Species-Level Vegetation Mapping in a Himalayan Treeline Ecotone Using Unmanned Aerial System (UAS) Imagery

Understanding ecological patterns and response to climate change requires unbiased data on species distribution. This can be challenging, especially in biodiverse but extreme environments like the Himalaya. This study presents the results of the first ever application of Unmanned Aerial Systems (UAS) imagery for species-level mapping of vegetation in the Himalaya following a hierarchical Geographic Object Based Image Analysis (GEOBIA) method. The first level of classification separated green vegetated objects from the rest with overall accuracy of 95%. At the second level, seven cover types were identified (including four woody vegetation species). For this, the suitability of various spectral, shape and textural features were tested for classifying them using an ensemble decision tree algorithm. Spectral features alone yielded ~70% accuracy (kappa 0.66) whereas adding textural and shape features marginally improved the accuracy (73%) but at the cost of a substantial increase in processing time. Contrast in plant morphological traits was the key to distinguishing nearby stands as different species. Hence, broad-leaved versus fine needle leaved vegetation were mapped more accurately than structurally similar classes such as Rhododendron anthopogon versus non-photosynthetic vegetation. Results highlight the potential and limitations of the suggested UAS-GEOBIA approach for detailed mapping of plant communities and suggests future research directions.


Introduction
Species composition of vegetation assemblages is a critical parameter for accessing biodiversity and ecosystem health [1,2]. Therefore, species-level maps of vegetation can provide useful information for monitoring ecosystem health and dynamics in terrestrial ecosystems [3]. Such maps can be of tremendous value especially for areas with high biodiversity that are expected to face a disproportionately higher impact of climate change. Generating an accurate map of vegetation composition at stand level is, however, a challenging task compared to broad scale vegetation characterization with remote sensing, which has emerged over the last few decades as an important tool for biodiversity monitoring [4,5]. For the Himalayan region, space-borne coarse to medium spatial resolution imagery has been utilized for broad biodiversity characterization and assessment [6,7], sample selection and classification of segments into classes of interest [34]. GEOBIA approach-based classification has lower likelihood of 'salt-and-pepper' speckle, which is often an issue with per-pixel analysis and GEOBIA also offers the possibility to include neighborhood, geometric and textural variables in improving classification of features [35]. GEOBIA-based tree species classification of UAS derived imagery has shown potential in recent studies [36,37] and it needs to be extensively tested for under-explored areas such as the Himalaya.
Determining optimal classification feature space is an important pre-classification step and has been investigated by previous studies using hyperspectral imagery for species level mapping of vegetation [29]. Following GEOBIA, the availability of hundreds of features (e.g., spectral, geometrical, textural, and contextual) makes feature space selection a time-consuming and subjective process [34]. Due to the large data size of UAS imagery, inclusion of higher/larger number of features may significantly increase processing time and inclusion of more features may not necessarily increase the mapping accuracy. Therefore, finding the most suitable classification features that maximize the mapping accuracy using the UAS + GEOBIA approach in the context of mountain vegetation mapping needs to be investigated for practical applicability of this data and method for species level mapping over larger areas.
This study is the first ever application of UAS remote sensing for vegetation mapping in the Himalaya. While UAS remote sensing applications for vegetation characterization have been conducted in other ecosystems such as the African savanna [38], temperate grasslands [39] and tropical forests [40], such applications for studying mountain vegetation are currently very few. In the Himalaya, a handful of UAS photogrammetric applications focusing on studying glacier dynamics [41] have been conducted, while UAS application for studying vegetation properties remains untested. The present study utilized data collected from a multi-spectral sensor (i.e., Parrot Sequoia) [42] to map tree and shrub species along the treeline ecotone in the Nepal Himalaya. Specifically, the objectives of this study are: (a) map the tree and shrub species of the treeline ecotone using multi-spectral UAS imagery and assess mapping accuracy, and (b) determine the optimal image parameters (spectral, geometric, textural matrices) for mapping tree species at treeline ecotone in the Himalaya.

