Acacia longifolia (Andrews) Willd. is easily identifiable both during the flowering season due to the bloom of yellow flowers as well as during the fruitification when its seed pods begin to brown and remain on the branches. It can develop as a shrub (often up to 4 m in height), when occurring in primary dunes, and as small tree (up to 8 m in height) when occurring more inland, also below the canopy of other species (e.g., maritime pines) (Cesar de Sa et al., 2017; Invasoras, 2017). During fieldwork we observed a high variability in flower patterns amongst plants depending on its morphology (tree or shrub) and on the position in the dune system (Figures 1B,C). For example, while A. longifolia generally produces flowers throughout the broad terminal part of the branches (Figure 1B), it occasionally produces less flowers in the tip of the branches when in the primary dunes (Figure 1C); and produces much lesser amount when under pine canopies and often closer to the tip of the branch (Figure 1D).

This study was carried out on seven sites distributed along ca. 80 km of the central-northern coast of Portugal (Figure 2 and Table 1). Sites were selected in order to include the most common coastal habitats invaded by A. longifolia: (1) several plant communities in sand dunes (SD), including primary dunes, interdunes and secondary dunes (Cesar de Sa et al., 2017); maritime pine occurs in these areas but generally does not dominate (SD1, SD2, SD3, SD4, SD5); and (2) pine forests plantations (hereafter pine forests, PF), where maritime pine dominates and A. longifolia occurs in the understory (PF1, PF2).

The dunes of Quiaios (SD1 and SD2, Figure 2) are a coastal strain of primary and secondary sand dunes with associated pine forests that extends inland. These sites are characterized by a mosaic of several well-established plant communities, including native perennial herbs (e.g., Ammophila arenaria and Corynephorus canescens) alternating with chamaephytic shrublands (e.g., Corema album, Cistus salviifolius, Artemisia campestris). Between the different plants assemblages there are large areas dominated by invasive plant species (e.g., Carpobrotus edulis, Arundo donax, Acacia saligna, and A. longifolia).

Tocha secondary sand dunes (SD3, Figure 2) are characterized by a singular sand dune string conformation, perpendicular to the shore line. Like Quiaios dunes (SD1 and SD2) it is mainly dominated by diverse well adapted littoral plant communities (with A. arenaria, C. album and A. campestris as the most common species) alternating with areas dominated by alien plant species, such as C. edulis and A. longifolia. SD3 burned in 1993 which triggered the expansion of A. longifolia, leading to the replacement of maritime pine stands by monospecific stands of A. longifolia. São Jacinto (SD4) and Torreira (SD5) sand dunes are located further north and are characterized by primary dunes highly invaded by A. longifolia. SD4 shows maritime pines and large patches of invasion by A. longifolia. In SD5 the maritime pines co-occur with secondary dune communities and are also mixed with invasive Acacia dealbata and Acacia melanoxylon. On the other hand, the pine forests of Quiaios (PF1 and PF2, Figure 2) include sparse plantations of Pinus pinaster with scattered patches of Halimium halimifolium and C. salviifolius. PF1 also has Eucalyptus globulus plantations. Both PF1 and PF2 show a varying degree of invasion by different Acacia species, including A. longifolia.

The field work was divided in two distinct main tasks performed as closely as possible in time (Table 1): (1) remote imagery acquisition by low cost digital cameras mounted on UAV, and (2) fieldwork sampling to quantify the number of spikes (hereafter referred as flowers) of A. longifolia produced per m² in randomly selected areas within each site.

Both field sampling and the UAV flights were performed in two seasons: (1) during flowering peak and (2) at the end of the flowering season of A. longifolia (Table 1), in order to test the accuracy in detecting flowers in both periods. By studying different phenological stages and habitats we expected to obtain a broad overview of the possibilities and limitations of this method.

Unmanned aerial vehicles flights were executed by Firemap SA¹, between March and April 2015, using the senseFly eBee² mounted, alternately, with one of two versions [regular RGB or filtered Colour-infrared (CIR)] of the Cannon IXUS/ ELPH, 16.0 MP, 6.16 mm × 4.62 mm sensor. Since two different cameras were operated with the same UAV, it was necessary to perform two flights over the same area and therefore the RGB and CIR acquisition had on average 1 h of difference between them. All flights were performed between 100 and 160 m above ground altitude depending on the meteorological and local terrain conditions and this resulted in different resolution of the final products (Table 1).

All the UAV data was pre-processed and delivered by Firemap SA using Structure-from-motion (SFM) available in Agisoft Photoscan version 1.2.0. Besides delivering the orthophotos of each location with 4 bands (RGB and NIR), the SFM by-products of Digital Surface Model (DSM) and Digital Elevation Model (DEM) were also delivered. The DEM was obtained by cloud-point semi-supervised classification methods which are readily available in Agisoft Photoscan. The Canopy Height Model (CHM) was obtained by subtracting the DEM from the DSM (Matese et al., 2017). This resulted in 5 different covariates used for the classification exercise: Red, Blue, Green, and NIR bands and the CHM.

