The essence of acquisition time of airborne hyperspectral and on-ground reference data for classification of highly invasive annual vine Echinocystis lobata (Michx.) Torr. & A. Gray

ABSTRACT Invasive alien species are one of the biggest threats to biodiversity today. Identifying their locations are mandatory parts of the strategies being developed to control them. Remote sensing along with machine learning are already proven and effective tools for monitoring invasive species, especially trees, shrubs, and tall perennials. However, annual vine species are particularly difficult to map using remote sensing because of their dynamic plant growth and the movement of shoots during the growing season. Therefore, the phenological phase in which the data is acquired, and the synchronization of airborne data acquisition with on-ground reference data, may be key factors for correct plant classification. This research aimed to answer the following questions: (i) What is the impact of acquiring synchronized on-ground data and hyperspectral data in different phases of plants’ phenological development on the annual vine IAPS (Invasive Alien Plant Species) classification results? (ii) How does the lack of synchronization while obtaining hyperspectral and on-ground data collection impact annual vine IAPS mapping results? (iii) Does multitemporal image fusion improve the results of annual vine IAPS classification? For this purpose, research was carried out on Echinocystis lobata, an annual vine species considered highly invasive in Europe. The obtained results indicate that the phenological phase in which the data is acquired has a very strong influence on the quality of the classification result. The period of flowering (summer) with the greatest coverage of the area with shoots was optimal for the classification of Echinocystis lobata with F1 classification accuracy of 0.87 ± 0.04. The accuracy of classifications was significantly less for spring (F1 = 0.64 ± 0.04) and autumn (F1 = 0.75 ± 0.05). Obtaining on-ground reference data that is mismatched temporally with hyperspectral data causes a decrease in the accuracy of the result up to 0.08 (from F1 = 0.64 to 0.56) in relation to data obtained synchronously. In the multitemporal image fusion method, using hyperspectral data linked from different phases of plants’ development to classify an image had a minimal improvement in classification accuracy compared to classifications trained on images from one phenological stage. The main conclusion is that mapping an annual vine using remote sensing and machine learning is possible and highly effective, provided the remote sensing and on-ground data are obtained in strict synchronization and the appropriate phenological phase. For the most efficient classification results, a single data acquisition per year is enough, even in the case of annual vine IAPS. Further research is needed to explore the possibility of mapping Echinocystis lobata using, i.e. multispectral or hyperspectral satellite data (e.g. EnMAP).


Introduction
The uncontrolled spread of invasive alien plant species (IAPS) is currently one of the main aspects affecting global change, resulting in the loss of biodiversity, degradation of ecosystems, and impairment of ecosystem services worldwide (Rai and Singh 2020).In addition to serious ecological effects, IAPS have enormous economic costs and negatively impact human health, such as allergies (Pyšek and Richardson 2010;Pejchar and Mooney 2009).Despite increasing efforts aimed at monitoring and controlling (European Commission, 2014), the threat from IAPS continues to grow (Hulme et al. 2009).Therefore, one of the most important objectives of modern nature conservation is to develop new and precise techniques of IAPS monitoring to implement effective strategies for controlling invasion.Increasing numbers of studies indicate that remote sensing (RS) and machine learning (ML) techniques may be useful and valuable tools (Underwood, Ustin, and DiPietro 2003;Joshi, De Leeuw, and van Duren 2004;Walsh 2018;Khare, Latifi, and Ghosh 2018;Khare, Latifi, and Rossi 2019).Compared to traditional survey methods such as hand mapping (Cooksey and Sheley 1997), they are spatially continuous, objective, and reproducible.Moreover, RS and ML techniques make it possible to identify IAPS in a short time in very large areas that are often difficult to explore with traditional survey methods.
Successful IAPS mapping using RS methods depends on many factors, one of which is data acquisition time.The time of data collection must correlate to the point in the growing season that the plant can be distinguished from the surrounding vegetation with the highest precision (Rocchini et al. 2015;Müllerová et al. 2017;Singh et al. 2018).It must be adapted to the target species' characteristics (Bradley 2014).Studying phenological variations of different plant species and finding the correlation with their optimal spectral response could determine the best data acquisition time for a given species (Niphadkar and Nagendra 2016).
Most studies used RS and ML methods to identify and map invasive trees, shrubs, and perennial herbaceous plants.Those plants can form large, monotypic stands, which are strongly distinctive from other co-existing plants.(Bradley 2014).In these cases, it is possible to use sensors with moderate spatial/spectral resolution data, such as satellite imagery (Müllerová et al. 2016;Huang and Asner 2009).For accurate detection of smaller invasive herbaceous plant species, high spatial resolution (often below 1 m) imagery obtained from airborne-level is often used (Chance et al. 2016;Kopeć et al. 2019;Sabat-Tomala, Raczko, and Zagajewski 2022).Annual vine species are examples of species that are difficult to identify using RS and ML, therefore, highresolution images are indicated for their study.A high detection efficiency of Polygonum perfoliatum was obtained using RGB data (orthophoto map in Red-Green-Blue colors), which is only one example of annual vine IAPS mapping (Zhou et al. 2021).
An example of an annual vine species that threaten natural ecosystems is wild cucumber Echinocystis lobata (Michx.)Torr.&;.Gray (Grašič et al. 2019;Stanković et al. 2022).It is a vine species, which can spread over a vast area during one growing season, contributing to the displacement of other plant species and the loss of biodiversity (Silvertown 1985;Kostrakiewicz-Gierałt et al. 2022).In general, very little research exists on mapping invasive annual herbaceous plants in heterogeneous habitats, where the invasive plant is similar to different cohabitant plants (Michez et al. 2016;Tesfamichael et al. 2018).In contrast, there are more studies focused on identifying annual weed in cultivated monocultures (de Castro et al. 2012;Gibson et al. 2004;Gray et al. 2008;Peña-Barragán et al. 2007;Hamouz et al. 2008;López-Granados et al. 2006).Many studies focused on identifying annual weeds in cultivated monocultures, but annual vine IAPS in the natural environment, which can pose a threat to biodiversity, are rarely studied.Further examples can be found which demonstrate the potential use of RS and ML concerning perennial lianas.Species such as Pueraria montana (Cheng, Tom, and Ustin 2007;Peters 2016;Liang et al. 2020), Mikania micrantha (Chen, Lin, and Sun 2014;Dai et al. 2020) and Lygodium microphyllum (Wu et al. 2006) were studied using satellite data.Still, none of these studies addresses, how phenology and synchronization of RS and on-ground reference acquisition affect mapping accuracy.
One way to identify IAPS using RS techniques is by implementing ML methods (Maxwell, Warner, and Fang 2018).It is a widely used solution to assess plant distribution based on a single time point in a season.Currently, numerous studies indicate that the classification of plant species or vegetation using a fusion of multiple raster data from various stages of plants' development (also called multitemporal image fusion) produces good results, improving the accuracy of classification (Marcinkowska-Ochtyra et al. 2019;Labonté et al. 2020).However, considering that some IAPS go through the entire development cycle in one year (from the seedling to the senile stage), classification maps made on RS images from different periods of the plant's development must be checked.
Given the particularity of the life cycle of annual species, the appropriate time for data acquisition is limited, due to rapid changes in plant traits during the vegetation season.Mapping species using ML also requires a dataset that trains and validates the model.For many plant species (like Echinocystis lobata), it is not possible to determine a set of train and validation pixels using photointerpretation of image data (Marshall and Lee 1994).The studied plants may be visually too similar to other plants.For these species, it is necessary to collect on-ground reference data (i.e.ground truth) during field surveys.The synchronization of RS and on-ground reference data seems to be an extremely important issue for such variable species as annual vines.The synchronization means gathering on-ground reference data simultaneously with airborne data and it is performed before any changes in the study species occur.Classification maps made with RS techniques require collecting on-ground reference data, possibly spectrally clean, with a significant percentage coverage of the study species in the reference polygons (Kopeć et al. 2019).In the case of annual vine species, the dynamics of changes in species coverage of a given reference polygon may be rapid.Therefore, it should be assumed that the gathering of on-ground reference data before or after the RS data acquisition will have a negative effect on the classification result, especially when the interval of time is significant.Demonstrating how negatively the lack of synchronization affects the classification result has not yet been investigated.In our research, we used a unique set of data that made it possible to conduct such an analysis.
The objective of this research was to establish the classification method for a highly invasive annual vine, using an example of Echinocystis lobata.This article addresses the following questions: (i) What is the impact of acquiring synchronized on-ground data and hyperspectral data in different phases of plants' phenological development on the annual vine IAPS (Invasive Alien Plant Species) classification results?(ii) How does the lack of synchronization while obtaining hyperspectral and on-ground data collection impact annual vine IAPS mapping results?(iii) Does multitemporal image fusion improve the results of annual vine IAPS classification?