Study Area
The site for this study was located at 28.10 • N, 85.3905 • E and was inside Langtang National Park in central Nepal ( Figure 1). Within the 0.81 ha study area the highest and lowest elevation recorded were 3791 m and 3766 m above mean sea level respectively and are part of the treeline ecotone. Abies spectabilis was a major treeline forming species while a tall shrub Rhododendron campanulatum was the spatially most dominant woody species followed by small shrubby species such as Rhododendron anthopogon and Berberis sp. Phenology of Rhododendron species appears to be highly sensitive to climate warming [43,44]. Therefore, these species have the potential to be used as indicator species for monitoring climate change impacts. Climate data from the nearest weather station showed that the average monthly total rainfall was highest in August (172.76 mm) and lowest in December (2.57 mm). Mean monthly temperature of the hottest (July) and coldest (December) month was 9.5 • C and −0.1 • C, respectively [45]. The understory was characterized by bare soil/rocky outcrops and non-photosynthetic vegetation (NPV) in the form of dry twigs and woody vegetation branches, possibly as a result of extraction of fuel wood by the local population and grazing by domesticated animals of nearby settlements.

UAS Flight Missions and Ground Control Points
A consumer-grade Phantom-3 professional multirotor UAS (DJI Inc., Shenzhen, China) was used to carry a Parrot sequoia camera (MicaSense Inc., Seattle, WA, USA). Sequoia is a narrow band multispectral camera that acquires images in four wavelength bands at a resolution of 1.2 million pixels and has a global shutter rather than a rolling shutter ( Table 1). The sequoia comes with a sunshine sensor which is mounted on top of the UAS (Figure 2c) and records irradiance in the same wavelength bands as the camera [42]. This information is utilized to calculate per-pixel reflectance values by accounting for the variability in ambient light conditions during the UAS mission [46]. UAS images were acquired on 16 June 2017 between 11 AM and noon under cloudy weather. To minimize the impact of terrain variability, the images were acquired in two flights missions each with separate launching location/height ( Figure 1c). UAS data acquisition for both the missions were conducted in an automated manner using the Micasense Atlas Flight application with predefined parameters (i.e., flying height: 30 m above ground, velocity: 3 m/s and forward/side overlap of 80 percent). The two flights resulted in a total of 1240 geotagged images (310 images × 4 spectral bands) taken from nearnadir viewing geometry. Flights were conducted at stable illumination and fixed camera settings, aiming to acquire images at uniform conditions. Photos of a calibrated reflectance panel produced by the camera manufacturer ( Figure 2a) were acquired immediately after the UAS flights. The average reflectance values of this panel for each waveband of the camera were utilized for calculating perpixel reflectance.

UAS Flight Missions and Ground Control Points
A consumer-grade Phantom-3 professional multirotor UAS (DJI Inc., Shenzhen, China) was used to carry a Parrot sequoia camera (MicaSense Inc., Seattle, WA, USA). Sequoia is a narrow band multi-spectral camera that acquires images in four wavelength bands at a resolution of 1.2 million pixels and has a global shutter rather than a rolling shutter ( Table 1). The sequoia comes with a sunshine sensor which is mounted on top of the UAS ( Figure 2c) and records irradiance in the same wavelength bands as the camera [42]. This information is utilized to calculate per-pixel reflectance values by accounting for the variability in ambient light conditions during the UAS mission [46]. UAS images were acquired on 16 June 2017 between 11 AM and noon under cloudy weather. To minimize the impact of terrain variability, the images were acquired in two flights missions each with separate launching location/height ( Figure 1c). UAS data acquisition for both the missions were conducted in an automated manner using the Micasense Atlas Flight application with predefined parameters (i.e., flying height: 30 m above ground, velocity: 3 m/s and forward/side overlap of 80 percent). The two flights resulted in a total of 1240 geotagged images (310 images × 4 spectral bands) taken from near-nadir viewing geometry. Flights were conducted at stable illumination and fixed camera settings, aiming to acquire images at uniform conditions. Photos of a calibrated reflectance panel produced by the camera manufacturer ( Figure 2a) were acquired immediately after the UAS flights. The average reflectance values of this panel for each waveband of the camera were utilized for calculating per-pixel reflectance. Prior to the UAS flights, five Ground Control Points (GCPs) in the form of bull's eye targets were deployed in the imaged area (example shown in Figures 1 and 2). The three-dimensional co-ordinates (WGS 84, UTM zone 45N) were determined with a Trimble GeoXH 6000 handheld GPS connected to a Zephyer 2 antenna [47] mounted on a monopod (average positional accuracy < 0.5 m). GCPs were occupied for 15 min each (5 s interval) and received error correction from GAGAN, an operational satellite-based augmentation system (SBAS) [48]. As a post-processing step to reduce uncertainty in GCPs location; spatial mean was calculated from the 180 logged positions (over 15 min) to determine the accurate 3-D position of each GCP.   Figures 1d and 2). The three-dimensional coordinates (WGS 84, UTM zone 45N) were determined with a Trimble GeoXH 6000 handheld GPS connected to a Zephyer 2 antenna [47] mounted on a monopod (average positional accuracy < 0.5 m). GCPs were occupied for 15 min each (5 s interval) and received error correction from GAGAN, an operational satellite-based augmentation system (SBAS) [48]. As a post-processing step to reduce uncertainty in GCPs location; spatial mean was calculated from the 180 logged positions (over 15 min) to determine the accurate 3-D position of each GCP.

