Introduction

Invasive alien tree species threaten ecosystem integrity by modifying the structure and function of ecosystems and have negative impacts on ecosystem services delivery and native biodiversity (Le Maitre et al. 2011; Mooney 2005). In particular, three genera of invasive alien trees—acacias, pines and eucalyptus species are often listed as the worst offenders in South Africa (Gaertner et al. 2017; McConnachie et al. 2015), Brazil (de Abreu and Durigan 2011), Argentina (Zalba et al. 2008), Uruguay (Six et al. 2013), Kenya (Pellikka et al. 2009), New Zealand (Ledgard 2001), Hawaiian Islands (Daehler 2005) and India (Arasumani et al. 2019). These trees are native to Australia and were introduced to several tropical and subtropical countries in the nineteenth century (Richardson 1998). These trees were largely established on grasslands and shrublands which were historically classified as wastelands (Joshi et al. 2018; Rundel et al. 2014). All of these species are fast-growing, have high water use and show potential for encroachment into native grasslands and scrublands. Acacias, in particular form dense stands, maintain a high leaf area all through the year and have high evapotranspiration which impacts water yields of infested catchments (Dye and Jarmain 2004). In South Africa, these species have also had negative impacts on grassland birds (Allan et al. 1997; Armstrong and Van Hensbergen 1995), small mammals (Armstrong and Van Hensbergen 1995), invertebrates (Donnelly 1985) and plants (Richardson and Van Wilgen 1986).

In the Western Ghats, tropical montane grasslands, also known as Shola grasslands have been lost to exotic tree invasion at a rapid pace (Arasumani et al. 2018, 2019). This widespread invasion has impacted various faunal communities, including threatened species like the Nilgiri Pipit (Lele et al. 2020) and Nilgiri Tahr (Alempath 2008) in the Western Ghats. These trees were primarily established outside wildlife preserves and protected areas since the 1950s. However, today, the invasion of exotic trees is rapidly encroaching critical grassland landscapes, including protected areas (Arasumani et al. 2019; Joshi et al. 2018). Our prior research using Landsat data (Arasumani et al. 2019) indicates that 23% (340 sq. km.) of montane grasslands have been lost to invasive exotic tree stands within the past five decades. This data, however, had included all exotic trees as a single class due to the limitations with the spectral and spatial resolution of LANDSAT imagery. Maps classifying multiple invasive species into a single-class have limitations as some species, like Acacia and Pine, may invade remnant grasslands more rapidly than others. In the Western Ghats, Acacia mearnsii and Pinus patula, both listed among the most invasive species in the World (IUCN-GISD), co-occur with less invasive timber species as well as with native tropical forests (Joshi et al. 2018). Such a complex landscape makes the detection and estimation of the spread of the invasive trees challenging. One of the major challenges for conservation managers is to detect the invasion front, which typically consists of sparsely dispersed seedlings and saplings that are difficult to detect using remote sensing imageries (but see Arasumani et al. 2020). Detecting this invasion front is, however, critical in managing the invasion front as is identifying the best combination of algorithms and remote sensing platforms that can be used by conservation agencies in tropical areas across the globe.

Choice of imageries

Field-based, landscape-scale, invasive species mapping techniques are known to be challenging in tropical forests. Many researchers have attempted to identify broad forest type categories using space-borne and airborne images (Arasumani et al. 2018; Erinjery et al. 2018; Foody and Hill 1996; Shimizu et al. 2019). Discriminating between native and non-native species, however, remains challenging with medium resolution satellite data due to extensive intermixing of species and canopy heterogeneity. Such complex landscape scenarios may require high spatial and spectral resolution imageries for appropriately classifying invasive species. Researchers have also employed microwave data (Chen et al. 2018; Laurin et al. 2013; Wheeler et al. 2017), recently launched Sentinel-2 multispectral images (Laurin et al. 2013; Wheeler et al. 2017), and combinations of Sentinel-1 SAR (Synthetic Aperture Radar) images with Sentinel-2 multispectral images to improve the classification accuracy of tropical forests types and woody invasive species (Erinjery et al. 2018; Kattenborn et al. 2019; Zhang et al. 2019). A few studies have also suggested that hyperspectral images could produce high accuracy for mapping tree species compared to the multispectral datasets (Awad 2018; George et al. 2014; Thenkabail et al. 2004).