Study species
One example of an invasive annual species occurring in species-rich natural, semi-natural, and anthropogenic habitats is the wild cucumber, Echinocystis lobata (Hulme et al. 2009;Weber 2017).It is an annual vine listed as one of the 100 most invasive species of plants, animals, and fungi in Europe (Vilà et al. 2009).
The species comes from North America, where its natural range extends between 35ºN and 53ºN and from the Atlantic coast to 110ºW.It was brought to Europe at the turn of the 20 th century both intentionally, as an ornamental plant, and unintentionally, with the transport of cotton during the same time period (Botta-Dukát and Balogh 2008).Within its secondary range, E. lobata usually occurs first in anthropogenic habitats, i.e. along roadsides and in ruderal spots.From there it encroaches on semi-natural and natural habitats, usually wet and fertile areas such as river valleys, and especially in the immediate vicinity of riverbeds and ecotone zones (Protopopova et al. 2015), which are protected as NATURA 2000 habitats (code 6430) in the European Ecological Network (European Commission 2013).This is a very unfavorable and undesirable phenomenon because ecotones (referred to as riparian buffer strips) are particularly important for the preservation of biodiversity, particularly in the agricultural landscape (McCracken et al. 2012;Stutter, Chardon, and Kronvang 2012).E. lobata growth in ecotones most often coincides with the common reed Phragmites australis and tall nitrophilic plants such as Urtica dioica and Cirsium arvense.It competes mainly with native vines, forcing out Calystegia sepium and Humulus lupulus (Kołaczkowska 2016).
E. lobata is particularly widespread in Central and Eastern Europe (Protopopova and Shevera 2014;Protopopova et al. 2015).In Poland, it is one of the fastest spreading IAPS, which requires the implementation of strategies to limit further growth and control established populations.The transport of seeds by flowing water, combined with their rapid growth rate, facilitates the colonization of shorelines and spread in river valleys (Hulme et al. 2009;Protopopova et al. 2015;Zając, Tokarska-Guzik, and Zając 2011) (Figure 1).E. lobata has stems up to 5-6 m long (Hulme et al. 2009).The stems feature numerous coiling, branched tendrils (formed from modified leaves), and light green, palmately lobed leaves.Flowers are inconspicuous (corolla ca. 5 mm long) and dioecious with a white-greenish color.Female flowers grow singly while male flowers are numerous, collected into long, and erect inflorescences (racemes or panicles) with each measuring 5-30 cm long by 2-8 cm wide (Tokarska-Guzik 2022).
In the climate conditions of Central Europe, the plant begins to germinate in late April/early May (Zając, Tokarska-Guzik, and Zając 2011).In the spring (May), the vegetative growth of juvenile individuals is observed (Figure 1A).At the beginning, the size of the individuals and their abundance are small, but later, during the next two months, a very fast biomass increase is observed.Being a photophilous species, it twines around other herbaceous plants, bushes, and trees, significantly limiting their access to light (Grundy 2004).The species reaches the generative phase of development in July (Figure 1B).The vines bloom from the middle of July to the end of August.Flowering is very abundant.Flowers are greenish to white, and both sexes can be found on the same plant (Hulme et al. 2009).The leaves begin to turn yellow at the end of September (Figure 1C).At the same time, the individuals differ from each other by the percentage of dead and living shoots (Figure 1).E. lobata shoots wither and die in the autumn (October/ November), while the fruit may remain on withered shoots through the winter (Tokarska-Guzik 2022).

Study area
Research on the classification of E. lobata was carried out in an area of 6 km 2 (7.5 km x 0.8 km) in the Bzura River valley in Central Poland (N 52°07"47"/E19° 34"26;" Figure 2), in a protected NATURA 2000 site: Pradolina Bzury-Neru (Bzura-Ner Ice-Marginal Valley) PLH10006.The selected river valley had representative vegetation for the majority of the transformed lowland river valleys of Central and Eastern Europe (Forysiak 2012;Kopeć et al. 2014), which are currently the most threatened by the invasion of E. lobata.
The area is characterized by an average annual air temperature of about 9°C.The average daily temperature reaches its minimum value in January and maximum in July.The total annual rainfall in the study area is about 650 mm.The highest monthly rainfall amount occurs in July and is lowest in January (Ustrnul et al. 2021).

