Mapping invasive alien plant species with very high spatial resolution and multi-date satellite imagery using object-based and machine learning techniques: A comparative study

ABSTRACT

Invasive alien plant species (IAPS) have negative impacts on ecosystems, including the loss of biodiversity and the alteration of ecosystem functions. The strategy for mitigating these impacts requires knowledge of these species’ spatial distribution and level of infestation. In situ inventories or aerial photo interpretation can be used to collect these data but they are labor-intensive, time-consuming, and incomplete, especially when dealing with large or inaccessible areas. Remote sensing may be an effective method of mapping IAPS for a better management strategy. Several studies using remote sensing to map IAPS have focused on single species detection and were conducted in relatively homogeneous natural environments, while other common, more heterogeneous environments, such as urban areas, are often invaded by multiple IAPS, posing management challenges. The main objective of this study was to develop a mapping method for three major IAPS observed in the urban agglomeration of Quebec City (Canada), namely Japanese knotweed (Fallopia japonica); giant hogweed (Heracleum mantegazzianum); and phragmites (Phragmites australis). Mono-date and multi-date classification approaches were used with WorldView-3 and SPOT-7 satellite imagery, acquired in the summer of 2020 and in the autumn of 2019, respectively. To estimate presence probability, object-based image analysis (OBIA) and nonparametric classifiers such as Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) were used. Overall, multi-date classification using WorldView-3 and SPOT-7 images produced the best results, with a Kappa coefficient of 0.85 and an overall accuracy of 91% using RF. For XGBoost, the Kappa coefficient was 0.81 with an overall accuracy of 89%, whereas the Kappa coefficient and overall accuracy were 0.80 and 88% for SVM classifier, respectively. Individual class performances based on F1-score revealed that Japanese knotweed had the highest maximum value (0.95), followed by giant hogweed (0.91), and phragmites (0.87). These results confirmed the potential of remote sensing to accurately map and simultaneously monitor the main IAPS in a heterogeneous urban environment using a multi-date approach. Although the approach is limited by image and reference data availability, it provides new tools to managers for IAPS invasion control.

KEYWORDS:

• Invasive alien plant species
• remote sensing
• multi-date
• satellite imagery
• OBIA
• machine learning

1. Introduction

Invasive alien plant species (IAPS) are currently considered to be one of the main drivers of biodiversity loss worldwide (Langmaier and Lapin 2020; Paz-Kagan et al. 2019; Guido, Pillar, and Souza 2017; Early et al. 2016; IUCN 2000). The main negative impacts of IAPS are native species replacement, alteration of chemical soil properties, and modification of the fire regimes (Kumar Rai and Singh 2020). IAPS are most often more competitive for resources than native species (Martínez-Izquierdo et al. 2016), causing a reduction or extinction of native vegetation, a reduction of primary productivity, and an alteration of ecosystem functions (Davies and Johnson, 2011). For example, the replacement of native species by Japanese knotweed (Fallopia japonica) can alter the quality and stability of the soil and thus accentuate flooding and erosion (Lavoie, Guay, and Joerin 2014; Kumar Rai and Singh 2020). In addition to environmental impacts, management and eradication costs, loss of invaded land value, and health problems caused by certain IAPS also represent important invasion consequences (Kelsch et al. 2020; Roy et al. 2019; Lavoie, Guay, and Joerin 2014). These impacts could be limited by control strategies designed for the prevention of new invasions, the detection of early stage invasions, rapid response, and management of established or spreading IAPS (IUCN 2000). The management of established IAPS consists of eradication, containment, and control, requiring important financial resources that can be reduced by early detection. In order to succeed, the strategy needs the most accurate information possible regarding the spatial distribution and levels of infestation of IAPS.

In the past, datasets of the territorial distribution of IAPS were mainly acquired from field surveys, in situ inventories collected by GNSS (Global Navigation Satellite Systems) or via the interpretation of aerial photos (Jombo, Adam, and Odindi 2021; Royimani et al. 2019; Hartling et al. 2019; Lawrence, Wood, and Sheley 2006; Müllerová et al. 2005). However, these methods are labor-intensive and often not financially and technically suitable, especially when dealing with large or inaccessible areas. They can also be relatively subjective because they are observer-dependent (Royimani et al. 2019; Lawrence, Wood, and Sheley 2006). The use of remote sensing may be a more favorable alternative for monitoring and managing IAPS (Dash et al. 2019; Niphadkar and Nagendra 2016) despite limiting factors related to the availability and resolution of images (Robinson et al. 2016). Easily replicable approaches make it possible to map spatial distribution over large areas much more quickly than in situ inventories, enabling early detection, which is one of the success factors of IAPS eradication before large areas are invaded (Dash et al. 2019; Early et al. 2016; Ustin et al. 2002). Multispectral images from very high to medium spatial resolution are those most frequently used in the visible and near infrared electromagnetic domains (Royimani et al. 2019; Vaz et al. 2018). A review of the characteristics of widely used data for remote sensing of IAPS is presented in Appendix 1.

Automatic classification methods have been developed in recent decades to overcome the main limitations of field inventory-based methods (Royimani et al. 2019; Lu and Weng 2007; Lass et al. 2005). In particular, object-based image analysis (OBIA) and nonparametric approaches (e.g. machine learning techniques) (Royimani et al. 2019; Asner et al. 2008; Lass et al. 2005) have been widely used.

OBIA consists of first grouping adjacent pixels, fulfilling certain criteria of similarity between them to form objects (i.e. homogeneous groups of neighboring pixels) which will be considered as the basic unit of classification, and then to assign them a class of interest (Chen, Zhao, and Powers 2014). OBIA has the advantage of using several types of spectral and spatial features (e.g. geometry, texture, topology) in the classification process (Hantson et al. 2012). Since the processing units are objects, the approach avoids the problems frequently observed in pixel-based classification approaches, such as salt-and-pepper noise (Hirayama et al. 2019; Chen, Zhao, and Powers 2014; Jones et al. 2011).

Nonparametric machine learning techniques do not require a normal distribution of training samples and are therefore able to combine multiple data sources often characterized by distinct statistical distributions (Benediktsson and Sveinsson 1997). These methods are also suitable for small samples, especially in heterogeneous environments (e.g. urban areas) where it may be difficult to find enough samples for certain classes (Masse 2013).

