1. Introduction
Invasive alien plant species (IAS), exceeding the biogeographic borders of their natural habitats [
1,
2], destroy natural ecosystems and cause ecological and economic dysfunction [
3,
4,
5]. Therefore, it is crucial to prevent the further spread of IAS, especially in Natura 2000 sites, established to ensure the long-term sustainability of valuable species and habitats in Europe. One of the widespread invasive species that threatens the biological richness of mountain meadows and grasslands is
Lupinus polyphyllus Lindl., also known as garden lupin or large-leaved lupine. By creating dense patches and producing allelopathic substances, this species may limit the germination of native plants and be harmful to farm animals, which is why it was chosen as the subject of this study [
6,
7].
Remote sensing and machine learning methods have been increasingly used for monitoring the spread of invasive species [
8,
9]. Currently, various types of remote sensing imagery are available, and their selection involves a compromise between spatial, spectral, and temporal resolution and the spatial extent of the image [
8,
10]. Satellite data have been successfully used to identify invasive trees, shrubs, and tall perennials [
11,
12]. However, to detect herbaceous plants in the initial phases of invasion, it is helpful to use data with higher spatial resolution from airplanes or unmanned aerial vehicles (UAVs, [
13,
14]), but such analyses are oriented toward local case studies. Hyperspectral data are useful for morphologically similar plants [
15,
16,
17] due to the identification of unique spectral signatures of plant species [
18]. The use of HySpex hyperspectral data has enabled the accurate identification of wild cucumber (
Echinocystis lobata, OA
RF: 97%, F1: 0.87) [
9], steeplebush (
Spiraea tomentosa, OA
RF: 99%, F1: 0.83) [
19], purple moor-grass (
Molinia caerulea, F1
RF: 0.86), and wood small-reed (
Calamagrostis epigejos, F1
RF: 0.72) [
20]. The Cubert S185 imaging spectrometer was successfully used to identify bitter vine (
Mikania micrantha Kunth, OA
RF: 88%, OA
SVM: 84%) [
21], and the Cubert UHD-185 hyperspectral camera was used for mapping common milkweed (
Asclepias syriaca, OA
SVM: 92%, OA
ANN: 99%) [
22].
Hyperspectral data consume a lot of space on a hard drive and contain noise and redundant information [
23,
24]; the solution is the use of dimensionality reduction methods, such as a principal component analysis (PCA, [
25]), independent component analysis (ICA, [
26]), or minimum noise fraction (MNF, considering image noise and principal component analysis data variation, [
27]). The MNF method is the most commonly used, because it quickly and effectively compresses data and removes noise [
23]. In several studies [
28,
29], the use of MNF-transformed data yielded higher classification accuracy than the original hyperspectral bands. For example, using nine MNF bands to classify herbaceous plant species in Hortobágy National Park provided higher overall accuracy for the support vector machine (SVM; OA: 82%) and random forest (RF; OA: 79%) classifications than the 128 original AISA spectral bands (OA: 73% for both algorithms) [
30]. Many researchers have observed that the date of data acquisition also has a significant impact on the accuracy of invasive species identification [
9,
31,
32].
Different algorithms produce maps with different accuracies for heterogeneous spatial systems [
24]; the following algorithms have been used most often to identify invasive plants: spectral angle mapper (SAM; [
33,
34]; OA: 63–95%), mixture tuned match filtering (MTMF; [
35,
36]; OA: 64–90%), random forest (RF; [
37,
38]; OA: 84–97%), support vector machine (SVM; [
22,
38]; OA: 92–98%), and neural networks (NNs; [
22,
39]; OA: 97–99%). Owing to the high accuracy of invasive vegetation classification, the most popular methods are machine learning (ML) algorithms, such as SVM [
40], RF [
41], and NNs [
24]. For example, in identifying the kudzu vine in Georgia (USA), the SVM, RF, and NN methods gave high overall accuracies of 92%, 96%, and 97%, respectively, using AVIRIS data [
39]. ML methods are also less sensitive to unbalanced training datasets (common in invasive species identification) because they do not make assumptions about the distribution of input variables [
42]. Despite their high precision [
43,
44], neural network methods have certain limitations, such as high dependence on the amount of training data, long training times, high-performance hardware requirements, and lower interpretability of the outcomes, resulting from the use of hidden layers in the network structure [
13,
45,
46]. The use of convolutional neural networks (CNNs) and UAV images enabled the identification of seven invasive plant species with high accuracy (OA: 93%; F1 score oscillated between 0.87 and 0.99) [
13]. However, SVM and RF algorithms are used more often in the case of hyperspectral images because of their high accuracy and computational efficiency [
22,
47]. In a study comparing SVM and CNN for hyperspectral image classification [