Remote sensing data
The RS data were collected with two hyperspectral cameras (HySpex VNIR-1800 and SWIR-384) integrated into a single aircraft (Sławik et al. 2019).The data collection was performed three times in 2016 under cloudless conditions: spring (24 th May), summer (22 nd July), and autumn (9 th September).The data acquisition time was determined based on phenological and structural traits, aiming to capture three main stages of the species development to assess the possibility of detecting the target species at different phases of the life cycle.The flyover was conducted by Cessna CT206H aircraft at an altitude of 730 m AGL (Above Ground Level) and an airspeed of 59.2 m/s, acquiring data in four flight paths (Figure 2).Both HS cameras varied in spatial resolution, HySpex VNIR-1800 0.4-0.9µm with 0.49 GSD (Ground Sampling Distance) and HySpex SWIR-384 0.9-2.5 µm with 1.07 GSD.
The collected HS aerial data were pre-processed to create a mosaic of images for each season: spring, summer, and autumn.The combined data from both sensors were subjected to the necessary procedures (Sławik et al. 2019): radiometric calibration, geometric correction in PARGE software (Schläpfer and Richter 2002), and atmospheric correction using the MODTRAN5 radiative transfer model in the ATCOR4 software (Richter and Schläpfer 2016).The processed images were mosaicked into a single file with a spatial resolution of 1.00 m using PARGE software.The "Center Cropped" option was used to find the middle of the overlapping areas between the images.The bands with wavelengths longer than 2350 nm (which contained high noise) were cropped, leaving 430 bands that were subjected to the spectral polishing process using the Savitzky -Golay algorithm with a range of 13 bands.With so many spectral bands, it was essential to perform dimensionality reduction (Green et al. 1988).In order to eliminate noise and select the most informative bands, the hyperspectral mosaics were subjected to MNF (Minimum Noise Fraction) transformation (Green et al. 1988) using ENVI software (v.5.5).
The thirty most informative channels were selected for further analysis.The data processing resulted in three HS data sets in the form of the first 30 MNF channels for each flight (Figure 3).This produced data sets ready for classification from spring (H1), summer (H2), and autumn (H3).Then, to check the informativeness of the multitemporal data fusion, layers in two or three seasons were combined, obtaining sets consisting of 60 or 90 MNF channels (H12, H23, H13, H123; Figure 3).Data processing resulted in seven HS data sets.

On-ground reference data
To classify E. lobata, on-ground reference data (reference polygons) were obtained simultaneously with the acquisition of HS data.The polygons were marked in situ by botanical specialists.Field surveys were conducted three times: spring − 24-26 May (reference dataset no. 1 = R1), summer − 22-24 July (reference dataset no. 2 = R2), and autumn − 9-11 September (reference dataset no. 3 = R3).During the field survey reference polygons were established for both E. lobata sites and for the background (absence of E. lobata).Each reference polygon had fixed-area plots (Stehman and Czaplewski 2003) in the shape of a circle with a radius of 2 m, which gave approximately 15 pixels of HS data per polygon.Before the field survey, a preliminary reconnaissance of the area was carried out to determine the distribution of E. lobata patches and the diversity of the surrounding vegetation.During this reconnaissance, 10% of the study area was noticed to be covered by E. lobata.To preserve the statistical rigor of assessment accuracy, sampling was placed in the design-based inference, involving stratified sampling with equal allocation.
During the field survey, geolocation and information about the coverage percentage of E. lobata in patches were noted and divided into dead and living shoots.Coverage by dead and living shoots was determined with an accuracy of 10%.The percentage of plant coverage in patches was estimated in a horizontal projection, assuming that the total coverage of all plants in each polygon (live shoots of E. lobata + dead shoots of E. lobata + other plants) was 100%.Only the parts of shoots which were visible from the top were taken into account to match what the RS sensors may register.Geolocation of onground reference polygons was recorded with a Global Navigation Satellite System (GNSS) using GPS MobileMapper 120 with real-time differential correction and measurement accuracy between 0.2 and 0.4 m.
During the first campaign (spring), 85 reference polygons for E. lobata and 140 background polygons were established.Places with the highest coverage of shoots were selected.During the subsequent summer and autumn campaigns, each of the 85 polygons for E. lobata were visited and reestablished in a place with the highest shoot coverage.Thus, the location of a given polygon may have shifted between survey campaigns (Figure 4).The localization of these polygons could change to encompass the area with the highest shoot density and keep the best purity of pixels possible (Figure 4).In addition to reestablished reference polygons for E. lobata, the percentage coverage of investigated invasive vine (and other plant species) were again recorded.On average, the reference polygons between spring and summer were shifted by 1.86 m, and between summer and autumn by 0.92 m.The average difference in the location between spring and autumn was the highest and amounted to 2.23 m.
E. lobata shows a high growth rate in established 85 reference polygons.In the spring, the plants were usually small, in the early stage of growth.For this reason, patches with only 30-40% coverage of E. lobata accounted for nearly 60% of the on-ground reference dataset (Supplementary Materials, Figure S1).The average percentage coverage of E. lobata in the reference polygon was 44%.All individuals were in the vegetative stage, and only live shoots were found.Other plant species occupied the remaining part of the polygons on which the E. lobata would climb.In the summer, E. lobata reached its maximum size and completely suffocated other plant species growing in the vicinity.Polygons with 90-100% of E. lobata coverage dominated (Supplementary Materials, Figure S1).The average coverage percentage of E. lobata in the reference polygon was 88%.All individuals were in the generative stage.A flowering peak was observed, and only live shoots were found.
The fruiting period and the early yellowing and senescence of E. lobata leaves were observed in autumn, which resulted in the reduced surface area covered by the studied species compared to summer.The average percentage coverage of E. lobata in the reference polygon was 65%.The process of dying was relatively advanced in most polygons.Only 25% of polygons were more than 30% covered by the living shoots.For comparison, in almost 35% of polygons, the percentage of dead shoots was above 40% (Supplementary Materials, Figure S1).During the autumn, the E. lobata was in the fruiting stage in all reference polygons.
In addition to determining polygons for E. lobata, 140 background polygons were established.Reference polygons for the background were placed randomly while considering the previously established sampling design, including stratified sampling with equal allocation.Diversity of background concerned all types of vegetation found in the study area, including pastures, meadows, wetlands (that may be invaded by E. lobata), and communities of tall herbs with Calystegia sepium, Humulus lupulus, Phragmites australis, Urtica dioica, and/or Cirsium arvense.All background polygons were located during the first campaign (spring).In subsequent summer and autumn campaigns, the location of the background polygons did not change, but they were checked if E. lobata shoots did not occur in these polygons.

