Invasive species are considered among the major drivers of global change and are threats to biodiversity, ecosystem services, the economy, and human health (Pyšek and Richardson, 2010; Vilà et al., 2010). Early monitoring and rapid action at the operational level are therefore needed to mitigate the consequences of plant invasions (Kaiser and Burnett, 2010). Thanks to the availability of high spatial-resolution data, remote sensing (RS) is being increasingly used to study and model the spread of invasive plant species (Pergl et al., 2011; Somers and Asner, 2013; Rocchini et al., 2015), their phenology (Ge et al., 2006), and impacts (Asner et al., 2008). Most of the RS studies focus on invasive trees and shrubs (e.g., Frazier and Wang, 2011; Masocha and Skidmore, 2011), but the accurate detection of herbaceous plant species is challenging owing to their smaller size and patchiness. In this sense, invasive plant species are easier to handle compared to native ones since they usually grow vigorously, produce a great amount of biomass, and dominate the canopy, often forming extensive monospecific stands (Maheu-Giroux and de Blois, 2004; Huang and Asner, 2009; Müllerová et al., 2013). There continues to be a need for data of sufficient quality and resolution, especially where the target species is less distinct from other vegetation, or covered by it.

It has been shown that the timing of data acquisition plays an important role (Laba et al., 2005) since plants are often more distinct from the surrounding vegetation at certain times of the vegetation season (Andrew and Ustin, 2008; Rocchini et al., 2015), mostly during flowering (Müllerová et al., 2005, 2013; Andrew and Ustin, 2006; Ge et al., 2006) or leaf coloring (Shouse et al., 2013). Using a change detection approach, some species can be recognized by the differences in their life cycle compared to the background vegetation (Peterson, 2005; McEwan et al., 2009). Such adaptation is quite common in plant invasions, since the mismatch of phenology between native and exotic species can provide benefit to invaders outcompeting the natives by using the time window available via extending or shifting the growing season (Wolkovich and Cleland, 2011; Fridley, 2012; Gioria et al., 2016). Phenological differences influencing the RS detection include also the seasonal development of broad-leaved tree canopies, because the target species in the understory may be detectable only before emergence and after the leaf senescence of the canopy vegetation (Resasco et al., 2007; Wilfong et al., 2009). Thus, the seasonal dynamics and spectral behavior of the target species and their surrounding constitute important parameters in the RS detection process.

To study the role of phenology on RS detection of invasive species, it is necessary to acquire data at certain phenological phases, as well as with suitable spatial and spectral resolution. Freely available satellite data such as MODIS, Landsat or Sentinel do not provide appropriate spatial resolution for a highly heterogeneous Central European landscape where the occurrence of invasive plant populations is rather patchy. Very high resolution (VHR) satellites such as Pleiades, QuickBird, Ikonos, or WorldView are costly to acquire, and in all cases the acquisition of suitable data is constrained by clouds and the regular trajectory of the satellite. Aerial campaigns offer some flexibility, but they must still be planned in advance, with uncertainty in weather forecasting being a factor, and they are also costly. If relying on the archives of aerial imagery, the choice of phenology phases is limited and spectral resolution is usually low. New means such as unmanned aircraft vehicles (UAV) provide high flexibility, low price, and easy deployment (Watts et al., 2012), but rather low spectral resolution (RGB + modified NIR bands with modified consumer cameras if cheap solution is considered), and due to geometric and radiometric irregularities they require complex processing (Rango et al., 2009; Laliberte et al., 2010; Salamí et al., 2014; Müllerová et al., 2017). At the same time, the large amount of data collected by UAVs brings about the need for automatic processing. Legal constraints limiting use of UAVs only to unpopulated areas further reduce their applicability (Watts et al., 2012). This is especially problematic for invasive species that tend to prefer urban areas (Pyšek, 1998; Kowarik, 2008). Moreover, extremely high detail of UAV data generates problems with precise location of validation data due to the limits of available GNSS instruments (especially in forested or hilly environments). Still, despite these limits, the UAV technology holds great potential for invasive species assessment and monitoring (Calviño-Cancela et al., 2014; Michez et al., 2016; Müllerová et al., 2016, 2017; Table 1).