Random Forest (RF), support vector machine (SVM), and artificial neural networks (Qian et al. 2020; Shiferaw et al. 2019; Paz-Kagan et al. 2019; Dash et al. 2019; Kganyago et al. 2018; Dorigo et al. 2012) are the most frequently used machine learning techniques for IAPS classification. Although these techniques require powerful tools related to optimization, training, and classification time as well as a high number of input feature requirements (Qian et al. 2020; Royimani et al. 2019; Lantz and Wang 2013; Kavzoglu and Mather 2004), a number of studies have highlighted their strong performance. As an example, Martin et al. (2018) achieved an overall accuracy of more than 90% using RF to map Japanese knotweed (Fallopia japonica) in the Anse and Serrières area (France). An overall accuracy of 95% was also obtained by Dorigo et al. (2012) using RF to map Japanese knotweed in Ljubljana, Slovenia. Artificial neural networks and SVM classification were used by Abeysinghe et al. (2019) for the detection of phragmites (Phragmites australis) in Ohio, USA, achieving 94% and 90% overall accuracy, respectively.

These machine learning techniques also have the potential to incorporate different types of arithmetic discriminative features to distinguish IAPS from native species. Their spectral properties in the visible, near-infrared, or mid-infrared range, their phenology, and their morphology are the most commonly used attributes (Ewald et al. 2020; Niphadkar and Nagendra 2016; Dorigo et al. 2012). Most IAPS have higher photosynthesis rates and concentrations of photosynthetic pigments compared to native species (Kalkman, Simonton, and Dornbos 2019; Sánchez-Azofeifa et al. 2009; Asner et al. 2008; Castro-Esau, Sánchez-Azofeifa, and Caelli 2004), enabling significant separability from native species. Castro-Esau, Sánchez-Azofeifa, and Caelli (2004) found a decrease of classification error from 16 to 4% for two invasive and other native species following the addition of chlorophyll content. Phenology is also used because some of the vegetative stages (e.g. flowering) of certain IAPS are not synchronous with those of native species (Labonté et al. 2020; Xu, Griffin, and Schuster 2007). This characteristic can be exploited by remote sensing using bi-temporal images or indices and has been successfully used (overall accuracy = 95%) to map Japanese knotweed, drawing upon aerial images acquired in winter and summer in Ljubljana (Slovenia) (Dorigo et al. 2012). Morphology is measurable using texture and shape indices of objects in an image (Kattenborn et al. 2019; Haralick 1979), making it possible to distinguish IAPS from native species when characterized by a particular shape or coloration (Müllerová et al. 2017; Michez et al. 2016; Jones et al. 2011). Texture indices were exploited in a mapping study of giant hogweed (Heracleum mantegazzianum) with an overall accuracy of 97%, and relevant discriminant texture indices appeared to be related to the white coloration of umbels, enabling the extraction of discriminating homogeneity indices (Michez et al. 2016).

Relatively few studies have focused on a simultaneous mapping of multiple IAPS in the same area, and most have been carried out in relatively homogeneous natural environments over small areas (Michez et al. 2016; Ustin et al. 2002). Heterogeneous environments such as urban areas are often invaded by multiple IAPS, leading to management and detection challenges using remotely sensed data. The most important challenges to mapping IAPS in urban areas concern the coexistence of several species (ornamental and native species) spectrally similar to IAPS. This similarity makes it difficult to conduct multi-IAPS mapping and to identify individual IAPS. The few studies conducted in urban environments have focused on single IAPS and have not exploited multi-date approaches that combine phenology and morphological-based features to carry out simultaneous multi-IAPS classification (Kazmi et al. 2021; Martin et al. 2018; Singh et al. 2018). The main objective of this study was therefore to perform a multi-species classification of IAPS in a complex urban area, using machine learning techniques and OBIA applied to multi-date images.

2. Materials and methods

2.1. Study area

The study area is the urban agglomeration of Quebec City (Quebec, Canada) and covers a total surface area of 557 km² (Figure 1). The natural environment consists of forest, urban woodlands, wetlands, and water bodies (Ville de Québec 2006). The forest, which occupies mainly the northern and western zones, represents about 35% of the total area, of which 44% is deciduous and 42% is mixed (deciduous and coniferous). Several areas of the urban agglomeration of Quebec City are invaded by IAPS (Lavoie, Guay, and Joerin 2014).

Figure 1. Study area limits and reference data locations of invasive alien plant species (giant hogweed, Japanese knotweed, and phragmites).

2.2. Studied species

The species considered in this study are terrestrial IAPS, namely Japanese knotweed, phragmites, and giant hogweed. Originally from East Asia and introduced to North America for ornamental purposes, Japanese knotweed is ranked among the world’s 100 worst invasive species (Lavoie, Guay, and Joerin 2014; Dorigo et al. 2012). The species forms dense and homogeneous colonies mainly along roads, railways, riparian areas, and other urban environments and can grow up to 3 m tall (Aguilera et al. 2010) (Figure 2a). This species causes a loss of biodiversity due to its proliferation and a modification of the hydrographic network as a result of the resistance that its root system imposes to runoff (Dorigo et al. 2012; Collingham et al. 2000). Its highly developed root system represents one-third of its biomass and explains the very high eradication costs, which requires large-scale excavation (Lavoie, Guay, and Joerin 2014).

Figure 2. Field photography of invasive alien plant species at two zoom levels: (a, b) Japanese knotweed; (c, d) Giant hogweed; and (e, f) phragmites.

Giant hogweed is a species native to the Caucasus region of Central Asia (Page et al. 2006). It can exceed a height of 3 m and mainly colonizes stream edges, agricultural ditches, and forest edges (Page et al. 2006) (Figure 2c). This species has a negative impact on biodiversity, accelerates water erosion, and causes human health threats due to inflammatory toxic secretions (Lavoie, Guay, and Joerin 2014; Page et al. 2006).

Originating in Eurasia, phragmites comprise a genus that develops dense colonies over 2 m tall primarily in marshes, along roads, and in agricultural drainage ditches (Abeysinghe et al. 2019; Lavoie, Guay, and Joerin 2014) (Figure 2e). Stems can sometimes damage road construction, reduce the presence of native species, and alter habitat for aquatic species such as fish (Abeysinghe et al. 2019; Lavoie, Guay, and Joerin 2014; Leonard, Wren, and Beavers 2002).

2.3. Field reference data collection

Field campaigns were conducted between June 26 and 26 August 2020. We collected points and polygons for previously invaded sites which were systematically visited. New sites were selected opportunistically during the visit of these designated sites and were also characterized. Points and polygons of IAPS and absences were collected using Xplore Bobcat (Xplore Technologies, Texas) and Blackview (Blackview, Shenzhen) electronic tablets connected to high accuracy (< one meter) GNSS receivers (Pro 6 H GNSS (Trimble, California) and Arrow Gold (EOS positioning system, Quebec). Polygons were digitized in situ based on RGB orthomosaic (red-green-blue, 15 cm spatial resolution, acquired in summer 2018) aerial imagery. Giant hogweed point data collected in 2019 by managers of the Cap Rouge River watershed council were also added to the database after validation of their presence in 2020 using photo interpretation. A total of 283, 128, 76, and 32 points and polygons were used for absences, Japanese knotweed, phragmites, and giant hogweed, respectively (Appendix 2).