Because the flights were performed between March and April (A. longifolia flowered very late in the Portuguese winter of 2015) and often near the coastline, it was not possible to deploy the UAV in all locations during an ideal timeframe (Table 1). The flights in each site covered approximately an area of 65 ha. Also, per area, a minimum of nine well spread-out ground control points were collected with sub centimeter accuracy using a Trimble GPS XH 6000 antenna receiver to assist in the processing of the final products (Benassi et al., 2017) as well as for alignment of the RGB and CIR acquisitions.

In each site, 30 areas of 50 m × 50 m were randomly selected for further field sampling which meant not all the pre-selected areas would have A. longifolia. Within each area with A. longifolia presence, a minimum of three 1 m² quadrats were sampled for flowering counts (Figure 1 and Table 1). A minimum of 60 1 m² quadrat samples were collected in each study area (Table 1). Due to the extremely high number of flowers present in 1 m², it was necessary to adapt the method of flower counting which would otherwise become unfeasible. So, within each quadrat, the flowers were counted in (at least) 3 smaller subplots of 30 cm × 30 cm (Figure 1A). Also, only the number of flowers in the first 30 cm starting from the tip of the branch were considered. These subplots were chosen in order to be representative of the flowering within that 1 m² quadrat (Figure 1). The smaller 30 cm × 30 cm subplots were then extrapolated to number of flower per m². This method was adapted from the methodology used by Morais and Freitas (2015) to provide an estimate of the number of flowers on a sampled area m². In total, data on 408 m² were collected. Each quadrat was georeferenced in the SW corner and oriented toward NE. A Trimble GeoExplorer 6000^® with post-processing differential correction was used to georeference the quadrat position to correctly match the field flower count sampling with the UAV observation.

To map the flower presence/absence in each image we used the Random Forest algorithm (Breiman, 2001). The Random Forest is a classification algorithm widely used in the field of Remote Sensing (Feng et al., 2015; Belgiu and Drăgut, 2016). It’s a non-parametric supervised classification which consists in an ensemble of Classification and Regression Tree’s to identify the target classes (Breiman et al., 1984; Breiman, 2001). The tree creation process consists in subsetting the training samples with replacement and using part of the data to train the tree while the remaining data is used to measure the prediction error (Belgiu and Drăgut, 2016). The final error estimate is known as the Out-of-Bag error (OOB) and is given by averaging the prediction error (Breiman, 2001; Belgiu and Drăgut, 2016).

To calibrate and validate the classification, 1000 points were randomly generated for each study area and visually identified as having flower presence or not (Figure 3). This is a common practice in Remote Sensing when very high resolution images are available and the distinction between the classes is clear (Feng et al., 2015; Lehmann et al., 2017; Müllerová et al., 2017b) as it is in our case (A. longifolia flowers vs. everything else, Supplementary Figure S9).

These 1000 points were then randomly separated into two groups, one used for calibration (70% of the points) and the remaining 30% used for validation which allowed assessing the Overall Accuracy and Cohen’s Kappa for each image (Lehmann et al., 2017; Müllerová et al., 2017b). To obtain the maps of A. longifolia flower cover/m² we applied a focal filter to the binary map where the window size varied according to the image resolution (Table 1), effectively resampling each image to 1 m².

The visual identification of the calibration/validation points was performed using ArcGIS^®10.2 while R^® 3.2.1 was used for: (i) data handling raster (Hijmans and van Etten, 2012), (ii) maptools (Bivand and Lewin-Koh, 2012), (iii) Random Forest classification (Liaw and Wiener, 2002) and (iv) Presence/Absence classification accuracy tests (Freeman and Moisen, 2008).

As we aim to evaluate the feasibility of using UAV imagery to monitor the expansion of T. acaciaelongifoliae (specifically through its effects on the flower cover of A. longifolia) we analyzed the statistical distributions of the flower cover per m² in each site.

In our study we followed the same approach used by Thorp and Dierig (2011). In controlled conditions these authors collected digital photographs of a small patch (0.125 m²) of Lesquerella fendleri (Gray) S. Wats and established a linear relationship between the area occupied by flowers in the photograph and the number of flowers present in the patch. We tested this method in the field for A. longifolia and by using UAV imagery. Acacia longifolia is structurally more complex than L. fendleri; as previously mentioned, it is a shrub/small tree which produces flowers in different heights while the latter is an herb.