Spectral variability of Echinocystis lobata
In order to show the spectral variability of the E. lobata and the background within established on-ground reference polygons (from spring, summer, and autumn), mean spectral signatures for aerial HS mosaic data (430 spectral bands) were performed for all three campaigns.The curves were averaged based on data obtained from reference polygons (from a pixel located entirely within a given reference polygon), which were separated from the studied species and eight background subclasses.In order to generate the averaged curves, 40 reference polygons of E. lobata with the highest shoot coverage were selected.For each background subclass, 15-40 polygons were used (depending on the availability of references for a given subclass).This was done to verify the wavelengths that differentiate E. lobata from both the main species comprising the ecotone (Calystegia sepium, Humulus lupulus, Urtica dioica, Phragmites australis, and Cirsium arvense) and the major types of vegetation prevailing in the studied areas (meadow, farmland, and wetland).The data distribution was checked by normal probability plots with the Shapiro-Wilk test.The significance of reflectance differences between the background and E. lobata was tested separately for each of the 430 spectral bands (from 400 to 2500 nm) using the Kruskal-Wallis and post-hoc tests.The analyses were conducted using Statistica software (v.13).For all the statistical analyses, the significance level was α = 0.05.

Classification of airborne data and accuracy assessment
Supervised binary classification with the Random Forest (RF) algorithm was applied as a method of classification (Breiman 2001).All the processing was designed on a pixel basis.The reference dataset was prepared in the vector format of polygons.Two legend classes were used in the classification: the E. lobata class and the background class (absence of E. lobata).The classifications were carried out with the following parameters: (i) 100 trees learned for each model; (ii) the minimum number of samples per split = 2; (iii) the minimum number of samples at leaf node = 1.
The trained RF model was then applied to all pixels contained by the validation polygons.The confusion matrix was built from the reference and predicted classes of all validation pixels (Wulder et al. 2006).A confusion matrix reported with reference sample proportions 10% E. lobata: 90% background (See Section 2.4; Stehman and Czaplewski 1998) was the basis for estimation of overall accuracy (OA), F1, producer's accuracy (PA), and user's accuracy (UA) for E. lobata and the background class, taking into account the unequal inclusion probabilities.To compare the classification results between the different scenarios we primarily used the F1 measure, because it allowed us to evaluate the classification result separately for each of the two classes in a synthetic way (considering both underestimation and overestimation error).For each scenario, cross-validation of the reference dataset was carried out using the Stratified Random Sampling method (Rudnicki et al. 2006) in a 50/50 split with 50-fold repetition.In each iteration of polygon draws, all pixels from a particular polygon participated in only a training or validation dataset.
To answer all the objectives of the research, a series of classifications were carried out according to 18 scenarios that differed in the set of reference polygons and that used various HS datasets (Figure 5).
In the case of reference data, the following sets were distinguished, depending on which campaign they came from: R1 (from spring), R2 (summer), and R3 (autumn).This was similar in the case of HS data, where the following sets of data were distinguished: H1 (spring), H2 (summer), and H3 (autumn).
To answer the first research question, which was to determine the impact of acquiring data at different phases of phenological development of the plant on the classification results, the following scenarios were designed: R1H1 (where R1-indicates reference data from spring, H1-indicates HS data from spring), R2H2, and R3H3.
To answer the second research question, which concerned the impact of the lack of synchronization between the acquisition of reference data and HS data, scenarios R1H2, R1H3, R2H1, R2H3, R3H1, and R3H2 were created.Their results were compared with scenarios from the first research question (R1H1 with R2H1 and R3H1; R2H2 with R1H2 and R3H2; R3H3 with R1H3 and R2H3).
To answer the third research question, related to the impact of using multitemporal data for the result of classification, the following scenarios were planned: R1H12, R1H13, R1H123, R2H12, R2H23, R2H123, R3H13, R3H23, and R3H123.Their results were compared with the results from the first research question (R1H1 with R1H12, R1H13, and R1H123; R2H2 with R2H12, R2H23, and R2H123; R3H3 with R3H13, R3H23, and R3H123).
Statistical tests determined the significance of the difference between individual scenarios.The normality was checked using the Kolmogorov-Smirnov test, and the homogeneity of variance by using the Levene's test.After that, the parametric ANOVA test and the post-hoc Tukey's tests, with the significance level set at 0.05 were performed.The above comparisons were carried out using the Statistica software (v.13).
The whole classification workflow was performed in the R programming language, using YAML-based experiment definition language.Utilizing its multiclassification functionality allowed the automatic execution of thousands of individual classifications with consistently applied settings, thus enabling reproducibility and comparison of the results.

Spectral variability of Echinocystis lobata during the growing season
Both in spring and autumn, there were no bands in the range from 400 to 2500 nm where E. lobata reflectance values differed significantly from allidentified background subclasses (Figure 6).Differences were found for data collected in the summer but they don't apply to two of all analyzed classes: Calystegia sepium, Humulus lupulus.E. lobata showed significantly higher reflectance values in the ranges: 0.51-0.58nm; 0.60-0,62 nm; 0.70-1.12nm.The comparison of mean spectral reflectance curves for E. lobata among the three campaigns shows that reflectance values significantly differ for spring, summer, and autumn in the following ranges: 1) 0.52-0.57nm; 2) 0.66-0.68nm; 3) 0.74-0.93nm (Figure 7).