2.4. Image acquisition

This study used WorldView-3 (WV-3) satellite images (WorldView 110 (WV110) camera) and Satellite Pour l’Observation de la Terre-7 (SPOT-7) (New AstroSat Optical Modular Instrument (NAOMI) camera) acquired in summer (vegetation growing period) and autumn (senescence period), respectively, in order to explore the contribution of these phases to the classification. This image selection is based on the availability of archived multispectral satellite images at high and very high resolutions, over the study area, during our targeted periods. Due to the limited availability of suitable images, we had to select them from two different sensors (WV-3 and SPOT-7). The characteristics of the images are presented in Table 1.

Table 1. Characteristics of satellite images used in this study.

Image	Band	Wavelength range (center) (nm)	Spatial resolution (m)	Acquisition date
WV-3	Coastal	400–450 (425)	1.24	2020/07/06
	Blue	450–510 (480)
	Green	510–580 (545)
	Yellow	585–625 (605)
	Red	630–690 (660)
	Red Edge	705–745 (725)
	Near-infrared 1	770–895 (832.5)
	Near-infrared 2	860–1040 (950)
	Panchromatic	450–800 (625)	0.31
SPOT-7	Blue	450–520 (485)	6	2019/10/26
	Green	530–590 (560)
	Red	625–695 (660)
	Near-infrared	760–890 (825)
	Panchromatic	450–745 (597.5)	1.5

WV-3 : WorldView-3, SPOT-7 : Satellite Pour l’Observation de la Terre-7.

2.5. Image pre-processing and classification

Two main steps were performed for IAPS mapping: (1) pre-processing, including atmospheric correction, fusion of panchromatic and multispectral bands, orthorectification, and masking to remove irrelevant areas; and (2) classification of the segmented objects by machine learning techniques. The flowchart of the study is presented in Figure 3.

Figure 3. Methodological flowchart of the study.

2.5.1. Image pre-processing

Atmospheric corrections were performed using the ATCOR module (PCI Geomatics, 2018). Spatial resolution enhancement (i.e. pansharpening) of the multispectral bands by fusion with the panchromatic band was also applied. The Zhang (Zhang 2004); Gram-Schmidt (GS) (Belfiore et al. 2016); ratio component substitution (RCS); local mean and variance matching (LMVM); and Bayesian fusion (Grizonnet et al. 2017) methods were compared. The method that minimizes the relative dimensionless global error (ERGAS) index was selected. This index is calculated from the root mean square of error (RMSE) according to equations (1) and (2) and evaluates the distortion of the spectral information in the pansharpened images (Park et al. 2020; Li, Jing, and Tang 2017; Belfiore et al. 2016; Ranchin et al. 2003); values close to zero are more relevant to minimizing alteration of spectral information of pansharpened images (Mhangara, Mapurisa, and Mudau 2020; Li, Jing, and Tang 2017; Ranchin et al. 2003),

(1) $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{m\times n}{\sum }_{i=1}^m{\sum }_{j=1}^n{\left[\mathrm{M}\left(\mathrm{i},\mathrm{j}\right)-\mathrm{F}\left(\mathrm{i},\mathrm{j}\right)\right]}^2}$(1)

(2) $\text{ERGAS}\text{=}\frac{100}{\text{R}}\frac{1}{\text{N}}\sum \text{k}=1\text{N}\frac{{\text{RMSE}}_{\text{k}}{ }^2}{{\text{M}}_{\text{K}}{ }^2}$(2)

where M (i, j) is the spectral value of the original band pixel at position (i, j); F (i, j) is the spectral value of the pansharpened band pixel at position (i, j); R is the resolution ratio between the multispectral bands and the panchromatic band; N is the number of bands; k is the band number; and M_k is the mean of the original band k. To optimize the processing time, four of the fourteen WV-3 tiles representative of the different environments and occupying half of the study area were selected for this analysis. Spectral bands whose wavelength does not correspond to that of the panchromatic band (i.e., Coastal and Near-infrared 2) were not used for this analysis and were resampled to 0.31 m using the bilinear interpolation method. This method was selected because it is the most adapted to continuous data and it limits data alteration (Sa 2014; Patel and Mistree 2013). The fusion methods were each compared to the WV-3 images and the optimal method thereby identified was used for the SPOT-7 image fusion. The LMVM and RCS methods showed the best results compared to the other methods (Table 2).

Table 2. ERGAS (Relative Dimensionless Global Error) values for the fusion methods applied to the WorldView-3 image.

	Zhang	Bayesian fusion	RCS	LMVM	GS
Tile 1	2.80	2.31	2.03	2.03	2.87
Tile 2	4.24	3.72	3.40	3.08	4.31
Tile 3	6.17	7.43	4.80	4.34	5.40
Tile 4	3.98	3.13	2.95	2.59	3.53
Mean ± SD	4.30 ± 0.94	4.15 ± 1.64	3.30 ± 0.81	3.01 ± 0.70	4.03 ± 0.83

RCS: Ratio Component Substitution; LMVM: Local Mean and Variance Matching.

GS: Gram-Schmidt; SD: Standard Deviation.

While the LMVM method performed slightly better according to the ERGAS, the RCS method was selected because it presents a better compromise between minimizing spectral distortion and improving the geometric structure of the objects (see Appendix 3).

Orthorectification of the pansharpened images was performed using a digital terrain model (Varin, Allostry, and Chalghaf 2020), and a second-order polynomial function was performed before tile mosaicking using the Bundle color balancing method (PCI Geomatics 2018). Control points were identified from an orthomosaic of aerial images acquired in summer 2018 at 15 cm spatial resolution. The images were orthorectified with an RMSE of 0.63 pixels for WV-3 and 0.50 pixels for SPOT-7, and the respective spatial resolutions after fusion were 0.31 m and 1.5 m.

2.5.2. Masking

Three masks were applied to exclude areas not suitable for the selected IAPS. A normalized difference vegetation index (NDVI) threshold of 0.37 was used to eliminate non-vegetated areas (i.e. bare soil, buildings, and water bodies) using the minimum average of reference objects corresponding to the four classes. A second mask, based on a digital height model (Varin, Allostry, and Chalghaf 2020, was applied to exclude vegetation higher than 4 m based on field observations and a literature review showing a height lower than 4 m for these IAPS (Abeysinghe et al. 2019; Lavoie, Guay, and Joerin 2014; Barney et al. 2006). This threshold is based on field observations and the literature (Abeysinghe et al. 2019; Lavoie, Guay, and Joerin 2014; Barney et al. 2006). Finally, a shadow index (SI) was used (Zhou et al. 2018) to eliminate shaded areas (SI ≤ 13.13). This threshold was selected based on a sample of 60 objects representing shadows on the WV-3 orthomosaic (i.e. the maximum average identified for all 60 objects).

