The study area is located in the Atlantic Forest domain in Brazil, at coordinates 17°41′35″ S and 39°28′40″ W, in the municipality of Caravelas, southeastern Bahia State (Figure 1). The climate is classified as Tropical with no dry season (Af) in the Köppen classification system (Alvares et al., 2013), with annual mean precipitation around 1750 mm (Saporetti-Junior et al., 2012), and mean temperature around 26°C. Vegetation in the region is highly fragmented and predominantly composed of Atlantic rainforest remnants, Mussununga, eucalyptus plantations, pasture and crops. Mussununga is a heterogeneous savanna-type formation of different physiognomies, from grasslands, dominated by few species of monocots; savanna, composed of one layer of herbaceous plants and another layer of scattered woody plants; and woodland formed by a closed canopy (Saporetti-Junior et al., 2012; Ferreira et al., 2014; Lima et al., 2015). These vegetation patches are spread out in a matrix of Eucalyptus plantation and Brazilian Atlantic Forest (IBGE, 2004) having rounded to amoeboid shapes which greatly vary in size (less than one to hundreds of hectares). Mussununga vegetation is strictly associated with nutrient-poor, acidic, sandy soils formed through podzolization as a consequence of high humidity and hydromorphism (Saporetti-Junior et al., 2012; Ferreira et al., 2014). Our Mussununga study area was dominated by grasses and monocots with scattered patches of dwarf shrubs and small trees (Figure 2A).

Acacia mangium is indigenous to Australia, Papua New Guinea, and Indonesia. It is widely planted around the world, mostly with the purpose of soil rehabilitation due to its fast growth and nitrogen fixation (National Research Council, 1983). In Brazil, the species is recorded from North (Roraima and Amapa states) to South (Rio Grande do Sul state) (The Horus Institute National Database of Invasive Alien Species Brazil, 2017). Introduced around the 1970s as an alternative for the reclamation of degraded areas, it is still planted in forest stands for firewood, timber, fence posts, and support for black pepper crops (Halfeld-Vieira and Nechet, 2009). Acacia mangium is adapted to climates with a short dry period, mean annual rainfall between 1,000 and 4,500 mm, temperatures from 15 to 34°C, acidic soils with a low level of nutrients, and degraded areas (National Research Council, 1983; Delnatte and Meyer, 2012).

Besides being well adapted to these environmental conditions, A. mangium is a good competitor in high lighting conditions (Osunkoya et al., 2005) and is widely distributed in lowland Atlantic forest domains in the southeast of Brazil, invading degraded areas, pastures, eucalyptus plantations and Mussununga. In regenerating forest invaded by A. mangium where the canopy closed in short time and biomass increased in a few years, dominance by the invasive species delayed colonization by native species and had an adverse impact on biodiversity (Le Maitre et al., 2011). Likewise, invasion by A. mangium in Mussununga formations is expected to accelerate biodiversity loss and alter plant communities.

Acacia mangium is characterized by some unique physiognomies within the Mussununga formation making this invader a good candidate for mapping through remote sensing. Particularly, the bright green, large phyllodes (flattened leaf stalks; Figure 2B) are helpful in distinguishing A. mangium from the surrounding savanna-type vegetation. In addition, it usually grows taller than most native plant species in this formation (Figures 2C,D, 3).

A UAS consists of an aerial platform, associated sensors and control equipment (Valavanis and Vachtsevanos, 2015). The unmanned aerial platform, the autopilot control system and the camera sensors used to match the specific requirements (Table 1) of this remote sensing technique as an innovative invasive plant management tool are described in the following sections.

Figure 5 gives an overview of software modules and the steps of image processing and analysis used in this study. Before processing, we removed five images (out of 212) in the original RGB data set and eight images (out of 193) in the CIR data set due to their very poor overall quality caused by strong motion-blurring, high-angle obliqueness and under or overexposure effects. Images with further minimal quality issues were not removed to avoid potential gaps in the final, aligned orthoimagery. Furthermore, we preselected images with a specific focus in the AOI to optimize the subsequent processes by minimizing the total amount of images. These remaining images underwent an optional image equalization process (stretching) to reduce illumination differences.