The impact of data acquisition time on Echinocystis lobata classification results
The classification results showed that E. lobata occurred in the study area mainly along the riverbed.An example of the spatial structure of E. lobata population found on the resulting maps is shown in Figure 9 More extensive sites have not been identified.To assess the effectiveness of E. lobata mapping, in all three different phenological periods of its development (Figure 1), the following three scenario results were compared: R1H1 (where R1-indicates reference data from spring, H1-indicates HS data from spring), R2H2, and R3H3.
These scenarios assumed the classification of onground reference and HS data from the same campaign.Each of the analyzed accuracy measures (OA, F1, PA, UA for E. lobata, and the background class) assumed significantly different values in the tested scenarios.The results of PA and UA for E. lobata and the background for those scenarios were presented in Supplementary materials as Table S1.High accuracies for the background class (PA and UA within 0.96-0.99) in all analyzed scenarios indicate the exact precision of the background classification.The highest classification accuracies (OA = 0.97 ± 0.1, F1 = 0.87 ± 0.04 for E. lobata, and F1 = 0.99 ± 0.01 for the background) were achieved for the R2H2 scenario (E.lobata flowering period).Also in this scenario, the highest accuracies were obtained for PA (0.88 ± 0.26) and UA (0.85 ± 0.08) for E. lobata class.The lowest values of accuracy measures (F1 = 0.64 ± 0.04 for E. lobata, PA = 0.64 ± 0.39 and UA = 0.67 ± 0.37) were obtained for the R1H1 scenario (Table 1, Figure 8).The difference in F1 accuracy for E. lobata between R2H2 and R1H1 is 0.23.For autumn (scenario R3H3), intermediate values of classification accuracies were obtained (OA = 0.95 ± 0.01, F1 = 0.75 ± 0.05 for E. lobata, and F1 = 0.98 ± 0.01 for the background).The decrease in value was also observed in UA = 0.68 ± 0.12 for E. lobata class (Supplementary materials, Table S1).The F1 E. lobata accuracy for the R3H3 scenario was significantly higher compared to R1H1 (the difference between R3H3 and R1H1 is 0.11) and significantly lower than R2H2 (the difference between R3H3 and R2H2 is 0.11; Figure 8).Analyzing the map of the classification results confirms the results of the accuracy measures (Figure 9).

The impact of lack of synchronization while obtaining hyperspectral and on-ground reference data collection on Echinocystis lobata classification results
To assess the importance of convergence of the acquisition time of on-ground reference and airborne HS data on E. lobata classification, the results for scenarios with synchronized data acquisition were compared to scenarios with no synchronization.The effect of this lack of synchronization on classification results was checked for each airborne campaign separately.
Of the three measures of accuracy used -OA, F1 for E. lobata, and F1 for background -only F1 for E. lobata differed significantly between scenarios.
In the case of classifying the species with HS data from spring (H1), those scenarios produced the worst results in synchronization (see Section 3.2).Using the reference data from summer (R2) or autumn (R3) worsened the impact on classification results in relation to data obtained synchronously (F1 score dropped in R2 by 0.04 and in R3 by 0.08, compared to R1H1; Figure 10).This was observed in a comparison of analyzed accuracy measures (F1 Table 1.The comparison of average accuracy obtained from classification for three scenarios: R1H1, R2H2, and R3H3, according to values: OA -overall accuracy, F1 for Echinocystis lobata and F1 for background class.Abbreviations: on-ground data collection in: R1 (spring), R2 (summer), R3 (autumn).Airborne hyperspectral data acquisition: H1 (spring), H2 (summer), H3 (autumn).score for E. lobata) from scenarios R1H1, R2H1, and R3H1 (Table 2).Also, low PA and UA values were obtained for R2H1 (PA = 0.59 ± 0.35, UA = 0.63 ± 0.57) and R3H1 (PA = 0.58 ± 0.50, UA = 0.57 ± 0.39; Supplementary materials, Table S2).For the OA and F1 background accuracy measures, the decrease in accuracy of R1H1 to R2H1 and R3H1 was also significant, but there was no significant difference between the R2H1 and R3H1.
In the case of classification of HS data from summer (H2), which brought the best results in synchronization (see Section 3.2), lower F1 scores for E. lobata (0.07 significant difference) were obtained when reference data from the spring campaign (R1) were used, compared to R2H2 (Table 2).A statistically significant decrease in E. lobata F1 values (0.04 difference) was also observed when the reference data from autumn (R3) were used compared to R2H2.Equally high PA values were obtained for those scenarios: R1H2 (0.87 ± 0.31) and R3H2 (0.87 ± 0.24), but lower for UA values: R1H2 (0.75 ± 0.16) and R3H2 (0.78 ± 0.16; Supplementary materials, Table S2).For the OA and F1 background accuracy measures, the decrease in accuracy of R2H2 to R1H2 and R3H2 was also significant, but there was no significant variance between the R1H2 and R3H2 scenarios.
When analyzing the classification results obtained from the autumn HS data (H3), the lack of synchronization on the classification results did not have an impact on the accuracy of the results in each scenario (Table 2).Also, for UA and PA in E. lobata class, no significant differences were found.The usage of reference data from spring (R1) caused a significant decrease in accuracy while using data from the summer (R2) did not have a statistically significant impact on classification accuracy (Table 2, Figure 10).
When comparing all scenarios based on the use of single raster data, whether a given scenario assumed convergence in collected data, the best results were obtained for scenarios R1H2, R2H2, and R3H2, in which the airborne HS data from summer (H2) were used (Figure 10).In turn, the worst results were obtained for R1H1, R2H1, and R3H1, in which the airborne HS data from spring (H1) were used (Figure 10).

The effect of hyperspectral multitemporal image fusion on Echinocystis lobata classification results
To improve the classification results performed with on-ground reference and HS data obtained in spring (H1), adding HS data from summer (R1H12) was crucial (Figure 11).When comparing the results of the discussed scenarios, there was a significant increase in the value of all analyzed accuracy measures (OA increased by 0.04, and F1 score for E. lobata increased by 0.20; Table 3).A high increase in PA (0.91 ± 0.15) and UA (0.79 ± 0.14) was also observed in this scenario for E. lobata class (Supplementary materials, Table S3).Comparing the HS data set from spring and summer (R1H12), the data combination acquired in spring, summer, and autumn (R1H123) gave comparable results and did not produce any additional improvement in the values of used accuracy measures.The scenario in Abbreviations: on-ground data collection in: R1 (spring), R2 (summer), R3 (autumn).Airborne hyperspectral data acquisition: H1 (spring), H2 (summer), H3 (autumn).ANOVA was used to test the significance of the difference between the scenarios.Based on the post-hoc Tukey's tests, scenarios between which there is no difference were marked with the same letter.