2.5.3. Multi-resolution segmentation

WV-3 orthomosaic segmentation was used for object generation because it was acquired at higher spatial resolution than the SPOT-7 image. To eliminate redundancy of spectral bands (Bilgin, Erturk, and Yildirim 2011; Kermad and Chehdi 2000) and to optimize processing time, the Red-edge, Near-infrared 1, green, and red bands were selected for segmentation after analysis of spectral separability using the Jeffries-Matusita distance (JM) between classes (Oktorini et al. 2021; Massetti et al. 2016).

The Estimation of Scale Parameter method (ESP2) (Drăguţ et al. 2014; Drǎguţ, Tiede, and Levick 2010) was used to assist in identifying the optimal scale. Twelve tiles of equal size (20000 × 20000 pixels) with identifiable IAPS were selected to analyze the local variance at each object level (Drǎguţ, Tiede, and Levick 2010). Different combinations of scale, color, and shape (compactness) were tested until the optimal values of 25 for scale, 0.9 for color, and 0.5 for shape were reached, as well as a weight equal to 1 for each of the four bands used.

2.5.4. Extraction and pre-processing of discriminant features

Three categories of arithmetic features, as well as groupings of spectral, textural, and geometric features for classification modeling, were calculated using eCognition Developer 10.1 software (Trimble Geospatial, USA). The categories and formulas for the features computed for each of the images (152 for WV-3, 78 for SPOT-7) are presented in Table 3. These features were centered and scaled over all the study area dataset, and a correlation analysis was performed to eliminate redundancy. If two features were correlated with a Pearson coefficient greater than or equal to 0.85 (Varin, Chalghaf, and Joanisse 2020), the feature achieving better separability according to the JM separability distance was used.

Table 3. Extracted features (spectral, textural, and geometric) for classification modeling.

Category	Feature	Definition	Reference
Spectral	Mean	All
	Maximum	All
	Standard deviation	All
	95^th percentile	All
	Skewness	All
	NDVI (Normalized Difference Vegetation Index)	(NIR1 - red)/ (NIR1 + Red)	Carlson and Ripley (1997)
	RENDVI (Red-Edge Normalized Difference Vegetation Index)	(NIR1 - RE)/ (NIR1 + RE)	Waser et al. (2014)
	VARI (Visible Atmospherically Resistant Index)	(green - red) / (green + red -blue)	Gitelson et al. (2002)
	RGR (Red Green Ratio)	red/green	Zhou et al. (2021)
	SR (Simple Ratio)	NIR1 /red	Zhou et al. (2021)
	BTBR (Bi-Temporal Band Ratio)	(red, green)*	Dorigo et al. (2012)
	HSI (Hue, Saturation, Intensity)	red, green, blue*	Chaudhary et al. (2012)
	CI (Colour Index)	(red - green) / (red + green)	Escadafal and Huete (1991)
	BI (Brightness Index)	(red^2 + NIR1^2)^0.5	Zhou et al. (2018)
	NDWI (Normalized Difference Water Index)	(green - NIR1) / (green + NIR1)	Zhou et al. (2018)
	NIR-RE ratio	(NIR1–3.06 × RE)/(1.71× NIR1–2.42 × RE − 4.14)	Developed by the research team
Textural	GLCM (Gray Level Co-occurrence Matrix): contrast, correlation, entropy, homogeneity, Angular second moment, mean, standard-deviation) GLDV (Gray Level Difference Vector): contrast, entropy, mean, angular second moment	All	Haralick (1979)
Geometric	Width, length, compactness	-

NIR: Near-infrared, RE: Red-Edge.

*Because of limited space, only the bands used are indicated (see references for detailed definition).

2.5.5. Classification approach

Five classifications were compared, with two mono-date classifications using only features calculated from each of the two images (WV-3 and SPOT-7) performed first. Two additional classifications were then performed by adding the multi-date BTBR index developed by Dorigo et al. (2012) to each of these images (WV-3 + BTBR, SPOT-7 + BTBR). This index (Equation (3)(3) $\mathrm{B}\mathrm{T}\mathrm{B}\mathrm{R}=\frac{\frac{R_{off}}{R_{on}}-\frac{G_{off}}{G_{on}}}{\frac{R_{off}}{R_{on}}+\frac{G_{off}}{G_{on}}}$(3) ) combines the spectral properties of two phenological periods using the reflectance in the red and green bands,

(3) $\mathrm{B}\mathrm{T}\mathrm{B}\mathrm{R}=\frac{\frac{R_{off}}{R_{on}}-\frac{G_{off}}{G_{on}}}{\frac{R_{off}}{R_{on}}+\frac{G_{off}}{G_{on}}}$(3)

where R and G indicate the reflectance values in the red and green bands, respectively. The suffixes indicate the image acquisition periods (the growing season (on) and the senescence period (off)). Finally, a classification combining features from both images including BTBR (WV-3 + SPOT-7 + BTBR) was performed.

2.5.6. Machine learning techniques and optimization

Three machine learning techniques, namely RF (Breiman 2001); SVM (Guenther and Schonlau 2016); and XGBoost (Zarei, Hasanlou, and Mahdianpari 2021; Samat et al. 2020) classifiers, were tested and compared. The selection of important discriminant features for each classifier was performed using the recursive feature elimination (RFE) selection method (Yang et al. 2019; Guyon et al. 2002; Ambroise and McLachlan 2002). For each classifier, the first iteration of RFE consists of a classification using all features and the computation of mean accuracy of ten-fold cross-validations as accuracy assessment. The successive iterations consist of eliminating the least ranked feature followed by new accuracy assessment and feature ranking. The final step consists of selecting a subset of features that reached the highest accuracy for classification. Subsequently, the caret package in R (R core team 2021; Kuhn 2008) was used to optimize the other parameters of the three classifiers. The value of mtry (i.e. the number of random features used at each node) for RF was the one that minimized the out of bag (OOB) error. The number of decision trees (ntree) was the value at which the minimal OOB error was considered constant. For the SVM classifier, radial basis Kernel function (RBF) was used because of its performance in previous studies (Jombo et al. 2020; Qian et al. 2014; Huang, Davis and Townshend 2002). The cost and sigma parameters, as well as the XGBoost parameters, such as the learning rate, were automatically optimized by performing all possible combinations between 10 values of each parameter (i.e. tune length parameter was set to 10). The used values for each parameter were presented in Appendix 4. A classifier that optimizes the user and producer accuracies (i.e. for high F1-score averages) was used for the prediction of object membership in classes. The class with the highest prediction probability was assigned to the object.