48], it was shown that SVM with an RBF kernel gives better accuracies in land use classification (OA = 98.84%) and the result was more reliable than for the CNN method (OA = 94.01%). SVM uses a hyperplane to separate classes in a high-dimensional space with an optimal margin for class separation [
49] to distinguish spectrally similar classes well, even when noisy bands or a small training dataset are used [
50,
51]. In a study using five pixel-based classifiers to identify saltcedar (
Tamarix spp.) species on AISA hyperspectral data, the SVM algorithm achieved higher accuracies (OA: 86%) than the maximum-likelihood algorithms (MLC, OA: 84%), SAM (OA: 69%), NN (OA: 64%), and maximum matching feature (MMF; OA: 45%) [
52]. In contrast, RF is a machine learning algorithm based on decision trees and the principle of majority voting; the simple operating principle makes the RF algorithm require less processing time and less computational power in the case of heterogeneous systems [
53,
54] than other ML algorithms [
47,
55]. Both RF and SVM algorithms were successfully used to identify, among others, seven herbaceous plants (
Mikania micrantha,
Sphagneticola calendulacea,
Ageratum conyzoides,
Mimosa pudica,
Lantana camara,
Lpomoea cairica, and
Bidens pilosa) in China using 138 hyperspectral bands of spectrograph S185 (OA
SVM: 89%; OA
RF: 84%) [
56] and wood small-reed (
Calamagrostis epigejos), blackberry (
Rubus), and goldenrod (
Solidago) in southern Poland using 30 MNF bands of airborne HySpex images (OA
SVM: 91%; OA
RF: 93%) [
57].
The aim of this study was to create a reliable map of
Lupinus polyphyllus spatial distribution based on multiple classification and thresholding methods. The influence of the raster dataset on the obtained accuracies was analyzed by comparing the original hyperspectral image data (430 HySpex spectral bands) and a variable number of MNF bands (1–50 most informative bands). The second important task was to assess the informativeness of data obtained in different periods of the growing season because plant species characterize unique absorption features, which change due to phenology. The research area is located on valuable Natura 2000 meadows used for agriculture, i.e., cows graze and grass is mowed for hay, which significantly changes the plant species composition during the growing season. This is important because lupine has bacteria in its roots that synthesize nitrogen compounds, allowing the gain of a competitive advantage over surrounding plants. This study further expands the methodology presented in two previous articles, which focused on identifying goldenrods, reed grass, and blackberries in suburban areas of the Silesian agglomeration [
57,
58]. The aim of the present study was to verify whether the methodology used is repeatable in a completely different area and in the case of other invasive plants, thus confirming the potential of aerial hyperspectral data and machine learning in the identification of invasive plants regardless of their location and species composition.
3. Results
The classifications enabled identifying lupine, with a median F1 score of 0.85 in the area RJ1 and 0.83 in the area KA1. The accuracy increased with the number of MNF bands used in the classification of datasets comprising fewer than 20 bands (
Figure 4). The highest accuracies were obtained for classifications performed on about 20 to 40 MNF bands (median F1 score from 0.76 to 0.85 depending on the campaign and classification algorithm). In all analyzed cases (except RF classification in the second campaign and SVM in the third campaign), there were no statistically significant differences between classifications made for 30 and 40 MNF bands (
Figure A1 in
Appendix A). The use of more than 40 MNF bands did not significantly affect the RF classification accuracy. For the SVM algorithm, this did not significantly improve accuracy or resulted in a reduction in accuracy of up to 6 percentage points. Hence, it can be concluded that the set of 30 MNF bands is the optimal choice for lupine identification, and allows for obtaining the highest accuracies while reducing the data volume and classification time. For the RF algorithm, the interquartile distance for 25 classification iterations on different numbers of MNF bands was smaller than that for the SVM algorithm. Therefore, the results obtained were less variable between the iterations.
The analysis of the eigenvalue graphs generated during the MNF transformations also showed that approximately the first 30 bands of the MNF transformation were the most informative (
Figure 5). For the data with respect to each tested campaign, the curves flatten out for more than 30 MNF bands.
The use of 20–40 MNF transformation bands allowed us to obtain up to 0.17 higher F1 scores for lupine compared to that using 430 HySpex spectral bands (
Table 3). The accuracies obtained on 430 spectral bands using the SVM algorithm were higher (median F1 score from 0.72 to 0.81) than with the RF algorithm (median F1 score from 0.62 to 0.75,
Table 3,
Figure A2 in
Appendix A).