A total area of ~25 ha was covered by aerial RGB and CIR imagery with ground pixel resolution of less than 4 cm. However, to reduce image data processing and analysis time, the imagery outside of the AOI (i.e., the Mussununga under study) was discarded. Hence, the final stitched orthomosaics covered an area of ~13 ha (Figure 7). A visual inspection of the two entire orthomosaics revealed some blur and distortion effects (<5% of total area), as well as light fluctuations (<10% of total area). RGB orthoimagery was affected by strong cast shadows resulting in partial loss of image details. CIR orthoimagery, in contrast, was less affected by cast shadows but includes some underexposed areas due to the presence of clouds. Since partially the same GCP positions were used during acquisition of RGB and CIR images, spatial shifts could be detected between the two orthorectified mosaics. As a consequence, one-to-one pixel co-registration of both data sets failed due to glaring perspective distortions. Nevertheless, both aerial images showed no significant pixel gaps or missing parts. To illustrate image processing results and methods of analysis in more detail we selected an one-hectare subscene intensely invaded by A. mangium.

The interpolated DSM proved to be a valuable representation of surface elevation over the study area. As a result, elevated objects, in this case, larger trees or shrubs, were clearly recognizable visually (Figure 8). The linear regression model indicated a strong correlation (R² > 0.89 with N = 60) between the predicted DSM and field measurements of A. mangium heights (Figure 9).

The resulting grayscale maps of the calculated VIS-NIR-based vegetation indices are presented in Figure 10. The associated boxplots are shown in Figure 11. The GNDVI showed high aptitude to differentiate photosynthetic active vegetation (i.e., plant biomass) from other surface types such as bare soils (P < 0.001). However, the application of the GNDVI revealed some limitations in the distinction between different plant species or vegetation types. Particularly, A. mangium and grasses had similar pixel values (P = 0.776). In this case, the calculated GDVI was favorable. The applied Tukey post-hoc test showed significant pixel value differences between all classes (P < 0.001). The modified TVI was most promising in detecting A. mangium since pixel values for this distinct species differ significantly from the pixel values related to surrounding native grasses and herbaceous/woody vegetation (particularly if applied to A. mangium individuals with new, bright green phyllodes).

The resulting classification map is presented in Figure 12. The semi-automatic pixel-based classification revealed an overall accuracy level of 85.4% and overall KIA of 0.82 using all six classes (Table 5). Producer's accuracy achieved best results for shadow (98.6%), followed by bare soil (97.4%), road (97.1%), other vegetation (84.3%), grasses (74.5%), and A. mangium (73.1%). User's accuracy performed best for road (100%) and bare soil (99.3%) classes. The A. mangium class reached a user's accuracy of 95.9%, followed by shadow (83.6%), grasses (81.4%), and other vegetation (69.5%).

The final A. mangium distribution map obtained from multilayer visual image interpretation is shown in Figure 13. The image interpretation revealed an overall accuracy rate of 82.7% for the correct identification of A. mangium (Table 6). Using aggregated tree heights of A. mangium, the height classes 3–5 m and 5–7 m were detected with higher accuracy, 91.8 and 88.2%, respectively. In comparison, A. mangium trees lower than 3 m or between 7 and 10 m were more difficult to detect, with 77.1 and 66.7% accuracy, respectively.

In this study, we present an innovative, integrated remote sensing approach which combines low-cost UAS techniques with open-source GIS and freely available image processing software. The methodology described allowed us to map invasive A. mangium trees with high accuracy with a reasonable effort. The classification algorithm based on the training data can be applied to image data collected by the same camera sensor at new sites, and the resulting maps can improve the information base for invasive plant management. In general, this approach can allow environmental managers to monitor the distribution and spread of invasive plants in similar situations once the classification models are adjusted to the respective species. Environmental managers should be strongly encouraged to use low-cost UAS approaches to avoid high costs and time-consuming field surveys, especially in sites of low accessibility (Wan et al., 2014; Chabot et al., 2016; Hill et al., 2016; Müllerová et al., 2017). Minimizing in situ mapping and control operations is also an added advantage for using UAS, because reduction in habitat disturbance can encourage the recovery of native plants that are threatened by biological invasion (Mack and D'Antonio, 1998; Huston, 2004).