2.5.7. Accuracy assessment

The reference data were overlaid with the objects from the segmentations, and for each class, the corresponding segments were divided between training segments (70%) and validation segments (30%), chosen randomly. Cohen’s Kappa coefficient (Congalton 1991; Cohen 1960) and overall accuracy were calculated as performance indicators for models’ assessment. User accuracy, producer accuracy and F1-score were also calculated for individual class performance analysis (Costa et al. 2021; Yang 2001).

2.5.8. Post-classification analysis

Several criteria were used to exclude certain areas that were not suitable for the selected IAPS. Giant hogweed and phragmites do not colonize areas near residences (Lavoie, Guay, and Joerin 2014; Page et al. 2006; Barney et al. 2006), and so prediction of these species was not performed within 10 m of residential buildings. Agricultural areas were also excluded, except parcel edges (10 m or less from edges) that may be invaded by these species.

3. Results

3.1. Mono-date classification

Classification using just the WV-3 orthomosaic showed that the XGBoost classifier performed better (Kappa = 0.81) than the RF (Kappa = 0.77) and SVM classifiers (Kappa = 0.74) (Table 4). The performances obtained using the WV-3 orthomosaic are much better than those obtained from the SPOT-7 image classification (Kappa = 0.48 (SVM), 0.44 (RF) and 0.41 (XGBoost)).

Table 4. Performance measures of classifiers. The values in bold correspond to the maximum performance values.

		WV-3			WV-3 + BTBR			SPOT-7			SPOT-7 + BTBR			WV-3 + SPOT-7 + BTBR
		RF	SVM	XGBoost	RF	SVM	XGBoost	RF	SVM	XGBoost	RF	SVM	XGBoost	RF	SVM	XGBoost
	Kappa	0.77	0.75	0.81	0.85	0.80	0.81	0.44	0.48	0.41	0.64	0.65	0.49	0.85	0.80	0.81
	OA	0.87	0.85	0.89	0.91	0.88	0.89	0.70	0.72	0.66	0.80	0.80	0.70	0.91	0.88	0.89
Japanese knotweed	UA	0.90	0.87	0.90	0.88	0.83	0.90	0.67	0.74	0.47	0.77	0.81	0.58	0.88	0.85	0.85
	PA	0.97	0.94	0.97	0.97	0.97	1.00	0.50	0.47	0.58	0.67	0.72	0.64	0.97	0.92	0.94
	F1-score	0.93	0.91	0.93	0.92	0.90	0.95	0.57	0.58	0.52	0.72	0.76	0.61	0.92	0.88	0.89
Phragmites	UA	1.00	0.75	0.92	1.00	0.84	0.88	0.77	0.75	0.80	0.93	0.88	0.72	0.94	0.84	1.00
	PA	0.52	0.57	0.57	0.76	0.76	0.67	0.48	0.57	0.38	0.67	0.71	0.62	0.81	0.76	0.76
	F1-score	0.69	0.65	0.71	0.86	0.80	0.76	0.59	0.65	0.52	0.78	0.79	0.67	0.87	0.80	0.86
Giant hogweed	UA	0.78	0.82	0.90	0.82	0.82	0.80	1.00	0.75	0.33	0.80	0.60	0.50	0.82	0.83	0.80
	PA	0.70	0.90	0.90	0.90	0.90	0.80	0.20	0.30	0.20	0.40	0.30	0.20	0.90	1.00	0.80
	F1-score	0.74	0.86	0.90	0.86	0.86	0.80	0.33	0.43	0.25	0.53	0.40	0.29	0.86	0.91	0.80
Absence	UA	0.85	0.86	0.88	0.92	0.91	0.90	0.69	0.71	0.76	0.78	0.80	0.76	0.93	0.90	0.90
	PA	0.93	0.87	0.93	0.92	0.86	0.91	0.89	0.91	0.81	0.93	0.92	0.80	0.91	0.87	0.91
	F1-score	0.89	0.87	0.90	0.92	0.88	0.90	0.78	0.79	0.78	0.85	0.85	0.78	0.92	0.89	0.90
Mean (F1-score)		0.81	0.82	0.86	0.89	0.86	0.85	0.57	0.61	0.52	0.72	0.70	0.58	0.89	0.87	0.87
SD (F1-score)		0.10	0.09	0.08	0.03	0.03	0.07	0.12	0.11	0.13	0.09	0.15	0.15	0.03	0.03	0.03

OA: Overall Accuracy; PA: Producer Accuracy; UA: User Accuracy; SD: Standard Deviation; RF: Random Forest; SVM: Support Vector Machine.

XGBoost: Extreme Gradient Boosting

Classification performances for individual classes also showed that the WV-3 orthomosaic features succeed in discriminating classes better than SPOT-7, according to the F1-score. Japanese knotweed (maximum F1-score = 0.93, RF) and giant hogweed (maximum F1-score = 0.90, SVM) were better classified using the WV-3 orthomosaic than phragmites (maximum F1-score = 0.71, XGBoost) with the same orthomosaic. For all IAPS, the F1-score values were lower than 0.65 when the SPOT-7 image was used (Table 4).

3.2. Multi-date classification

3.2.1. Mono-date feature combination with BTBR (WV-3-BTBR and SPOT-7-BTBR)

RF was the best performing classifier when the BTBR index was added to the WV-3 orthomosaic (increased Kappa coefficient from 0.77 to 0.85) (Table 4). The Kappa coefficient value also increased, from 0.75 to 0.80 for SVM, while it remains constant for XGBoost (0.81) with the BTBR incorporated. The reduction in confusion between phragmites and absences after incorporation of the BTBR index (maximum F1-score = 0.86 versus 0.71 for WV-3) improved the performance indicators. The Kappa coefficients also increased for the SPOT-7 image, from 0.44 to 0.64 for RF, 0.48 to 0.65 for SVM, and 0.41 to 0.49 for XGBoost.

Individual classification performances between IAPS varied when comparing the addition of BTBR. For WV-3, the F1-score value increased by 0.02 for Japanese knotweed and 0.15 for phragmites, whereas the maximum F1-score value decreased by 0.04 for giant hogweed (Table 4).

This suggests that the use of BTBR provides more benefits for phragmites than for the other two IAPS. The use of BTBR also improved the IAPS classification performance with SPOT-7 according to the F1-score. For RF, for example, the improvement is 0.15, 0.19, and 0.20 for Japanese knotweed, phragmites, and giant hogweed, respectively (Table 4).

3.2.2. Multi-date feature combination with BTBR (WV-3 + SPOT-7 + BTBR)

The performance of the three classifiers after the selected features were combined with BTBR (Appendix 5) is similar to that obtained for WV-3 + BTBR, and better than for SPOT-7 + BTBR. RF remains the optimal classifier for classification of all three IAPS with the highest Kappa coefficient (0.85), compared to XGBoost (0.81), and SVM (0.80) (Table 4). The only difference observed between WV-3 + BTBR and WV-3 + SPOT-7 + BTBR is the F1-score value, which increases very slightly from 0.86 to 0.87 for phragmites. The combination of WV-3 + SPOT-7 + BTBR and the RF classifier was therefore used for the prediction and mapping of the three IAPS in the study area.