Choice of classification algorithms

Several classification algorithms have been used for classifying hyperspectral, multispectral, and SAR data: of these, support vector machine (SVM; Burgess et al. 2007), random forest (RF; Breiman 2001), and classification and regression tree (CART; Breiman et al. 1984) have been the most widely used especially on the publicly-available Google Earth Engine Platform (Gorelick et al. 2017). Although some studies pick the best classification algorithm for classifying hyperspectral data; all three algorithms are used for mapping landcover on both multispectral and SAR images (Lu et al. 2018) with perhaps differences in their training data requirements. The SVM is an iterative, non-parametric machine learning algorithm widely used for classifying hyperspectral images (Mountrakis et al. 2011). The SVM algorithm does not depend on the statistical distribution of the data but relies on training data adjacent to the class boundary to deliver high accuracy even with limited training data for classification (Melgani and Bruzzone 2004). The random forest (RF) algorithm is a non-parametric classifier that operates by generating a number of classification trees and selecting the mode of the predictions (Breiman 2001). CART models use recursive binary splits on predictor data in a decision tree framework to produce classifications at the end nodes of the trees. By nature of the classification process, CART models are considered somewhat easier to interpret compared to RF-based models (Lawrence and Wright 2001). Notably, RF and CART approaches are widely used to classify the remote sensing images but have higher training data requirements relative to SVM models (Delalay et al. 2019; Shaharum et al. 2020).

Objectives

Overall, this study aims to identify appropriate data sources and algorithms to identify exotic tree species on the Shola Sky Islands of the Western Ghats. Specifically, we test the ability of (a) AVIRIS-NG (hyperspectral), Sentinel-2B (multispectral) and Sentinel-1B (microwave) data with (b) SVM, Random Forest, and CART classification algorithms to discriminate invasive woody species (Acacia, Pine and Eucalyptus) from native tropical trees in the Shola Sky Islands. The overall goal is to test the efficacy of these imageries in detecting the invasion front—the ecotone where conservation efforts can be targeted.

Methods

Study area

For the purposes of this study, we selected an area admeasuring approximately 12 sq. km. in the Nilgiris mountains (Fig. 1) that has a gradient of woody invasive species infestation across this landscape. The study area contains natural habitats such as montane grasslands, montane forests and water bodies, and non-native woody trees—Acacia spp, Eucalyptus spp, and Pinus spp (Fig. 2).

Fig. 1
figure 1figure 1

a Study area—Nilgiri Hills. b A—AVIRIS-NG false colour composite image, B—Sentinel—2 false colour composite image, C—Sentinel—1 VV polarisation image with training data and D—Sentinel—2 VH polarisation with validation data

Full size image
Fig. 2
figure 2

Field photographs; a Montane grasslands, b Montane forests, c Pine, d Eucalyptus, e Acacia, and f Acacia invasion front

Full size image

Field data

In our study region, native forests are generally intermixed with non-native trees (Acacia spp., Pinus spp., and Eucalyptus spp.). However, only acacias were observed to be present at the actual grassland-forest invasion front (i.e., pines and eucalyptus were generally established in distinct stands). We collected the unique land cover GPS points using simple random sampling where target species land covered in excess of a 30 m*30 m footprint to obtain ‘pure’ endmember spectral signatures. We used 1049 ground-truth points of montane grasslands, forest, acacias, eucalyptus, pines and water bodies for image classification. For all locations, we ensured that the positional dilution of precision (PDOP) was lower than the AVIRIS-NG, Sentinel-1, and Sentinel-2- spatial resolutions (4 m and 10 m). To detect the Acacia spp. invasion front, we collected 73 additional ground truth locations concurrent with satellite imagery. We assessed the efficiency of acacia invasion front mapping with classifications produced by the three different image datasets and algorithms on the Google Earth Engine Platform as described below.

Image data

We obtained an AVIRIS-NG (Airborne Visible InfraRed Imaging Spectrometer—Next Generation) apparent at-surface reflectance products (March 2018) from the Jet Propulsion Laboratory (JPL), National Aeronautics and Space Administration data portal (NASA—https://avirisng.jpl.nasa.gov/dataportal/). AVIRIS-NG data have high spatial (4 m) and spectral (5 nm) spectral resolution with 425 spectral bands spanning 380–2510 nm. We excluded noisy bands and water vapour absorption bands (bands 1–10, 195–207, 287–316, 325–329) from the AVIRIS-NG dataset.

The Sentinel-2 Level-2 ground reflectance product was not available for our study area on the GEE platform for 2018. We, therefore, substituted this with the Sentinel-2 Level-1 product from the USGS (United States Geological Survey) Earth Explorer portal (https://earthexplorer.usgs.gov/), and we performed atmospheric corrections using Sen2Cor v2.8 in the Sentinel Application Platform (SNAP) to convert at-sensor radiance imagery to apparent at-surface reflectance (http://step.esa.int/main/snap-supported-plugins/sen2cor/sen2cor_v2-8/). We standardized all bands to a 10 m spatial resolution for all subsequent analyses.