Several studies have already used UAS acquired image data successfully for detecting and monitoring invasive plants. However, these studies used either expensive high-end UAS (e.g., Michez et al., 2016) or commercial image processing and analysis software. For example, Müllerová et al. (2017) used a self-built, low-cost fixed-wing UAS to detect invasive Robinia pseudoacacia in a recently published study, but applied commercial stitching software. Such sophisticated software products are usually expensive, often exceeding 3,000 US dollars for a single license. Our results show that open source software may provide sound alternatives, from image data pre-processing to image analysis.

There are also drawbacks related to low-cost software and techniques. Compared to commercial photogrammetric processing software, Microsoft ICE image processing may lead to poorer geometric accuracy. However, it is the best free-of-charge option for UAS photogrammetric application and suitable for UAS applications unless high geometric accuracy (<2 m) is required (Gross and Heumann, 2016). Consumer-grade camera sensors have spectral limitations and provide poorer image quality compared to commercial-grade camera sensors: in RGB and CIR orthomosaics, both imagery blurring and partial distortion effects are noticeable. Blurring can be caused by fast movements of the aerial platform during image acquisition, e.g., due to wind or turbulence, as well as by faulty images used in the mosaicking. This can generate errors in the visual or automatic analysis of the image data and reduce the accuracy of automatic photogrammetric processing algorithms. In this study, only a few photos had to be removed because of their poor quality. Some of the remaining images, however, also showed some quality loss. Thus, using such Do-it-Yourself approaches always includes a tradeoff between removing low quality images and avoiding potential gaps in the final orthoimagery. A high overlap of individual images is necessary to provide flexibility in the use or removal of single low-quality images. The detection and removal of blurred images are currently done manually, and is a time-consuming, error-prone process, particularly with large image sets. Further, the removal of low quality images can negatively influence subsequent post-processing steps since less data are available. Fortunately, new studies dealing with automatic detection of blurred images in UAV image sets have been published recently (Ribeiro-Gomes et al., 2016; Sieberth et al., 2016).

Some spectral limitations of the CIR imagery are a consequence of the camera sensor modification. Since we used a cyan filter (which blocks the visible red light), the sensor recorded only in the NIR, green and blue channels. Most traditional vegetation indices, however, include the red band. As a consequence, we used the green band instead in order to calculate modified VIs. This limits comparability of indices based on images from modified low-cost cameras and from professional sensors (or literature values). Nevertheless, UAS-based application of modified low-cost cameras are promising for detecting and mapping invasive plants, although the results may vary among species of interest because vegetation is characterized by strong reflectance in the NIR.

Remotely-sensed vegetation indices (VI) have been extensively used for monitoring and detecting invasive tree species (Huang and Asner, 2009; Huang et al., 2010; He et al., 2011; Bradley, 2014). VI related to physical leaf features like greenness, chlorophyll or water/ moisture content, richness in lignin, nitrogen, and cellulose have especially proved to be highly effective in differentiating Acacia species from surrounding vegetation (Larsson, 1993; Oldeland et al., 2010; Große-Stoltenberg et al., 2016). The vegetation indices applied in this study were highly efficient in distinguishing invasive A. mangium trees from surrounding native Mussununga vegetation. In particular, the modified Triangular Vegetation Index (mTVI) separated A. mangium well. This index originally belongs to narrowband VI (Broge and Leblanc, 2001). Nevertheless, prior applications of the mTVI in a remote sensing study by the authors (Lehmann et al., 2016) also revealed high potential to distinguish plant species using imagery acquired with a self-modified CIR camera sensor.

In addition to spectral features, the DSM computed from the UAS-based aerial RGB images provided clues for the recognition of invasive A. mangium because it grows taller than the surrounding vegetation. We found a strong correlation between predicted and measured tree heights. This was facilitated by the flat terrain in the study area. However, in areas with substantial elevation differences, an underlying terrain model must be considered. Similar studies showed the potential of canopy-height models generated from UAS image data. Zarco-Tejada et al. (2014) used high-resolution image data to generate a precise DSM and a Digital Terrain Model (DTM) of their study site. By subtracting the DSM from the DTM, an informative canopy height model of the forest in the study was created. They concluded that surface models derived from UAS-acquired high-resolution image data can be as accurate as those generated with expensive, computationally more complex light detection and ranging (LIDAR) systems currently used in agricultural and environmental applications. Still, the Structure from Motion-based generation of canopy-height models that we applied in this study can be challenging to non-experts in remote sensing. The use of open-source software for this purpose also involves some obstacles as it requires a stepwise combination of various software modules, in contrast to expensive commercial software in which all required functions are combined in one product.