3.3. IAPS mapping

Some representative examples of IAPS mapping after membership prediction using the RF classifier (WV-3 + SPOT-7 + BTBR) are shown in Figure 4. The objects identified as phragmites are mainly located along highways, drainage ditches, and around water bodies (Figure 4a,b). Identified giant hogweed plants are usually located along streams and woodland edges and in clearings under the canopy (Figure 4a–c). Large concentrations of Japanese knotweed were generally observed along trails, railroads, and vacant lots (Figure 4d). The particularity of this species is its presence along the fences of houses where it had often been planted for ornamental purposes (Figure 4d).

Figure 4. Examples of invasive alien plant species detected in four types of environment: (a) riverbanks; (b) roadsides; (c) agriculture ditches; and (d) residential areas. Species are indicated in yellow (giant hogweed); orange (phragmites); and purple (Japanese knotweed).

The total area of the three IAPS detected is 455 ha and represents 0.8% of the study area. The area of phragmites (378 ha) is much larger than that of Japanese knotweed (68 ha) and giant hogweed (9 ha). However, this area could be underestimated for phragmites and overestimated for Japanese knotweed and giant hogweed, based on the accuracy assessment measures (Table 4). The omission error (1-PA) for phragmites was 19%, while the commission error (1-UA) was 6% (RF, WV-3 + SPOT-7 + BTBR) (Table 4 and Appendix 6). Some phragmite patches were not detected and a visual post-processing verification using field photos showed that undetected phragmites patches correspond to low densities of phragmites (Figure 5a) and areas where phragmites were not yet in the flowering stage (Figure 5b).

Figure 5. Examples of undetected phragmites stands. (a) Phragmites at low density and (b) a phragmite stand before the flowering stage (field photography from July 30, 2020).

Omission errors (1-PA) for Japanese knotweed and giant hogweed were 3 and 10%, respectively, while errors of commission (1-UA) were 12% and 18%, respectively (Appendix 6). There were more absence objects incorrectly classified as these IAPS than such objects in the phragmites class. The presence of several plant species, including species spectrally similar to Japanese knotweed and giant hogweed (e.g. staghorn sumac (Rhus typhina) and wild red raspberry (Rubus idaeus)) could be considered as sources of the errors observed for these two IAPS (Figure 6).

Figure 6. Examples of non-invasive alien plant species classified as Japanese knotweed Field photography of (a) staghorn sumac (Rhus typhina) from July 30, 2020, and (b) wild red raspberry (Rubus idaeus) from July 29, 2020.

4. Discussion

4.1. Global performances of multi-species classification

4.1.1. Mono-date classification

According to the range of performances based on the Kappa coefficient (Monserud 1990), very good performance was achieved with the mono-date classification (maximum Kappa coefficient = 0.81, XGBoost) for WV-3, while the classification with the SPOT-7 image was fair (maximum Kappa = 0.48, SVM).

Compared to the few studies that have mapped several IAPS over the same territory, the performances obtained for WV-3 (maximum overall accuracy = 91%) were comparable to that obtained by Ustin et al. (2002) (maximum overall accuracy>90%), who mapped four IAPS (different from those in the present study) in California (USA) from AVIRIS hyperspectral images acquired in summer.

However, our results showed better performances compared to those obtained by Michez et al. (2016) who mapped Japanese knotweed and purple jewelweed (Impatiens glandulifera) using aerial images acquired in autumn (50 cm spatial resolution) in a riparian area in southern Belgium (overall accuracy = 72%). The variability of luminosity in autumn (between acquisition campaigns), the presence of shadows, and a low coverage of targeted species were identified as the main factors related to their errors of classification. However, they obtained better results (overall accuracy = 72%) compared to ours with the use of SPOT-7 images (overall accuracy = 70%), also acquired in autumn. In our study, this fair performance could be related to the acquisition period (autumn) and the low separability between classes at this time of year (senescence for all species studied). Another source of classification errors could be related to low spatial resolution (6 m before pansharpening) compared to WV-3 (1.24 m before pansharpening). Spectral mixing increases for low spatial resolution images (pixel size greater than IAPS patch size) and would therefore be higher in the SPOT-7 than WV-3 image (Wang et al. 2022; Labonté et al. 2020). The spectral mixing could then explain low classification performances when SPOT-7 was used. The classification results using WV-3 show that remote sensing of IAPS at early stages of invasions should be based on very high spatial resolution imaging.

4.1.2. Multi-date classification

The combination of multi-date features including the BTBR index resulted in excellent performances (maximum Kappa = 0.85), higher than those obtained for mono-date classification (maximum Kappa = 0.81). These results highlight the significant contribution of the BTBR, which has also been shown in other studies carried out in similar environments. As an example, Martin et al. (2018) obtained a 61% detection rate for Japanese knotweed mapping by adding the modified BTBR (MBTBR) to the bands of the Pléiades images used. This detection rate was 50% and 59% when mono-date bands from autumn and summer were used, respectively. Dorigo et al. (2012) also obtained very good performance with an overall accuracy of 93% for mapping Japanese knotweed using multi-date drone images combined with the BTBR.

In addition to the BTBR, the most relevant features are the transformations of the initial bands, mainly vegetation indices and features from the IHS (Intensity, Hue, Saturation) transformation. These features are mostly WV-3-derived and represent 70% of the features used by the best classifier (RF). The under-representation of the features calculated from the SPOT-7 image once again confirms the low contribution of this image type.

The textural features and means of the spectral bands were the least-used features in the classifiers. Apart from the entropy of the red band of the SPOT-7 image, no other textural features are present among the 25 most important features identified by the RF method. However, some studies have demonstrated the importance of texture in the classification of IAPS such as Japanese knotweed and giant hogweed (Michez et al. 2016; Dorigo et al. 2012). In our study, this result is not surprising, although colonies of some species, such as Japanese knotweed, are easily visible in the images. The spatial resolution of the images, the size of the IAPS patches, and the level of separability from other native species are the main factors that influence the effectiveness of these features (Dorigo et al. 2012). Given that the WV-3 orthomosaic is at very high resolution and would be adequate for the use of texture to discriminate IAPS, the low contribution of texture features in this study could therefore be attributed to the presence of several native plant species with high intraspecific variability that are spectrally similar to the IAPS studied (e.g. staghorn sumac and wild red raspberry).