To demonstrate the role of timing and resolution for the RS detection of invasive plants, we chose two model herbaceous species with markedly different phenological, structural, and spectral characteristics. They are the giant hogweed (Heracleum mantegazzianum), a species with a very distinct flowering phase, during which it forms large white circular inflorescences, and knotweeds (Fallopia japonica, F. sachalinensis, and their hybrid F. × bohemica), lacking such distinct features. High flexibility of the data acquisition provided by small UAV (in our case a custom-built one producing imagery of 5 cm spatial and multispectral resolution of RGB + modified NIR) enabled us to study the role of phenology. The results were compared with those of multispectral satellite imagery of 50 cm resolution using several approaches (pixel- and object-based) and data from different parts of the vegetation season to find a repeatable, efficient, and low-cost monitoring strategy (in terms of data source, processing approach and target species' phenological stage). Our aims were to: (a) assess the role of timing and both the spectral and spatial resolution in detection of the two structurally and phenologically different invasive species, (b) define the optimal phenological stage (time window) and data resolution, and (c) describe the computer-assisted workflow of the image-processing to achieve the highest detection rate. Our recommended methodology, operational in environmental monitoring and applicable over large areas, can be used for nature conservation, with the potential to reduce the costs of extensive field campaigns.

Our results show that both the detection accuracy and the best classification approach depend strongly on the phenological stage of plants, and on the spectral resolution of data (Table 2). This was true for both model species regardless of their markedly different phenological, structural and spectral characteristics, and seasonal dynamics. While giant hogweed was best detected when plants were flowering, thanks to its large white inflorescences, knotweeds were distinct at the phase of senescence because of reddish-brown coloring of decaying plants (Figure 2). In other parts of the season, detection of any of the species was more problematic and higher spectral resolution of satellite imagery improved the detection. The object-based approach was successful for giant hogweed if (i) the species was sampled in the right phenological stage and (ii) the spatial resolution enabled the distinction of individual compound umbels. For less distinct knotweeds spectral resolution of the data and pixel-based approach played more important role.

Giant hogweed was detected with a very high accuracy when flowering (up to 100%), dropping down later in the vegetation season to ~60% (September; Table 4). Flowering, undamaged individuals were detected with high accuracy, whereas non-flowering, fruiting, cut, sprayed, or grazed ones were more difficult to recognize due to their similarity with the surrounding vegetation. Further, the leaves surrounding flowering umbels were difficult to map (Figures 2, 3). The best results were obtained for full flowering (UAV imagery; 9 July; site 2) using an object-based approach—contrast split segmentation followed by rule-based classification (Figure 3). Contrast split segmentation itself separated very well the white objects on the imagery, meaning that all flowering hogweeds were detected. They were, however, mixed with other bright objects such as some artificial surfaces or harvested fields and had to be filtered out by further rule-based classification. At peak flowering, the inflorescences were so distinct that they could be detected with reasonable success (42/100% PA/UA of the hogweed class) even from a low spectral resolution aerial orthophoto (25 cm) using the same approach. However, the template-matching approach incorporated in eCognition software failed to recognize the umbels even at the peak of flowering. Depending on the threshold, it either picked all the hogweed umbels and many other objects of similar shape, even of different color or, if the threshold was lower, it missed half of the plants. It could work well if the infested area was previously masked out and the algorithm ran with high threshold only inside the infested polygons. This approach would require substantial manual input and would, thus, not be feasible to apply over larger datasets. In case of site 1, the summer UAV imagery was obtained later in the season (15 and 20 July) and part of the umbels in hogweed compound inflorescences had already started to ripen, slightly decreasing the accuracy. The contrast split rule-based approach of 15 July 2015 was the most effective, but failed completely for the imagery from 20 July 2016, when hogweed was mostly out of bloom. In the case of ripening hogweed (site 1), pixel-based approach and ML algorithm of the satellite imagery provided higher accuracy measures when compared to those derived from UAV. Still the accuracies were lower compared to full flowering hogweed on UAV imagery from site 1. In autumn, the UAV imagery depicted only leaves and dry stems, and the accuracy was very low (Table 4, Figure 3). In this case, no distinct objects could be identified and the object-based approach failed.