The classification strategies used in this study denote a further problem: only partial tree crown areas were recognized and correctly classified as A. mangium. This is possibly due to the low illumination angle of the morning hours when the UAS was flown. In general, pixel-based clustering algorithms only use the spectral values of each pixel, but neglect texture and, hence, often result in image speckle and overall inaccuracies when applied to fine-resolution imagery. Object based image analysis (OBIA) can be suitable to improve classification results and reduce these speckle (“salt-and-pepper”) effects (Blaschke, 2010). In contrast to pixel-based analysis, OBIA includes the geometric properties of the target plant species, such as dimension or texture. Open-source software already shows high potential in using this method (ORFEO Toolbox for Grass GIS), but the application is still complex and availability of algorithms is limited compared with commercial software (e.g., eCognition). In this study, visual image interpretation, in addition to the pixel-based classification, improved the identification of A. mangium, even in large areas.

Evidently UAS offer capabilities for mapping the occurrence, dynamic or extent of invasive plant species in an area. Furthermore, UAS acquired remote sensing data generate additional valuable geospatial information for management. Bare soil cover, for example, can be quantified from vegetation index maps. We were able to map bare soil areas with high accuracy using simple semi-automatic classification. As mentioned before, disturbance increases opportunities for the spread of invasive plants (Lorenzo et al., 2010; Le Maitre et al., 2011) in open areas, on bare soil and other modified sites such as roadsides, trails, and degraded vegetation. These areas should, therefore, be a priority for preventative monitoring and early detection of invasive species followed by eradication or control.

There are three major approaches for managing invasive alien species. The first one refers to the prevention of species introductions to countries. Despite recent efforts to discriminate species for deliberate introductions and improve control of involuntary introductions, the rate of species introductions has not declined at the global level (Seebens et al., 2017). The second approach refers to early detection and rapid response for eradication or containment, and the third to sustained control over time once populations have grown too large or too widespread to be eradicated (Wittenberg and Cock, 2001). Species that invade forests can unfortunately not be mapped from the air, but most invasive plants thrive in open ecosystems, belonging to the group of pioneer species. The UAS technology can strongly contribute to improve current methods of inventory and prioritization for the eradication or control of these plants, as covering large areas on foot, often in difficult or inaccessible terrain, can be counterproductive or simply not viable in terms of cost, safety or practicability. Detection at early stages of invasion provides the best chance for successful eradication or control, as both cost and effort are still low and more feasible. UAS can greatly increase the opportunities for early detection in open ecosystems such as Mussununga, coastal scrub (“restingas” in Brazil), wetlands, grasslands, savannas and degraded areas, especially when the invasive plants are taller or somehow stand out in the native vegetation matrix—as acacias in Mussununga.

In any given area the control of invasive plants is initially easier, as there tend to be more individuals to locate and remove. As work progresses, finding isolated or few remaining plants becomes more challenging, leading to increased costs and work hours to locate them. It is often relevant to locate and remove mature plants first to prevent the seed production and renew the seed bank in the soil. Finding these plants is often difficult, especially in sloping terrain, coastal dunes or in patches of high and low vegetation where visibility on the ground is limited. UAS have high potential in facilitating the location of such plants. Even if not 100% precise, excluding areas where invasive plants clearly do not occur can still save much time and effort in the field.

Open areas of high susceptibility to invasion can be identified in UAS-based imagery, then surveyed by targeted field sampling. Further, aerial imagery can be processed into geo-referenced orthomosaics right after the completion of UAS flights for immediate control. Such UAS-based aerial image maps are more up-to-date and offer higher spatial resolution compared to freely available products, such as imagery from Google Earth. UAS can, thus, be used to support immediate management actions in the field as well as monitor the results of control efforts, especially when the ideal time to revisit specific sites to repeat control operations on remaining plants or on new plants grown from the seed bank is not clear, which is often the case. Revisiting control sites is crucial to prevent new plants from reaching maturity, replenishing the seed bank in the soil and renewing the invasion process.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.