Accuracy measures
Reference scenario The differences between reference scenarios which the HS data set derived from spring and autumn (R1H13) also increased accuracy (by 0.13) but not as much as R1H12 (Figure 11).The results of the classification made on the reference and HS data from summer (R2H2) had the highest values in OA and F1 scores for E. lobata and background (Section 3.2).Adding additional HS data from other campaigns (R2H12, R2H23 and R2H123) improved the classification results.However, this increase in accuracy reached a maximum of 0.04 for F1 E. lobata (Figure 11).Generally, the highest mean values were obtained for the R2H123 scenario (for E. lobata F1 = 0.91 ± 0.03; Table 3).Also, there was a slight increase in PA for E. lobata for scenarios compared to R2H2.The highest Table 3.The comparison of scenarios R1H1 with R1H12, R1H13, and R1H123; R2H2 with R2H12, R2H23, and R2H123; and R3H3 with R3H13, R3H23, and R3H123 according to values: OA -overall accuracy, F1 for Echinocystis lobata and F1 for background class.Abbreviations: on-ground data collection in: R1 (spring), R2 (summer), R3 (autumn).Airborne hyperspectral data acquisition: H1 (spring), H2 (summer), H3 (autumn).

Accuracy measures
Reference scenario The differences between reference scenarios was recorded for R2H123 and reached PA = 0.96 ± 0.09.
No statistically significant changes in UA were observed for these scenarios (Supplementary materials, Table S3).
In the case of classifications made by using onground reference and HS data set from autumn (R3H3), the addition of HS data from other campaigns (H1 and/or H2) had an impact on the classification results.The addition of data from spring (R3H13) slightly improved the results (F1 for E. lobata increased by 0.03; Table 3).Larger changes were observed when the HS data set from summer (R3H23) was added to the classification (F1 score for E. lobata increased by 0.11).In this scenario, PA was 0.93 ± 0.23 and UA reached 0.81 ± 0.07.Using HS data from all three campaigns showed comparable results between R3H123 and R3H23.In both cases (R3H123 and R3H23), statistically significant differences were obtained for accuracy measures of F1 for E. lobata, compared to the R3H3 scenario (Figure 11).In contrast, for OA and F1 for background, the changes in accuracy between R3H3 and R3H13 were not significant.Only the addition of data from the summer (H2) was significant.
Comparing the results of all the scenarios, the highest values of all analyzed measures were obtained for the R2H123 scenario where F1 = 0.90.However, it should be emphasized that the accuracy of F1 for E. lobata was only 0.04 better than those obtained in summer (R2H2; Figure 11).

The impact of the development stage of annual vine on the possibility of classification
This paper presents the results of Echinocystis lobata classification.There are no other examples available in the literature related to E. lobata mapping with RS methods.At the same time, the potential possibility of distinguishing the E. lobata from the surrounding vegetation based on spectral signatures was indicated by Grašič et al. (2019).Generally, there is relatively little data on the mapping of invasive annual climbing plants, which can be ascribed to the complexity of this issue.One example is the object-based classification of the annual vine Persicaria perfoliata (Zhou et al. 2021).The analyses were performed using features obtained from UAV-based RGB images.In the best scenario, the RF algorithm with 53 selected features was used, and the following results were obtained: OA = 90.73,PA = 80.00, UA = 100.Our results are slightly lower, but it is possible that using additional raster data, e.g.spectral indices or products from ALS, would improve the results.This will be the subject of further research.Perennial invasive vines, such as Pueraria montana (Cheng, Tom, and Ustin 2007;Peters 2016;Liang et al. 2020), Mikania micrantha (Chen, Lin, and Sun 2014;Dai et al. 2020) and Lygodium microphyllum (Wu et al. 2006), are more often the subject of mapping.Usually, they occupy vast areas and their shoots are visible throughout the growing season.For this reason, they can be classified using multi-or hyperspectral data obtained from the satellite level (Cheng, Tom, and Ustin 2007;Chen, Lin, and Sun 2014;Peters 2016;Dai et al. 2020;Liang et al. 2020).The results obtained for the listed species are as follows: Pueraria montana: OA = 80.18, UA = 83.02,PA = 73.26(Cheng, Tom, and Ustin 2007), PA = 0.94, UA = 0.96 (Liang et al. 2020); Mikania micrantha: OA = 87.5,UA = 81.1,PA = 83.3(Chen, Lin, and Sun 2014).
The obtained results indicate that the appropriate dates (in terms of plant development and phenology) of on-ground reference data and HS data collection are crucial for the effective mapping of species with such dynamic growth during the year, such as E. lobata.In the spring, E. lobata was in the vegetative stage, did not form compact carpet-like patches, and was outgrown by other species.At this phenological stage, the classification of E. lobata based on airborne hyperspectral images was the least effective (Table 1).Due to the late development (Hulme et al. 2009), E. lobata reached smaller cover-abundance values in the polygons in spring (usually below 50%) and cooccurred with other larger native perennials.Consequently, polygons with high spectral variability within the species class dominated among the training polygons as a result of mixed characteristics in pixels, i.e. those derived from E. lobata as well as those derived from the surrounding vegetation and litter.This is represented in the mean spectral signatures (Figure 6A).For spring, there were no wavelengths for which the reflectance coefficient significantly differentiates E. lobata from background subclasses.
The best results of E. lobata classification came from the summer (Figure 4; Figure 8).Due to the intense growth of shoots, it successfully outgrew and suffocated other co-occurring plant species.This had a positive effect on the purity of the spectral signature used for training, which is very important to get a good classification result (Kopeć et al. 2019;Gholizadeh et al. 2022).There is a possibility that two factors were crucial in E. lobata classification: mass flowering and growth of shoots.Spectral reflectance curve analysis from summer indicates that E. lobata significantly differs from the background (except Calystegia sepium, Humulus lupulus) with higher reflectance for wavelengths: a) 0.51-0.58nm and 0.60-0,62 nm indicating higher chlorophyll content in the studied species compared to the background (Gitelson and Merzylak 1997); b) 0.70-1.12nm, indicating high biomass, good health condition of the plants (Datt 1999) and higher tissue water content compared to background subclasses (Figure 6B; Aldakheel and Danson 1997).The features of E. lobata found from the analysis of mean spectral signatures align with the main features of alien invasive species described in the literature, giving them a competitive advantage over native species (Richardson and Pyšek 2012).Despite the mass flowering of E. lobata during the summer (small white flowers in relatively abundant inflorescences), there is no clear representation of this in the mean reflectance coefficient values (Figure 6B) for visible spectrum wavelengths.Therefore, it is impossible to confirm that flowering was a crucial factor in the species facilitating its classification based on hyperspectral imagery, as mentioned in other research (Hunt et al. 2007;Somodi et al. 2012;Hunt et al. 2004).Another factor that positively affected the classification results during the summer was that all individuals of E. lobata were in the same development phase.Based on other species, it has been shown that phenological variation and differences in the advancement of the development phase may reduce the species' spectral uniqueness and reduce the accuracy of the classification results (Andrew and Ustin 2008;Somodi et al. 2012;Dudley et al. 2015;Meerdink et al. 2019).
The classification results based on autumn data indicate that there is also the possibility of effective detection of the studied annual vine in this period.In the autumn, the species was in the fruiting stage, and its leaves became discolored and withered (Figure 1).The phenomenon of plant discoloration is often used to increase the effectiveness of IAPS mapping (Hunt et al. 2007).The described species discolors and dies earlier than other co-occurring plant species from the background (Kołaczkowska 2016).However, these differences are not visible in the mean curves of the spectral reflectance (Figure 6).The reason for the lack of difference may be due to variations in the advancement of the withering and dying processes in individual polygons.Researchers emphasize strong spectral differences between green and senescent parts of plants (Andrew and Ustin 2008).However, the differentiation of phenological stages of a given species creates intraspecific variation that may reduce the spectral uniqueness between co-occurring species (Hestir et al. 2008).Another reason was the secondary outgrowth of the studied vine plant by co-occurring species that became exposed after the withering of E. lobata shoots.This resulted in the decreased average cover of E. lobata in the polygons during the autumn compared to summer (Figure 4).The obtained results clearly indicate how important it is to know the individual functional traits of a given plant species before planning classification work.This allows the indication of the time of the growing season in which a given species can be effectively distinguished from the surrounding vegetation (Niphadkar and Nagendra 2016).
The use of raster data with a high spatial resolution (pixel size minimum 1 m 2 ) is extremely helpful for mapping E. lobata in the early stages of the invasion, where patches of this species usually reach small sizes.However, in places where the invasion of this species is more advanced, raster data with lower spatial resolution can also be used.For example, in Ukraine, E. lobata sometimes creates large thickets, even 250-300 m in diameter (Protopopova et al. 2015).Such a large variation in the size of patches of the studied species enables mapping at a spatial resolution, which can be collected from aerial or satellite levels.The choice of an alternative image data source should also include spectral data resolution.The results from summer (Figure 6) indicate that the wavelengths from 0.5 to 0.9 nm, i.e.VIS and NIR, are crucial for the correct mapping of the studied species.The EnMAP hyperspectral satellite (Guanter et al. 2015) seems to be the most dedicated to these applications.This instrument covers a 30 km swath width with a spectral range from 420 to 2450 nm using two spectrometers: VNIR (420-1000 nm, with 89 bands) and SWIR (900-2450 nm, with 155 bands), and is characterized by a relatively short return time (27 days;Guanter et al. 2015).The wavelength range crucial for E lobata mapping is also recorded by Landsat-8, Sentinel-2 and WorldView-3 satellites.Further research is needed to explore the possibility of mapping E. Lobata using satellite data.