The bands were ranked using the Gini index (Breiman 2001) (Appendix 5), and among the means of the spectral bands, only the red-edge band was selected and ranked 1^st out of the 25 relevant features. In contrast, derived features such as maximum, skewness, and standard deviation were selected (i.e. 11 out of 25). In this study, absence of the mean features from the spectral bands is not congruent with the results obtained by other authors. Michez et al. (2016) noticed that the mean features from the RGB drone images were among the most important features in the classification of IAPS, including Japanese knotweed. This difference could be explained by the different composition of native species (i.e. many native species in our study area are similar to IAPS in the visible bands). It is therefore important to note that the use of this methodology in another environment should be adapted, taking into account the specificity of each environment, especially the composition of native plant species.

4.2. Individual class performances

4.2.1. IAPS spatial distribution

One of the major challenges in the classification of IAPS is to collect enough reference data to build robust and easily generalizable models. For the phragmites, for example, very dense patches are often located along highways, which limits access to these samples for modeling. Some of the reference data used are thus less representative (e.g. lower density), reducing the classification performances of these species due to spectral mixing. More specifically, a verification showed that several omission errors for these species were related to low density phragmites objects, which was also observed by Rupasinghe and Chow-Fraser (2021), who found that low cover negatively affected phragmites’ classification performance. Giant hogweed reference data are also difficult to access because of the systematic control (removal) conducted annually in the study area. This lack of reference data creates an imbalance between absence data and the other classes, affecting classification performances (Pranto and Paul 2021; Richhariya and Tanveer 2020; More and Rana 2017). The ratio of absence to IAPS data is approximately 2, 4, and 10 for Japanese knotweed, phragmites, and giant hogweed, respectively. One reason for the poor performances of SVM compared to RF and XGBoost could be related to this imbalance between classes. Several authors have found that some classifiers, such as SVM, are more sensitive to highly unbalanced data (Gašparović and Dobrinić 2020; Richhariya and Tanveer 2020; Lemnaru, Potolea, and Cordeiro 2012; Nguyen, Cooper, and Kamei 2011; Akbani, Kwek, and Japkowicz 2004) than classifiers based on decision trees such as RF and XGBoost (Gašparović and Dobrinić 2020; Zhao et al. 2018; More and Rana 2017).

In addition to the use of less sensitive learning classifiers, other alternative techniques that involve undersampling the majority class or oversampling the minority class by creating new synthetic samples (e.g. Synthetic Minority Oversampling Technique) can be used (Patil, Framewala, and Kazi 2020; Lin et al. 2018; Fernandez et al. 2018; More and Rana 2017). However, these techniques should be used with caution, as undersampling can reduce the diversity of absence data, especially in diverse environments such as our study area (Patil, Framewala, and Kazi 2020; Xu, Chen, and Sun 2019). Oversampling can also create highly correlated and interdependent synthetic data, which could suggest a reduction in the efficiency of certain methods for selecting relevant features (Blagus and Lusa 2013). Collaborative data collection and a sharing of IAPS distribution platforms are alternatives to improve access to reference data (e.g. the SENTINELLE platform of the Quebec government’s Ministry of Environment and Climate Change).

4.2.2. Acquisition date and very high spatial resolution satellite images availability

The high performance for Japanese knotweed and giant hogweed with the WV-3 orthomosaic alone seems to be related to the optimal acquisition period. The month of July corresponds to the period when Japanese knotweed patches are well developed and delimited, which can facilitate detection using an OBIA approach. Giant hogweed was also in flowering season and the white umbels visible in the WV-3 image help to discriminate them from other species (Michez et al. 2016). This result has financial and technical implications, as the use of multi-date images often requires additional financial costs and potentially energy-consuming image co-registration and calibration processing. This study therefore shows the effectiveness of using a single very high spatial resolution image (WV-3) for these two IAPS when acquired at the optimal period.

For phragmites, the poor performance compared to the other two IAPS can be related to a non-optimal acquisition period. The SPOT-7 image is less efficient for reasons mentioned in the previous sections, and the WV-3 image was acquired in July before its flowering that occurs in August. It is during the flowering period that the spectral signature of phragmites is most distinct from the other surrounding vegetation (Rupasinghe and Chow-Fraser 2021). For example, a study using WV-2/3 satellite imagery acquired during the optimal flowering period to detect phragmites resulted in an overall accuracy of 93% (Rupasinghe and Chow-Fraser 2021). When a mono-date image acquired at the optimal period for phragmites is not available, multi-date images can therefore be an interesting alternative to improve phragmite detection, as illustrated by the increase of the F1-score performance in our study from 0.71 to 0.87. Other studies have also demonstrated that using multi-date images can improve classifier performance for this species (Rupasinghe and Chow-Fraser 2019; Abeysinghe et al. 2019; Poulin, Davranche, and Lefebvre 2010). Abeysinghe et al. (2019) achieved 94% overall accuracy for phragmites detection using multi-date very high spatial resolution drone images. Rupasinghe and Chow-Fraser (2019) also obtained good performances (overall accuracy = 88%) using medium spatial resolution (Landsat 7/8 and Sentinel-2) imagery to map phragmites.

However, the dearth of both very high spatial and spectral resolution images is a limit in a context where multi-date images are needed to simultaneously map several IAPS with different optimal phenological periods. On the other hand, more easily accessible and available medium spatial resolution imagery (e.g. Landsat and Sentinel-2) can be used when the objective is to detect patches whose size is greater than or equal to that of the image pixel (Wang et al. 2022). For small IAPS patch detection, especially in the case of early detection, fusion techniques (e.g. super-resolution) between medium and high spatial resolution images could be useful (Chen et al. 2020) and should be explored.

Super-resolution is a technique of deriving high-resolution images from low spatial resolution images (Wang et al. 2022; Müller et al. 2020; Shermeyer and Van Etten 2019). Some of the most widely used techniques are based on deep learning (Wang et al. 2022; Bashir et al. 2021). For example, PlanetScope (3 m) and Sentinel-2 (10 m) images, acquired every day and every five days, respectively, can easily be available at optimal times. New multispectral band images with better spatial resolution can therefore be created to improve the classification obtained from initial low/medium spatial resolution images. Shermeyer and Van Etten (2019) were able to achieve a 36% improvement in the detection accuracy of several objects (e.g. cars, airplanes, and boats) using super-resolution to increase the resolution of WV-3 images (30 cm to 15 cm). However, this technique has been used less frequently for IAPS detection (Chen et al. 2020). A few studies have highlighted the potential of this technique to improve detection (Chen et al. 2020; Shermeyer and Van Etten 2019; Pouliot et al. 2018). Although this technique is demanding in computational resources, expertise, and processing time, it should be explored for IAPS classification from medium spatial resolution images when very high spatial resolution images are not available or accessible. In fact, this technique was tested in our study for phragmites detection using the fusion between Sentinel-2 (10 m) and Planetscope (3 m) images acquired during the flowering period (August 2020), and between SPOT-7 and GeoEye-1 images acquired in autumn (November 2020, 0.5 m). We used two different deep-learning-based techniques, namely residual convolutional neural networks (Latte and Lejeune 2020) and enlighten-generative adversarial networks (Gong et al. 2021). The performances were poor (Kappa<0.5) compared to the multi-date classification approach used in this study. Again, we hypothesize that the low density of colonies could explain these low accuracies for Sentinel-2 and Planetscope images, while the non-optimal acquisition period could explain the low accuracies obtained with SPOT-7 and GeoEye-1 images. This approach could therefore be improved by using a higher amount of representative reference data and by improving the timing for the images. It would also be interesting to explore classifications based on class density to properly identify the potential of applying super-resolution with medium (often available during optimal period) and high spatial resolution images.