The accuracies of knotweed classification varied considerably according to the phenology and spectral resolution (Table 5; Figure 4). The best results obtained were for the late-autumn UAV imagery (8 November; ML for both original and 50 cm resampled), when reddish-brown knotweeds stands were clearly visible (Figure 4). At other parts of the vegetation season, the PA derived from UAV imagery was low. Resampling of UAV data from 5 to 50 cm to simulate the Pleiades resolution slightly improved accuracies by eliminating salt and pepper effect, showing that spatial resolution does not limit species' recognition. Classification of summer Pleiades imagery using a RF classifier provided high accuracies (74/95% PA/UA of the knotweed class). Due to their architecture, knotweeds do not form objects of distinct shape or color, and object-based algorithms were therefore unsuccessful.

Our study addresses how different aspects of data quality (spatial, spectral, and temporal), costs, and operationality affect the methodology and practical application of the detection of invasive plants. Our two model species of different phenology and architecture, giant hogweed and knotweeds, showed different results reflecting their spectral characteristics and seasonal dynamics, still phenology played an important role in detection of both species (cf. Huang and Asner, 2009; Somodi et al., 2012; Müllerová et al., 2013; Bradley, 2014). The choice of the best data source (aerial color orthophoto, UAV with RGB + NIR modified consumer cameras, and Pleiades satellite MSS imagery) and classification approach (object or pixel based) depended strongly on the species' characteristics and their phenological stage.

We demonstrate that the spectral resolution of the low-cost solution, represented by an unmanned aircraft system, is limited (720–950 nm with variable response declining rapidly toward the longer wavelength; Müllerová et al., 2017), but that this limitation can to some extent be surpassed by identifying the right phenological stage when a more satisfactory detection rate is achievable even for the less distinct species such as knotweed. During the spring, when the knotweed stands are formed by the mixture of old red stems from the previous year and short emerging new regrowth stems, and summer acquisition, with dense, fully-developed indistinct stands, spectral resolution became more important for the detection. On the other hand, during the senescent phase the reddish color and better visibility of knotweeds under defoliated trees provided good detectability irrespective of the spectral quality. The problem with the canopy cover hiding knotweeds during the vegetation season was partly solved by using off-season imagery (from late autumn), but some individuals still remained hidden, covered by trunks or by their shadows.

The very fine spatial resolution of UAV imagery was not always beneficial. In case of hogweed, it actually overwhelmed the relevant spatial patterns and hampered the classification based on flowering objects since, in such a great detail, instead of the consistent white “dots” of hogweed inflorescences, individual umbels of each inflorescence were visible, eventually decreasing the classification accuracy (cf. Ustin and Santos, 2010). On the other hand, the fine detail enabled mapping of the hogweed leaves surrounding flowering umbels and, to some extent, non-flowering individuals. In case of knotweed (pixel-based approach), very fine spatial resolution of UAV data caused the salt and pepper effect, eliminated after resampling and being less pronounced in case of late autumn imagery and slightly blurred due to technical matters relating to cameras. Still, the resampling of summer UAV imagery lowered the success rate, probably because, in the forest, individual pixels of knotweeds visible among the canopy trees were merged into mixed pixels.