The impact of the synchronization lack between on-ground reference and hyperspectral data on the possibility of Echinocystis lobata classification
The effect of the lack of data synchronization on classification results was checked for each airborne campaign (H1, H2, and H3) separately (Figure 10, Table 2).
It is a crucial issue, because E. lobata is characterized by spreading shoots during the growing season, and on-ground data collected in the summer could have polygons in places where E. lobata shoots were absent in the spring (during H1 acquisition).This was proven by counting the number of meters the reference polygon had to be moved between campaigns to be in the right location relative to this vine plant (Section 2.4, Figure 4).The classification results for E. lobata on synchronized HS and on-ground reference data showed significantly higher OA and F1 values than offset HS with on-ground reference dates (Figure 10).The use of HS data obtained on a different date from on-ground reference data effectively reduced the classification accuracy measured by the F1 score for E. lobata by more than 4% (Figure 10; Table 2).In this case, the lower negative impact of lack of synchronization can be explained by the greater overlap of reference polygons between R2 and R3 (Section 2.4).The R2H2 scenario stands out as the best result, while the map produced from the R1H1 data classification has a clear overestimation of E. lobata, and the R3H3 data classification led to an underestimation of E. lobata distribution (Supplementary materials, Table S2).Meerdink et al. (2019) came to similar conclusions by analyzing individual species, plant communities, and various land cover, stating that using spectra from the same term gave high Kappa scores of 0.80 and 0.86.Where data from the other terms were combined, mean 0.78-0.83Kappa values were obtained.Despite the fact that these were satellite data with 18-m spatial resolution, the authors also concluded that using spectral images obtained without synchronization of RS with on-ground reference data cannot be reliably used for species mapping.Results in our study show comparable, or even higher, differences in F1 score values between different scenarios (Table 2).This may be because E. lobata is very dynamic in its development during one vegetation season.Similar conclusions were drawn by Hesketh and Sánchez-Azofeifa (2012), by analyzing 47 liana and tree species, stating that the results across seasons significantly lowered the classification results by lowering the factor value by 10, whereas a high level of accuracy within a given season was observed when airborne and on-ground data were synchronized.