5. Conclusion

This study was conducted in an urban heterogeneous environment and produced a multi-species classification of three IAPS using a multi-date approach, with good performances. These results demonstrate the potential of remote sensing to monitor IAPS in such an environment. For operational purposes, the mapped areas of presence can be used for the monitoring and control of these IAPS that colonize several types of environments in the study area. The main limitations of this methodology are linked to the small amount of reference IAPS data, the high cost of very high spatial resolution and multi-date images, and a dearth of these images at optimal phenological periods of detection. Further studies are therefore needed to analyze the mapping performance of IAPS in similar environments using easily available images (e.g. Sentinel-2 and Landsat acquired during optimal periods) combined with techniques such as super-resolution.

Acknowledgments

The authors wish to thank Nicolas Turcotte-Major (Ville de Québec), Camille Armellin (CERFO), Anne-Marie Dubois (CERFO), Jean Marchal (CERFO), Julie Deslandes (Ville de Québec), Mustapha Ramdani (Ville de Québec), Andréanne Hains (Cap Rouge River watershed council), and Nicolas Latte (University of Liège) for their support during the project.

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The data that support the findings of this study are openly available in the Open Science Framework data repository (link: https://osf.io/b3ufr/?view_only=b02f51fddf5f4c11bd7c457fe85c2710).

Additional information

Funding

This work was supported by the Mitacs Accélération program (award number IT16759) and by the Ville de Québec

APPENDIX Appendix 1:

The most used remote sensed data to map terrestrial invasive alien plants species

Table

Imagery source	Type of image	Used spectrum range	References
Landsat	Multispectral	Visible and NIR	Yu et al. (2020); Labonté et al. (2020); Becker, Zmijewski, and Crail (2013); Wilfong, Gorchov, and Henry (2009)
Sentinel-1	Radar	C band	Kattenborn et al. (2019)
Sentinel-2	Multispectral	Visible – NIR	Kattenborn et al. (2019)
SPOT-7	Multispectral	Visible – NIR	Abutaleb et al. (2020)
Ikonos	Multispectral	Visible – NIR	Lawrence, Wood, and Sheley (2006); Casady, Hanley, and Seelan (2005)
WorldView-2	Multispectral	Visible – NIR	Rizaludin Mahmud, Numata, and Hosaka (2020); Abutaleb et al. (2020); Paz-Kagan et al. (2019); Fernandes et al. (2014)
WorldView-3	Multispectral	Visible-NIR-SWIR	Shendryk et al. (2020)
Aerial imagery	Hyperspectral and LiDAR	Visible-NIR-SWIR	Ewald et al. (2020) Kattenborn et al. (2019); Marvin, Asner, and Schnitzer (2016); Ishii and Washitani (2013); Asner et al. (2008)
UAV	Multispectral	Visible – NIR	Qian et al. (2020); Kattenborn et al. (2019); Müllerová et al. (2017) Fernandes et al. (2014); Dorigo et al. (2012)
Terrestrial imagery	Hyperspectral	Visible-NIR – SWIR	Rizaludin Mahmud, Numata, and Hosaka (2020)

UAV: Unmanned Aerial Vehicle; SPOT-7: Satellite Pour l’Observation de la Terre-7; NIR: Near Infrared; SWIR: Short Waves Infra-red; LiDAR: Light Detection and Ranging

Appendix 2:

Distribution statistics of reference data used for classification

Table

	Number of polygons	Area (m^2)
		Min	Max	Mean	Median	SD	Total area
Japanese knotweed	128	13	919	124	93	115	41.235
Phragmites	76	24	417	160	146	86	16.934
Giant hogweed	32	29	203	84	74	47	2.707
Absence	283	3	1466	143	100	149	12.181
Total	519	-	-	-	-	-	73.057

Note: SD: Standard Deviation.

Appendix 3:

Illustration of fusion results: (a) initial image; (b) local mean and variance matching fusion; (c) Bayesian fusion; (d) Gram-Schmidt fusion; (e) ratio component substitution fusion; and (f) Zhang fusion

Appendix 4:

Used hyperparameters for Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Random Forest (RF) (WV- 3,+ SPOT- 7,+ BTBR)

Table

Classifier	Hyperparameter	Optimal Value
		WV-3	SPOT-7	WV-3 + SPOT-7 + BTBR
XGBoost	Nrounds	250	100	400
	Max depth	6	10	5
	Learning rate	0.4	0.4	0.3
	Gamma	0	0	0
	Colsample_bytree	0.6	0.6	0.6
	min_child_weight	1	1	1
	Subsample	0.55	0.55	0.66
	Number of features selected	35	15	35
RF	Mtry	7	7	7
	Ntree	400	400	400
	Number of features selected	25	15	25
SVM	Sigma	0.05	0.07	0.03
	Cost	2	2	4
	Number of features selected	25	15	35

Appendix 5:

Random Forest feature importance calculated using mean decrease in the Gini coefficient

RE: Red-Edge; BI: Brightness Index; NIR: Near-infrared; CI: Colour Index; GLDV: Gray Level Difference Vector; HSI: Hue, Saturation, Intensity; VARI: Visible Atmospherically Resistant Index; SR: Simple Ratio; BTBR: Bi-Temporal Band Ratio; RGR: Red Green Ratio; RENDVI: Red-Edge Normalized Difference Vegetation Index; NDWI: Normalized Difference Water Index; WV-3: WorldView-3; SPOT-7: Satellite Pour l’Observation de la Terre-7

Appendix 6:

Omission and commission errors obtained by the combination of WV- 3+ SPOT- 7+ BTBR and the best classifier (Random Forest)

Table

	Producer accuracy	Omission error (1- producer accuracy)	User accuracy	Commission error (1-user accuracy)
Japanese knotweed	0.97	0.03	0.88	0.12
Phragmites	0.81	0.19	0.94	0.06
Giant hogweed	0.90	0.10	0.82	0.18
Absence	0.91	0.09	0.93	0.07