We obtained the Sentinel-1 SAR Ground Range Detected (GRD) product (March 2018) directly from the GEE platform. This data was available as calibrated and ortho-corrected and was pre-processed using the Sentinel-1 toolbox for thermal noise removal, radiometric calibration, and terrain correction (using the SRTM 30 m spatial resolution digital elevation model). The final terrain-corrected data were log-transformed to decibels. We used Sentinel-1 VH (vertical transmit and horizontal receive) and VV (vertical transmit and vertical receive) polarization for image classification. We fused the Sentinel-1 (VV, VH) bands (spatial resolution 10 m) and Sentinel-2 (2, 3, 4, 5, 6, 7 and 8) bands (spatial resolution 10 m and 20 m) for all subsequent analyses.

Classification algorithms

All image classifications were conducted using the Google Earth Engine (GEE) platform. We chose GEE to enable the creation of a workflow that can be utilized by conservation managers in other regions to track and manage invasions. We tested a combination of Random Forest (RF; Breiman 2001), Classification and Regression Tree (CART; Breiman et al. 1984), and Support Vector Machine using radial basis functions (SVM; Burges 1998) to assess the skill of classification. We iterated the random forest trees from 0 to 50. We find that the RF classification accuracy improves as the number of trees increases; however, this pattern is seen in up to 6 trees. The classifier’s performance continues almost identical to 50 trees. For the SVM, parameter values (gamma and cost) were determined by using an iterative grid search; gramma 0.6 and cost 220 produced the high classification accuracy.

We assessed the accuracy of the classifications using an additional set of 300 ground truth points held out from all preceding analyses and estimated accuracy metrics (overall accuracy, user accuracy, producer accuracy and, Kappa coefficient) from the confusion matrix (Congalton 1991; Congalton and Green 2019).

Results

Comparison of different classification

We observed the highest overall accuracy (OA; 98.7%) and kappa coefficient (Kappa; 0.984) in AVIRIS-NG dataset with SVM classification (Table 1, Fig. 3), followed by S1 + S2 datasets with CART (OA—91.3%, Kappa—0.896), Sentinel-2 with SVM (OA—91.0%, Kappa—0.892) and Sentinel-2 with RF (OA—91.0%, Kappa—0.892).

Table 1 Image classification results
Full size table
Fig. 3
figure 3

The best-classified maps of native and non-native habitats. a AVIRIS-NG with SVM, b S1 + S2 datasets with RF, c Sentinel-2 with SVM, and d Sentinel-2 with RF

Full size image

The highest overall accuracy in the AVIRIS-NG dataset was recorded with SVM classification (OA—98.7, Kappa—0.984), followed by Random Forest classification (OA—88.7 Kappa—0.864) and CART classification (OA—85.7 Kappa—0.828). (Table 1, Fig. 4).

Fig. 4
figure 4

The AVRIS-NG classified maps of native and non-native habitats. a Classified with SVM, b RF, c CART

Full size image

The highest overall accuracy for Sentinel-2 data was also obtained with SVM classification (OA—91.0, Kappa—0.892) and RF classification (OA—91.0, Kappa—0.892), followed by CART classification (OA—86.3, Kappa—0.836) with S2 bands of 2, 3, 4, 5, 6, 7 and 8 (Table 1, Fig. 5).

Fig. 5
figure 5

The Sentinel-2 classified maps of native and non-native habitats. a Classified with SVM, b RF, and c CART

Full size image

The S1 + S2 datasets provided the highest overall accuracy with CART classification (OA—91.3, Kappa—0.896) and RF classification (OA—90.0, Kappa—0.88), followed by SVM classification (OA—82.0, Kappa—0.784) with Sentinel-2 bands 2, 3, 4, 5, 6, 7 and 8 and Sentinel-1 VV and VH polarisations (Table 1, Fig. 6).

Fig. 6
figure 6

The Sentinel-1 and Sentinel-2 classified maps of native and non-native habitats. a Classified with SVM, b RF, and c CART

Full size image

The Sentinel-1 dataset resulted in low accuracy of all classifiers compared to the other datasets with the highest overall accuracy with RF classification (OA—49.3, Kappa—0.392), and CART classification (OA—48.0, Kappa—0.376), followed by SVM classification (OA—48.3, Kappa-0.38) with VV and VH polarisations (Table 1, Fig. 7).