The impact of using multitemporal image fusion on improved Echinocystis lobata mapping
Selecting the time of optimal imaging acquisition is a key factor in determining the correctness of the species classification.It is important to capture the features that distinguish the spectrally analyzed element of plant cover from its background (Müllerová et al. 2017).The practice of using multitemporal image fusion in vegetation and plant species classification is well known and is mainly used in satellite data analyses.These studies mostly concern the classification of forests and trees (Somers and Asner 2012;Evangelista et al. 2009), plant communities (Roelfsema et al. 2014), grains (Vuolo et al. 2018), and invasive herbaceous perennials (Singh and Glenn 2009).There are also studies using airborne data in multitemporal data to classify vegetation and plant species.In this way, perennial species of invasive and expansive grasses were studied, like Bromus tectorum (Noujdina and Ustin 2008) and Calamagrostis epigejos (Sabat-Tomala, Raczko, and Zagajewski 2022).Other researchers who used a multitemporal approach found that image fusion improved the results, using both satellite (Somers and Asner 2013) and airborne data (Dudley et al. 2015).
In this study, an attempt was made to classify the annual vine species E. lobata with high growth dynamics using multitemporal image fusion.The biggest difference (increase) in results was observed only when H2 was added to H1. Adding H2 data produced the greatest positive improvement in classification results, but ultimately these values were comparable to a single H2 (Figure 11; Table 3).In other cases, increases in accuracy were not statistically significant.We conclude that it is unnecessary to make a multitemporal image fusion, especially in the case of E. lobata.It is more important to reach the peak of species development and obtain references synchronously as this can present a significant additional cost, especially in the case of airborne data.However, this method may be helpful when free of charge satellite data (like Sentinel or Landsat) are used in plant monitoring.Since the acquisition of satellite data is not as convenient as airborne in terms of synchronization with the accurate plant's phase development, multitemporal data fusion may be a solution for highly effective plant classification.For example, Somers and Asner (2013) showed a systematic increase in invasive tree detection by comparing multitemporal fusion results (Kappa = 0.78) with those of the traditional singleperiod approach (Kappa = 0.51-0.69).It was confirmed that redundant spectral information can be avoided for better plant detection accuracy.Also, Dudley et al. (2015) achieved high values on vegetation classification using satellite data and multitemporal data fusion.Results indicate that multitemporal, seasonally mixed spectral libraries achieved similar overall classification accuracy compared to single-date libraries, and in some cases, resulted in improved classification accuracy (Dudley et al. 2015).Despite the high results in the classification of invasive trees and general vegetation, this method still needs improvement in the detection of annual invasive grasses.Weisberg et al. (2021) showed that the combination of all bands and all data acquisition dates resulted in OA = 64.3% and Kappa = 0.56 for two spectrally similar invasive grasses distinguished from native vegetation.The availability of near-infrared images turned out to be less important than true-color RGB images collected at appropriate time periods.Thus, multitime information can be a substitute for more extensive spectral information, as presented in this study (Figure 6).

Conclusions
The main conclusion is that mapping annual vine IAPS using RS and ML is possible and highly effective, provided the hyperspectral and on-ground reference data are obtained in strict synchronization and the appropriate phenological phase.Even in the case of E. lobata, which is a very expansive plant and grows rapidly during one vegetation season, according to our research, it is the period of flowering (summer) with the greatest coverage of the area with shoots.The results from summer indicate that the wavelengths from 0.5 to 0.9 nm (VIS and NIR) are crucial for the correct mapping of the studied species.This information may be essential when satellite data would be used in IAPS monitoring.
Multitemporal image fusion of HS did not have a significant impact on the improvement of the classification results.However, this method can be used in practice, if the data are obtained from the satellite (like EnMAP), but RS data are not synchronized with a particular phase of the plant's development.Multitemporal image fusion is a solution, e.g. when RS data was obtained at a different time due to the problem of acquisition of satellite scenes aligned with the peak of the plant's phenological development.It could be a frequent problem because of the time acquisition of satellite data and the high cloudiness of the study areas in Central Europe.However, further research is needed to explore the possibility of mapping E. lobata using, e.g.satellites, like EnMAP.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
The

Figure 2 .
Figure 2. Location of the study area and distribution of the reference polygons (based on the Digital Terrain Model), highlighting planned flight paths in the study area (based on the RGB orthophoto map).

Figure 3 .
Figure 3. Diagram showing the process of preparing hyperspectral data from three single flights (spring, summer, autumn) to multitemporal data fusion.

Figure 4 .
Figure 4. Examples of changing the location of reference polygons in subsequent measurement campaigns (A -spring, B -summer, C -autumn), in relation to the direction of plants' shoots growth.

Figure 5 .
Figure 5.A flowchart of the classification process.

Figure 7 .
Figure 7. Mean spectral signatures of Echinocystis lobata in spring (A), summer (B), and autumn (C) campaigns from hyperspectral images using reference polygons.Wavelength ranges where the reflectance coefficient differed significantly for spring, summer and autumn are shown in grey (Kruskal-Wallis test: n.S. α = 0.05).

Figure 8 .
Figure 8.The comparison of scenarios R1H1, R2H2 and R3H3 according to F1 values for Echinocystis lobata for a given scenario.Boxplots were created from fifty Random Forest classifications for each scenario.Line -indicates the average value; box -represents the first and the third quartile; plot -the highest and the lowest values in each scenario.Abbreviations: on-ground data collection in: R1 (spring), R2 (summer), R3 (autumn).Airborne hyperspectral data acquisition: H1 (spring), H2 (summer), H3 (autumn).ANOVA was used to test the significance of the difference between the scenarios.Based on the post-hoc Tukey's tests, scenarios between which there is no difference were marked with the same letter.

Figure 10 .
Figure 10.The comparison of scenarios R1H1 with R2H1 and R3H1; R2H2 with R1H2 and R3H2; and R3H3 with R1H3 and R2H3 according to F1 values for Echinocystis lobata for a given scenario.Boxplot was created from fifty Random Forest classifications.Lineindicates the average value; box -represents the first and the third quartile; plot -the highest and the lowest values in each scenario.Abbreviations: on-ground data collection in: R1 (spring), R2 (summer), R3 (autumn).Airborne hyperspectral data acquisition: H1 (spring), H2 (summer), H3 (autumn).ANOVA was used to test the significance of the difference between the scenarios.Based on the post-hoc Tukey's tests, scenarios between which there is no difference were marked with the same letter.

Figure 11 .
Figure 11.The comparison of scenarios R1H1 with R1H12, R1H13 and R1H123; R2H2 with R2H12, R2H23 and R2H123; and R3H3 with R3H13, R3H23 and R3H123 according to F1 values for Echinocystis lobata for a given scenario.Boxplot was created from fifty Random Forest classifications.Line -indicates the average value; box -represents the first and the third quartile; plot -the highest and the lowest values in each scenario.Abbreviations: on-ground data collection in: R1 (spring), R2 (summer), R3 (autumn).Airborne hyperspectral data acquisition: H1 (spring), H2 (summer), H3 (autumn).ANOVA was used to test the significance of the difference between the scenarios.Based on the post-hoc Tukey's tests, scenarios between which there is no difference were marked with the same letter.
work was co-financed by the Polish National Centre for Research and Development (NCBR) and MGGP Aero under the program "Natural Environment, Agriculture and Forestry" BIOSTRATEG II: The innovative approach supporting monitoring of non-forest Natura 2000 habitats, using remote sensing methods (HabitARS); project number DZP/BIOSTRATEG-II/390/ 2015.The consortium leader is MGGP Aero.The project partners include University of Lodz, University of Warsaw, Warsaw University of Life Sciences, Institute of Technology and Life Sciences, University of Silesia in Katowice, and Warsaw University of Technology.ORCID Dominik Kopeć http://orcid.org/0000-0003-0831-2992