We achieved slightly higher giant hogweed detection rates compared to previous aerial photography analysis (Müllerová et al., 2013). Michez et al. (2016) used a similar approach, applying multiresolution segmentation (an object-based approach) and RF classifier on UAV imagery gaining higher overall accuracy; still since the authors do not provide the hogweed class accuracy, it is difficult to compare their results with ours. In contrast to our research, earlier studies of knotweed detection did not achieve the accuracies sufficient for operational application (Dorigo et al., 2012; Michez et al., 2016). Dorigo et al. (2012), who detected knotweed using a ratio of spring and summer aerial photography, applied a random-forest classifier based on pixels combined with textural information derived from a moving kernel. Jones et al. (2011) performed a rule-based object-oriented classification of aerial photography but provide no assessment of accuracy so their results are difficult to compare. Michez et al. (2016) applied multiresolution segmentation (an object-based approach) and RF classifier of UAV imagery, but provided only overall accuracies and Kappa index, and admitted that they did not reach sufficient accuracies for operational application. These authors show that the best results were provided by using very small objects (~30 cm in size) and spectral indices, which is actually similar to the pixel-based methods. In some cases, ML performed slightly better compared to the learning algorithms, still the differences were minor. The commonly accepted principle of machine learning algorithms outcompeting ML cannot therefore be taken as a rule of thumb (but see Andermann and Gloaguen, 2009; Otukei and Blaschke, 2010). For object-based classification, a rule-based algorithm was more successful compared to the machine learning, still the method is highly case-specific, less universal, and requires substantially more expert knowledge.

Trade-offs exist between spatial, spectral, and temporal resolution while minimizing cost and making an approach operationally viable (Wiens et al., 2009; Willis, 2015). The demands for both spatial and spectral resolution depend largely on the target plant species; good results can be achieved with substantially lower resolution if the vegetation is sampled at the phenological phase when the species is at its most distinct. Our study shows that proper timing and high spatial resolution can to some extent compensate for the lower spectral resolution of a low-cost, unmanned aerial system, and help to decrease the error of omission. This is important for practical applications as it minimizes the number of missed plants (Müllerová et al., 2013). Multispectral or even hyperspectral UAV sensors available on the market would definitely improve the resulting accuracy; our low-cost solution of modified consumer cameras provides substantially lower spectral quality (Müllerová et al., 2017). We believe, however, that such instruments would make the resulting methodology too costly and complex to be puts into operation for nature conservation or land management.

UAV technology with its high flexibility and low costs can provide an appropriate sampling method at a high spatio-temporal scale. We do, however, need to be aware of several limits for application in the management of invasive plants. Among these are legal constraints in those urban areas that are typically invaded by alien plants and under the focus of land managers (Pyšek and Hulme, 2005). UAV is well-suited for experimental studies, targeted monitoring and eradication control, yet the satellite imagery (if available at right phenological phases) provides more appropriate solution for larger areas. For species with distinct appearance aerial imagery should also be considered.

The variety of data and especially flexibility of the UAV approach enabled us to assess the effect of timing of the data acquisition (i.e., the phenological stage of the target invasive species) and both spectral and spatial resolution on the detection success. Our research establishes a methodology for targeted timely monitoring of invasive species, provides land managers and nature protection with information on spatial distribution of invaders and serves as a baseline for assessment and modeling of spatial patterns and future spread of invasive plants. Nevertheless, it is necessary to be aware of limits of the RS detection, for example, if the target plant is under dense tree canopies or heavily damaged by mowing, grazing, or spraying. Our results indicate that the choice of the best classification method is case-specific, depending largely on both the target plant and imagery characteristics. The results from one methodological comparison should not, therefore, be mechanically transposed to other target species and data types. A detailed methodological analysis and testing of possible approaches such as ours is therefore a necessary step in finding optimal mapping strategies. For practical implementation of the proposed monitoring approach, the recommended time window should be wide enough to allow for organizing a flight campaign, and highly-flexible, small UAVs are a great choice, providing an opportunity to carry out the mission at the right time and in a cost-effective manner.

JM contributed substantially to the conception of the work, data acquisition, processing, analysis and interpretation, and composing the manuscript. JB contributed substantially to the data acquisition, processing, analysis and interpretation, and composing the manuscript. TB contributed to the data acquisition and processing, and composing the manuscript. PD contributed to the data acquisition and processing, and composing the manuscript. PP contributed to the conception of the work and composing the manuscript. MV contributed to the field data acquisition, interpretation, and composing the manuscript.