Fig. 7
figure 7

The Sentinel-1 classified maps of native and non-native habitats. a Classified with SVM, b RF, and c CART

Full size image

Best producer and user accuracy

The highest producer accuracy and user accuracy were obtained from the AVIRIS-NG dataset with SVM of all classes. Both Sentinel-2 imageries with SVM classification, and S1 + S2 datasets with CART classification produced high producer & user accuracy in detecting water, forests and grasslands, but had lower accuracy in identifying acacias, eucalyptus and pine. However, Sentinel-1 performed poorly, and the user and producer accuracy of all the classes was low (< 50%) except for water (Table 1).

Number of training samples vs classification accuracy

With limited data (25% of training samples), the AVIRIS-NG dataset with SVM classification performed well (OA—94%, Kappa—0.93) but RF and CART classifications yielded a lower OA (81%; Table 2). The Sentinel-2 with RF classification performed modestly (OA—89%, Kappa—0.85) with 25% of training samples. We observed that AVIRIS-NG dataset with SVM classification produced very high OA (97%) even with the 50% per cent of training samples; however, Sentinel-2 and S1 + S2 datasets yielded an overall accuracy of less than 90%. We noted that the classification accuracy for S1 + S2 datasets did not vary much with partial training datasets of 75% when compared to the complete training data set (Table 2).

Table 2 Accuracy assessment with 25%, 50%, 75% and 100% of training samples
Full size table

Detecting the acacia invasion front

We observed high accuracy in detecting acacia invasion front with AVIRIS-NG dataset with SVM classification (97%). The RF and CART classification, however, produced lower accuracy (< 40%). Moderate accuracy was observed in the Sentinel-2 with RF classification (60%) followed by SVM classification (58%) and then CART classification (38%). The S1 + S2 datasets and Sentinel-1 alone produced less than 40% accuracy across all classifiers.

Discussion

The systematic monitoring and mapping of invasive alien species are essential for the conservation and restoration of tropical grasslands. Therefore, spatially-explicit information on native habitats and non-native species is critical for sustainable forest management and forecasting landscape changes into the future.

Image data sources

In this study, we found that the hyperspectral dataset (AVIRIS-NG in this study) was the ideal platform for discriminating between native trees and non-native invasive trees with high precision. The AVIRIS-NG dataset accurately delineated the edges of non-native trees and native habitats, likely due to its high spatial and spectral resolution. The S1 + S2 datasets comprising multispectral and radar data proved to be a reasonable alternative but were only marginally better than using Sentinel-2 images alone. Finally, Sentinel-1 data did not produce sufficient accuracy for classifying and differentiating the invasive species from the native species in the tropical montane habitats. Similar results have been reported in classifying forest types with Sentinel-1 data in these tropical regions. Sentinel-1 data may, however, be probably useful when the Sentinel-2 images are affected by clouds (Erinjery et al. 2018).

Our study supports results obtained by others using hyperspectral remote sensing data for tropical tree species mapping efforts (Hyperion data—30 m spatial resolution 293 and 10 nm spectral resolution; George et al. 2014; Thenkabail et al. 2004). While Sentinel-1 images have shown to be useful for detecting and classifying water bodies in numerous studies (e.g. Bioresita et al. 2018; Hu et al. 2020), we were unable to map invasive woody species accurately. This may be due to similarities in the backscattering SAR signatures of native and non-native species. Sentinel-1 also has a shorter wavelength (C-band) that may not be able to differentiate the forest types based on the height and volume information where longer wavelengths (L-band) may be of advantage (Mitchell et al. 2014).

Although our study reports a relatively high accuracy of the combined Sentinel-1 and Sentinel-2 imageries, the difference with the use of only Sentinel-2 was marginal; we believe that this is due to our specific land cover types since Sentinel-1 image showed similar texture of all three invasive species. This, however, is different from other studies that have reported relatively higher accuracy in the classification from the combined Sentinel-1 and Sentinel-2 imageries for mapping different forest types (Erinjery et al. 2018; Kattenborn et al. 2019), agricultural applications (Veloso et al. 2017), mapping wetlands (Slagter et al. 2020), extracting urban extents (Iannelli and Gamba 2019), and delineating water bodies (Ahmad et al. 2019).

Classification algorithms

We observed the highest classification accuracy when using the SVM algorithm for classifying the hyperspectral data relative to RF and CART algorithms. This is in agreement with several other studies that have also observed SVM to be highly suitable for classifying hyperspectral data (Burai et al. 2015; George et al. 2014; Lim et al. 2019; Raczko and Zagajewski 2017). Both RF & SVM algorithms produced good classification accuracy with Sentinel-2 images compared to CART, as also observed by Lu et al. (2018). In our tests, we observe that CART is reasonably accurate, but some grassland and acacia invasions around the water bodies tended to be misclassified with pine trees; this was also the issue with a S1 + S2 classification. Although SVM performed well with Sentinel-2 data (91%), it produced low accuracy for the classification with the S1 + S2 datasets (82%). This is perhaps because the SVM algorithm does not do well with noisy data typical to Sentinel-1 data, and when target land cover class may have similar backscattered textures.