In both study areas, the highest lupine identification accuracy was obtained during the second campaign (F1 score: 0.85 for RJ1, F1 score: 0.83 for KA1). Considering the scenarios listed in the table above, a median F1 score above 0.8 was most often obtained for this campaign (C2). At the beginning of August (C2), lupine was at its peak of growth; it bloomed and filled large, compact patches, which made it easier to identify. Statistically significant lower accuracies were obtained in the spring (F1 score: 0.81) and autumn (F1 score: 0.80) campaigns (
Table 3,
Figure A3 in
Appendix A). During the first campaign in May (C1), lupine was identified before the flowering period and some specimens were still small and visually similar to co-occurring plants. In the third campaign (September C3), lupine inflorescences were withered, and some of the plant patches were attacked by a fungal disease of lupine, i.e., powdery mildew (
Erysiphe pisi). The use of meadows and pastures, especially mowing and grazing of animals, also made identification difficult. For campaign C1, the highest median F1 score (SVM: 0.79, RF: 0.81) for lupine was obtained for scenarios using 30–50 MNF bands (
Figure 4,
Table 3) for both research areas. For campaign C2, the highest F1 scores (SVM: 0.85, RF: 0.84) were obtained for 20–40 MNF bands in KA1 and 30 MNF bands in RJ1. Campaign C3 showed similar tendencies, with scenarios using 30 and 40 MNF bands achieving the highest F1 scores (SVM: 0.79, RF: 0.80). The results show that, regardless of the time of data acquisition (campaign) or study area, the datasets containing 30 and 40 MNF bands performed best. Space used on a hard disk (HDD) for 30 MNF bands was more than 5 times lower than the full set of original hyperspectral bands. Moreover, in the best-case scenarios, both the RF and SVM algorithms obtained comparable results (a maximum difference of two percentage points).
Based on the above results, 30 MNF bands from the second campaign (C2) were used to prepare the final lupine maps for both research areas. Pixels classified by RF and SVM classifiers as lupine a minimum of 95 times (out of 100 iterations) are marked in red on the maps below (
Figure 6 and
Figure 7).
Maps obtained using thresholding frequency images in the KA1 area achieved similar map accuracies for both classification methods (OA: 89%, Cohen’s kappa: 0.73, F1 score for lupine: 0.80,
Table 4). Lupine invasions were mainly located in the meadows and pastures in the northwest and southeast regions of the area.
Both post-classification images had an omission error of approximately 29%, especially in places where the lupine density was lower (over 20% of the reference polygon for lupines covered with co-occurring species). Some patches of lupine were poorly regrown after mowing or grazing by farm animals. A possible reason for the underestimation was also the use of a high threshold, which resulted in the rejection of less reliable pixels classified as lupins from the final map.
The overestimation error on both maps was approximately 9% and occurred where visually similar plants, such as butterbur (Petasites hybribus), nettle, meadow thistle (Cirsium rivulare), and bulbous oat grass, grew.
In the Rudawy Janowickie area, the SVM method yielded a higher identification accuracy (OA: 94%, Cohen’s kappa: 0.82, F1 score for lupine: 0.86) than the RF algorithm (OA: 93%, Cohen’s kappa: 0.78, F1 score for lupine: 0.83,
Table 5). The overestimation error for the lupine class was lower in the support vector machine image (8%) than in the RF image (12%), and the lupines were mostly mixed with tall grasses, that is, bulbous oat grass.
4. Discussion
In this paper, we have presented the original processing chain for invasive species mapping using machine learning algorithms and hyperspectral data. The presented method of multiple classification and thresholding of frequency images allows the results of many inferred images to be included in the final map. The final maps show only those pixels that were classified as lupine in 95 out of 100 classification iterations, based on a random selection of training and validation patterns from field polygons. This increases the reliability of the results and reduces the “salt and pepper” effect. Despite the similarity of
Lupinus polyphyllus to co-occurring plants (similar leaf color and physiological characteristics to native plants, occurring in heterogeneous, small patches), it was possible to identify this species using the SVM and RF algorithms with satisfactory accuracy (F1 score from 0.8 to 0.86) in two research areas. The high potential and repeatability of the method were also confirmed in a different location in Poland (Malinowice village), where wood small-reed, blackberry, and goldenrod were identified with high accuracies (F1 score above 0.9, [
57,
58]). However, these species were more distinguishable from the surrounding plants than lupines because of their characteristic inflorescences and their occurrence in large and dense patches.