To adapt the Thorp and Dierig (2011) method to imagery acquired by UAV we used the 1 m² quadrats positioned in the field (as shown in Figure 1) to guide the extraction of the flower cover in that same location. This allowed the establishment of the linear regression between flower cover and number of flowers counted in the field. If this linear relationship between area of flower cover and number of flowers counted in the field is successfully established, then this method would allow mapping the number of flowers of A. longifolia using UAVs using only the area covered by flowers.

The Random Forest classification obtained very high accuracy when detecting presence of flowers of A. longifolia (Table 2). The Overall Accuracy and Cohen’s Kappa values were always higher than 0.95 and 0.85, respectively, which indicates that the color of the flowers was very well detected by UAV. The lowest Kappa values corresponded to the sites PF1 and PF2 where A. longifolia invades as an understory species and the flower visibility was hindered both by total canopy cover as well as shadows. Considering the imagery acquired during the peak of flowering on dune systems (SD1, SD2, and SD3), SD2 and SD3 showed the highest Kappa values of all, while the OOB and the Overall Accuracy were very high for all locations. Blue band was the most important variable for all study areas except for SD2 (Supplementary Figure S10) which can be indicative of the contribution the yellow flowering to the classification accuracy as yellow is a result of low blue value and a combination of red and green (Sulik and Long, 2016). In opposition, the Near-Infrared band was the least important variable for all study areas except for P01 (Supplementary Figure S10) which can be indicative that this band is not helpful to distinguish yellow flowers from the background.

Clear flowering distribution patterns of A. longifolia were visible in images of all sites, showing specifically more presence of flowering in the dunes perpendicular to the shoreline (SD2 and SD3) (Figure 4) while SD1 did not have these types of dunes and also had more areas with pines. For the imagery acquired after the peak of flowering (SD4, SD5) the accuracy was lower than that of the best scenarios (0.973; 0.939 and 0.967; 0.896 respectively – Table 2). The flowering in these locations was harder to identify even during the extraction of the calibration/validation datasets (Figure 4 and Supplementary Figures S3, S4). While still within the flowering peak, it was possible to detect flowers on the edges of open areas of the pine forests (for example PF2 in Figure 4 and PF1 Supplementary Figure S4).

The flower cover maps, i.e., the total area of flower per m², showed that dunes perpendicular to the shoreline (SD2 and SD3) did have more flower cover (Figure 5). The post peak imagery (SD4 and SD5 Supplementary Figure S7) showed less flowering intensity throughout the whole image (although very high rates of A. longifolia are known to be present), with more cover observed in the East of the study area and farther from the coastline. In pine forests (PF2) higher cover of flowers were mostly visible in the open spaces where pine canopy did not cover A. longifolia (Figure 5) but also still visible through the pine tree canopy.

The analysis of the distributions of flower cover per m² in each site showed that all distributions were skewed, i.e., had an unbalanced distribution. In sites where flower detection was difficult, either because flowers were covered by pine canopy (PF1, PF2) or because images were acquired after the peak of flowering (SD4, SD5), the distribution was skewed toward the origin, reflecting lower values of flower cover per m² with less variability (Figure 6). On the other hand, imagery obtained during the flowering peak period (SD1, SD2, SD3) showed not only higher average values of flower cover but also the highest variability which translates the high spatial variability of flowering patterns of the species (Figure 6 and Supplementary Figures S5–S8).

The relationship between the number of flowers quantified during the fieldwork and the percentage of flower cover detected by the UAV was weak (Figure 7). The highest values were found in the primary and stabilized dunes where the flights were performed at the peak of flowering (SD1 and SD2, Figure 7). Lowest and non-significant correlations were observed in those areas where UAV flights occurred after the peak of flowering (SD4 and SD5) or where A. longifolia invades the understory of pine forest (PF2, Figure 7).

Our results using pixel-based Random Forest classification confirm the ability of using UAV to map invasive plants since we obtained very high classification accuracy in all our experimental locations. These results with the Random Forest algorithm are in-line with other works in the field of Remote Sensing (Immitzer et al., 2012; Feng et al., 2015; Tamouridou et al., 2016) and with object-based classification (Michez et al., 2016a; Yuan and Hu, 2016). This was expected due to the unique bright yellow appearance of A. longifolia flowers in the winter season in juxtaposition to the lack of yellow flowering in most co-occurring native or exotic plants (Flora-On: Flora de Portugal Interactiva, 2014).

The use of UAV in the field of invasive alien plants has recently become more common (Hill et al., 2016; Müllerová et al., 2017a). This technology can be a major contribution to monitor any phenomena that induces changes on plant phenology or morphology such as the spread of a forest pest (Lehmann et al., 2015b) or the establishment of a biocontrol agent. In our study, T. acaciaelongifoliae expansion is expected to reduce flowering (i.e., decrease quantity of flowers through space and time as it spreads from the release points) and vigor of A. longifolia (Dennill, 1988), and as such a high resolution Remote Sensing protocol to monitor the plant flowering may be an excellent solution to be used as a proxy to monitor temporal and spatial impact of this biocontrol agent.