While Sentinel-2 did fairly well in detecting invasive species and natural habitats, we found that the combination of AVIRIS-NG data with an SVM classification model was the only sensor to detect acacia invasion-front with high accuracy. While we are not aware of other studies focusing on detecting invasion fronts, AVIRIS-NG data have been successful for mapping mangroves to the species (Chaube et al. 2019) and for crop type identifications (Salas et al. 2020).

In this research, we utilized medium (S1 and S2—10 m) and high spatial resolution (AVIRIS-NG—4 m) images for IAS mapping. We were, therefore, successful in mapping the species at the pixel scale. We suspect that sub-pixel classification methods may be more suitable for data with spatial resolution images over 30 m.

Influence of the number of ground truth points

We observed that a limited number of ground truth points (~ 25% of our points) were sufficient for obtaining high accuracies when using hyperspectral data (AVIRIS-NG). However, multispectral data (Sentinel-2) seemed to require a higher density of ground truth points (> 50%) to get reasonable accuracies. If the study area is large, Sentinel-2 data might need three times the number of ground truth points than that required to classify a similar hyperspectral image. Conversely, hyperspectral data (AVIRIS-NG) is not available for all places, but data can be captured using a UAV-borne hyperspectral sensor. Using an UAV, however, might require significant effort and investment for large study regions.

Influence of spatial and spectral resolution

We believe that the higher spatial resolution of AVIRIS-NG dataset over S1 and S2 provides a distinct advantage in distinguishing smaller native forest patches from non-native trees. We found that the S1 and S2 datasets were not able to detect acacia invasion due to the lower spatial resolution compared to AVIRIS-NG dataset.

The fine spectral resolution of AVIRIS-NG (spanning 425 spectral bands) compared to Sentinel-2 (8 spectral bands) helped discriminate between individual native and invasive species. Other studies have shown similarly high species-level classification accuracies in complex landscapes (George et al. 2014; Skowronek et al. 2017). Spectral signature overlaps in the Sentinel-2 dataset resulted in lower average accuracy in detecting pines and eucalyptus similar to findings from Pu et al. (2012) and Marshall and Thenkabail (2015).

Inferences for conservation managers

Mapping and distinguishing native trees from non-native trees is an essential task for land managers for conserving native and endemic species, assessing fire risks, and impacts on ecosystem services. We found that airborne hyperspectral imagery provides the best solution for detecting invasive species and the invasion front. However, the availability of hyperspectral images is a major limiting factor for most locations. These data were only available only for 12 sq. km. of our study area. Further, procuring such data can be prohibitively expensive. Where these data are not available, we recommend using Sentinel-2 satellite data with SVM or RF algorithms as it offers a reasonable compromise between accurately discriminating native and non-native trees while allowing the mapping of large spatial extents. Sentinel-2 images, however, require a high-density of ground truth points and may still be unsuitable for mapping young invasion because of the limited spatial and spectral resolution. Conservation managers may also consider using RapidEye images with an object-oriented classification approach if they need to map a mixed-species invasion front, and do not need to discriminate among the invasive species along that front (Arasumani et al. 2020). If the invasion front requires constant monitoring in a smaller study area, an unmanned aerial vehicle with a hyperspectral sensor may be indispensable. Finally, we also recommend that conservation managers and restoration NGOs use the online Google Earth Engine (GEE) platform because of the high processing power required for processing high spatial and spectral resolution data. We provide the GEE code used for this study in Appendix 1. We have detailed our recommendations for conservation managers in Table 3.

Table 3 Recommendations for conservation managers
Full size table

In this research, we demonstrate the advantage of remotely-sensed hyperspectral, multispectral, and radar data for mapping, and distinguishing native and non-native invasive species using AVIRIS-NG, Sentinel-1 and Sentinel-2 datasets. Although the high-resolution, hyperspectral AVIRIS-NG dataset proved superior, we were severely constrained by its spatial availability. The multispectral Sentinel-2 data, on the other hand, were useful in identifying native and non-native trees across a large landscape. Combined with the processing power of the GEE platform, this study demonstrates the opportunity for species-specific discrimination of invaded landscapes that can be replicated across the globe.