It can be concluded that the
Lupinus polyphyllus identification results using both machine learning algorithms (RF and SVM) were similar (F1 score for lupine: 0.8 and OA: 89% in the Kamienne Mountains area and F1 score
RF: 0.83, OA
RF: 93%, F1 score
SVM: 0.86, and OA
SVM: 94% in the Rudawy Janowickie area). The methods used by other researchers to identify lupines have yielded similar accuracies to those presented in this article (
Table 6); however, they refer to a different spatial scale [
32,
72,
73]. The identification of lupine in the UNESCO biosphere reserve “Rhön” in Germany using UAV RGB, thermal imagery, and an object-based image analysis (OBIA) with the RF algorithm gave a similar mean overall accuracy of approximately 89%, but some models highly overestimated the results (false positive rate up to 47%) [
72]. A further comparison of the results is difficult because of the lack of reported class-accuracy metrics for individual classes in the above-mentioned studies. The object classification method worked well for data with very high spatial resolution (0.5 m), but the research was limited to a small area (1.5 km
2) due to the capabilities of the drone (DJI-Phantom IV quadcopter). Lupine identification has also been carried out in the same reserve using WorldView-3 satellite data and the gradient-boosting machine method [
32], but a lower classification accuracy was obtained (F1 score for lupine: 0.76) than those presented in the present article. Panchromatic and multispectral data from this satellite enabled a prediction map to be obtained for a larger “Leitgraben” area, but the authors noted that only large patches of lupine (area at least 3 × 3 m
2) were detected. Similar conclusions were drawn when identifying another lupine species,
Lupinus nootkatensis, in Iceland using SPOT 5 images [
73]. The authors used a maximum likelihood classifier to achieve high accuracy (OA: 94% and Kappa: 0.88); however, they observed that sparse and freestanding lupine areas and patches with a combination of other vegetation were not detected. However, these small, solitary patches of lupine are indicated as the main factor causing the spread of this species into new areas. Low accuracies for white and yellow lupine (F1 score below 0.04) were also obtained when mapping annual crops in Portugal using Sentinel-2 data and the random forest method [
74]. This confirms that, in the case of lupine identification, high spatial and spectral resolutions of the images are important, especially if the beginning of the invasion is to be detected.
In this paper, the impact of the number of MNF bands used on the accuracy achieved was also analyzed. The accuracy of lupine classification increased with the number of MNF bands in the input set but stabilized for sets consisting of more than 20 transformed bands. Datasets containing about the first 30 MNF bands gave the highest classification accuracies (median F1 score for lupine from 0.77 to 0.85), while accuracies obtained for the 430 HySpex spectral bands were lower (median F1 score from 0.62 to 0.81). Improvements in species identification accuracy after using dimensionality reduction methods have also been noted in other studies [
28,
30]. The use of 30 MNF bands to classify small-reed wood, goldenrod, and blackberry in southern Poland resulted in higher average F1 score accuracies for the three species (F1: 0.86–0.91) compared to the results with 430 HySpex hyperspectral bands (F1: 0.93–0.95, [
57]). The use of 20 MNF bands resulted in higher classification accuracies for seven tree species for RF (OA: 87%) and a multi-class classifier (MCC, OA: 89%), compared to using 118 HyMap bands (OA: 46% for RF and OA: 79% for MCC) [
29]. Additionally, testing different raster data (MNF bands from HySpex data, lidar products, and vegetation indices) for the identification of steeplebush (
Spiraea tomentosa) in the Lower Silesian forests gave the highest RF classification accuracies for 25 MNF bands (OA: 99%, F1 score for steeplebush: 0.83) and the obtained accuracies were similar to those for lupine in this article [
19].
The study showed that the best period for mapping lupines was the second campaign (August). The F1 score for lupine obtained in the summer campaign (SVM up to 0.85; RF up to 0.84) was higher than that in the other campaigns. Other authors have also shown that the beginning of September is not the best time to identify lupines in central Europe (Germany) and recommended collecting data during the peak flowering period of
Lupinus polyphyllus [
32]. The campaign during the flowering period of the identified plants (August) was also optimal for classifying small-reed wood (F1: 0.90) and blackberry (0.98), whereas goldenrod was well identified in every campaign (F1 from 0.96 to 0.99) [
58]. Similar conclusions were reached during the classification of
Echinocystis lobata in the Bzura River valley in central Poland using the RF algorithm and HySpex images [
9]. The F1 score for the species was the highest in summer (0.87) and was lower in spring (F1: 0.64) and autumn (F1: 0.75). The blooming time was also the best period for identifying
Molinia caerulea; in August, the F1 score for this species ranged from 0.86 to 0.89, while the accuracy was lower in June (F1 score from 0.78 to 0.87) and September (F1 score from 0.84 to 0.88) [
20]. Significant differences in the accuracy obtained depending on the time of data acquisition were also observed during
Carduus nutans identification using the SVM algorithm and AISA data, when OA = 91% was obtained for the peak flowering period for musk thistle (June) and OA = 79% was obtained before flowering (April) [
76]. The variability in the spectral characteristics of co-occurring plants is also worthy of attention. In the case of steeplebush identification, a higher F1 for the species was obtained for the September campaign (F1: 82.96%) compared to the August campaign (F1: 77.25%) [
19]. The woody parts of the steeplebush were more visible in autumn, owing to changes in the spectral characteristics of the surrounding plants and lower biomass.