra used to train the target detection algorithm revealed that the brown Phragmites seed heads present within the stands were spectrally distinct from the leafy green portion of the plant (Figure 7). In order to include this part of the plant in the target detection results, which could be the dominant material present in some Phragmites pixels, additional target spectra were collected from the brown seed heads and used to train the target detection algorithm on this separate class. It was also noted that there were considerable spectral similarities and overlap between the brown seed heads and the agricultural field spectra, which was comprised of dry soil and vegetation: this similarity can be seen in Figure 7. Comparison of the Phragmites spectra to that of non-Phragmites vegetation revealed that the spectra were distinct from each other and therefore more separated by the target detection algorithm.

A total of 2,037 separate stands were detected. The final map showing the extent of detected Phragmites within the park is shown in Figure 5. The total core area of detected Phragmites in the park was 26.74 ha (0.267 km²), which represents approximately 3.28% of the park’s total area of 814 ha (8.14 km²). The average (±σ) polygon size of detected Phragmites is 131.29 ± 571.04 m² and the frequency distribution of the area of Phragmites polygons is shown in Figure 8. When accounting for the uncertainty indicated by the 2.25 m buffer around each Phragmites polygon, the total area of detected Phragmites was 59.17 ha (0.591 km²), which represents approximately 7.26% of the entire park area. In order to evaluate the performance of the target detection methodology, the calculated confusion matrix is shown in Figure 9.

The overall accuracy of the final map was calculated as 84.28%, with 134 of the total 159 validation points properly identified (true positives). Of the 25 misidentified points, 18 false negatives occurred where the Phragmites stand was not identified correctly. Investigation of these 18 points revealed that nine points occurred where Phragmites stands were a sub-pixel target or were smaller than the MMU (4 m²) and were therefore removed. Similarly, seven ground truth points were collected at the edge of a Phragmites stand where the contribution of Phragmites to the corresponding pixel’s signal was too low to be correctly identified, resulting in false negatives. . For example, the validation point ID = 66 was located 6.58 m from the nearest pixel of identified Phragmites: at this distance, the materials at that point would not contribute any substantial signal to the closest detected pixel of Phragmites, based on the net PSF. The remaining two Phragmites points that were misidentified occurred in Phragmites stands that were predominantly defined by the brown seed heads and brown stems of Phragmites along the margins of water bodies (also false negatives). There were no ground truth points collected for aquatic Phragmites due to lack of accessibility. This reveals the limitations of extrapolating the terrestrial conditions where Phragmites was sampled in situ to vastly different environmental conditions such as aquatic stands. Examples of these three scenarios where Phragmites was misidentified are shown in Figure 10.

For this study, the small sample size of target pixels meant that the calculation of a receiver operating characteristic (ROC) curve and associated metrics would not be appropriate (Cisz and Schott, 2005, June). Therefore, using the guidelines regarding the most commonly used accuracy statistics as given by Foody (2002), the sensitivity, specificity, misclassification rate, errors of commission and omission, negative predictive value and precision were calculated and are shown in Figure 9. With a high specificity of 91.57% and a low error of commission of 8.43%, the target detection algorithm distinguished areas that did not contain any Phragmites (Figure 9). The calculated error of omission of 23.68% indicates that there are areas of Phragmites that were not properly identified. These areas most likely occur at sub-pixel Phragmites stands or at the edges of Phragmites stands where the concentration of Phragmites is low (as described above). Overall, the calculated statistics show that the target detection methodology applied in this study had a high overall accuracy and is appropriate for detecting Phragmites within Îles-de-Boucherville National Park. However, a limitation of this method is that small, incipient stands of Phragmites are less likely to be detected. Figure 11 illustrates the differences in the range of SAM detection values for the Phragmites target pixels versus the background pixels. The SAM threshold for Phragmites was set to 0.65 radians, so the associated SAM pixel values fall within that threshold. For the background pixels, there is some overlap with the range of values for Phragmites but the majority of background pixels are associated with pixel values that are higher than 0.65. This shows that there is a separation of the SAM values associated with Phragmites from the values associated with background pixels, allowing for detection of the appropriate target pixels.

A total of 10 visual interpreters assessed the 159 validation ground truth points and surrounding areas within the 15 m buffers. Eight of the interpreters had previously been to Îles-de-Boucherville National Park, and six had some prior knowledge of Phragmites specifically, either from field work or aerial imagery. All visual interpreters were shown several example images of Phragmites and non-Phragmites prior to assessing the validation ground truth points. The overall accuracy of the visual interpretation was 69.18% for Phragmites. For all calculated accuracy statistics (Figure 12), the visual interpretation method underperformed when compared to the target detection results. The human accuracy overall was lower, with higher rates of misclassification (30.82%), error of commission (21.69%), and error of omission (40.79%). Similarly, the sensitivity (59.21%) and specificity (78.31%) were lower than the target detection results by 17.11 and 13.26%, respectively.

Of the same 18 validation points where Phragmites was incorrectly identified by the target detection algorithm, only two of those points were correctly identified by all 10 visual interpreters. Three of the points were completely misidentified (i.e., none correctly identifying Phragmites in the image). The remaining 13 points of Phragmites were not identified with complete agreement amongst visual interpreters; for seven points >50% of correctly identified Phragmites, for five points <50% correctly identified Phragmites, and at one point where 50% of the visual interpreters identified Phragmites correctly the other 50% did not. The visual interpreters were assessing higher spatial resolution imagery (60 cm) than the HSI (2 m), which might have provided an advantage as the Phragmites in the Satellite Streetview images would potentially be more distinguishable due to the finer spatial resolution (Figure 13). However, the visual interpreters still misidentified some of the same points that the target detection algorithm missed/incorrectly identified, where Phragmites was present in reality but was identified as no Phragmites present: this highlights the shortcomings of visual interpretation of Phragmites using aerial imagery, even amongst experts.