Additionally, and despite a detailed cost analysis was not performed, the costs of monitoring the flowering in the field versus the costs of just deploying the UAV and processing the image were clearly different: flower monitoring at each field site (which sampled some plots and not the entire area overflow by the UAV) required four researchers to work for about 8 h while each UAV flight was done in less than 2 h with 2 operators (at some locations 1).

In studied sand dunes, as well as in other dune systems (Elorza et al., 2004; Rascher et al., 2011), A. longifolia frequently replaces native plant communities where shrubs and herbs are more frequent, occupying large patchy areas and becoming dominant as an overstory species, and so is easier to detect. As expected (Perroy et al., 2017), when A. longifolia occurs in the understory of the pine forests, it is not detected by the UAV. Nevertheless, at the peak of flowering A. longifolia visibility increased in the more open areas of the pine forest canopy.

Accuracy of classification was dependent of flowering season and habitat type, showing better values during flowering peak and in sand dunes where A. longifolia dominates the overstory. Even so, when flowers were hidden by pine canopy or when UAV flights occurred at the end of flowering season and less or senescent (weak yellow coloring) flowers were present, highly accurate values were still obtained. Using UAV technology to map below the canopy remains a challenge (Perroy et al., 2017) but new developments of unmanned aerial systems are starting to address flights in forests understory (Brust and Strimbu, 2015; Liao et al., 2016; Michez et al., 2016b). These results highlight the potential of this approach to detect variations on A. longifolia flowering, confirming their plausible use as a proxy to detect the decrease in flower production due to the establishment of the biocontrol agent.

The flexibility in the acquisition of imagery with UAV allows capturing specific stages of phenology (Müllerová et al., 2017b); in our study it allowed to capture the peak of flowering for most of the study areas. We observed that shadows can have very negative effects on imagery captured at high latitudes during the days of clear-sky. This happens due to the low altitude of sun on the local horizon even during the solar noon. Acquisitions during full cloud coverage significantly reduced the impact of shadows in our experiment, and did not have a meaningful impact on the high accuracy of the image classification exercise.

In a broader perspective, the high accuracy in identifying A. longifolia flowering confirms the potential for using UAVs to access phenological changes due to biota attack or any other factor that induce changes in plant phenology or morphology. This complies with other studies, e.g., that used satellite imagery to assess the impact of pine beetles (Skakun et al., 2003) and confirms the importance of accurate evaluation of plant characteristics. Such studies highlight how UAV techniques can be flexible and applied in a wide range of study fields. Not only forestry, but also agriculture and environment, are increasingly evaluating Remote Sensing approaches for improving management efficiency (Wulder et al., 2006; Xiang and Tian, 2011; Burkart et al., 2014).

The variation of phenology and structure of A. longifolia added to the variability of the multiple habitats where the species occurs and this increased the challenge of quantifying the flowers by UAV-based imagery. Acacia longifolia has a high structural variation as it can develop flowers along its branches, at different heights and overlapping (or not) each other, as confirmed during field measurements: e.g., dense groups of A. longifolia trees with high numbers of branches and flowers overlapping each other several times vs. trees with dispersed flowers that rarely overlap; flowers at the top of the canopy which are hard to reach (small trees up to 8 m) vs. flowers at soil level that are easily quantified (shrubs lower than 2 m). This high level of variation occurred within the same trees as well as between different trees and was, most likely, the main factor that contributed to the low correlation found between the numbers of flowers measured in the field and those estimated using UAV imagery. Few works have tried to evaluate this type of relation.

Thorp and Dierig (2011) provided a good example of the application of this methodology achieving much better correlation for Lesquerella fendleri (Brassicaceae). However, L. fendleri is an herb/forb which may produce all the flowers at approximately the same level; additionally, their experiment was conducted in highly controlled conditions with only one unique area of 1 m² being observed through time, which was not the case in our variable field conditions.

While our field measurements could potentially benefit from some re-design to better characterize what is in fact perceived by the UAV sensor and the inherent variability of A. longifolia flowering which is probably the main cause of the observed lack of correlation. To address what is detected by the UAV we suggest that the quadrats should be marked on the field prior to the UAV flights to ensure that positioning errors from GNSS receiver are not responsible for the mismatch. Furthermore, the volume of A. longifolia should be also considered by establishing allometric equations relating the volume with above ground biomass (Cunliffe et al., 2016).

This research was one of the first attempts to quantify the number of flowers using UAVs and although this goal was not satisfactorily achieved, it brought valuable insights on the limitations and potentials of these systems. The quantification of flowers, fruits, galls and/or any other plant trait is of increasing importance to Remote Sensing of vegetation and thus, the use of low-cost UAV as well as new satellite observation platforms (e.g., Sentinel) deserve further exploration.