Field Survey and Identifying Vegetation Classes
Field data collection was conducted simultaneously with UAS flights. Within the study area, the location of 57 individual woody plants in total were recorded using the Trimble GeoXH 6000 system described earlier (2 min occupancy). Four major woody plant species identified in the study site were Abies spectabilis, Rhododendron campanulatum, Rhododendron anthopogon and Berberis sp. (Figure 2d). These four species along with herbaceous ground vegetation (grass, forbs, mosses, etc.) were photosynthetically active/green vegetation. Other cover types identified include bare soil/rocks and non-photosynthetic vegetation (NPV) were combined under and non-green vegetation class and sample locations representing them were also collected during field work.

Image Processing
An overview of data collection and analysis workflow is shown in Figure 3. The UAS-derived images, GCP coordinates and images of the calibrated reflectance panel were brought into Pix4Dmapper Pro software (Lausanne, Switzerland) for processing via Structure-from-Motion (SfM). Following a standardized workflow, we selected the Ag Multispectral image processing template for processing [49]. The processing yielded an orthomosaic map, reflectance map and Normalized difference Vegetation Index (NDVI) for the study area with a ground sampling distance of 1.8 cm. Both orthomosaic and reflectance outputs are corrected for perspective error using a digital surface model (DSM). However, color balancing is applied on the orthomosaic to adjust the intensity of colors and produce a visually pleasing result. The reflectance bands are produced without any color balancing and the pixel values more accurately represent a target spectral reflectance [49]. Therefore, for detailed vegetation and cover type mapping this study utilized the reflectance bands.

Field Survey and Identifying Vegetation Classes
Field data collection was conducted simultaneously with UAS flights. Within the study area, the location of 57 individual woody plants in total were recorded using the Trimble GeoXH 6000 system described earlier (2 min occupancy). Four major woody plant species identified in the study site were Abies spectabilis, Rhododendron campanulatum, Rhododendron anthopogon and Berberis sp. (Figure 2d). These four species along with herbaceous ground vegetation (grass, forbs, mosses, etc.) were photosynthetically active/green vegetation. Other cover types identified include bare soil/rocks and non-photosynthetic vegetation (NPV) were combined under and non-green vegetation class and sample locations representing them were also collected during field work.

Image Processing
An overview of data collection and analysis workflow is shown in Figure 3. The UAS-derived images, GCP coordinates and images of the calibrated reflectance panel were brought into Pix4Dmapper Pro software (Lausanne, Switzerland) for processing via Structure-from-Motion (SfM). Following a standardized workflow, we selected the Ag Multispectral image processing template for processing [49]. The processing yielded an orthomosaic map, reflectance map and Normalized difference Vegetation Index (NDVI) for the study area with a ground sampling distance of 1.8 cm. Both orthomosaic and reflectance outputs are corrected for perspective error using a digital surface model (DSM). However, color balancing is applied on the orthomosaic to adjust the intensity of colors and produce a visually pleasing result. The reflectance bands are produced without any color balancing and the pixel values more accurately represent a target spectral reflectance [49]. Therefore, for detailed vegetation and cover type mapping this study utilized the reflectance bands.  The four reflectance bands and NDVI were stacked and multiplied by a scale factor of 10,000 and converted to a 16 bit unsigned integer to reduce data volume. A hierarchical GEOBIA was implemented into a eCognition Developer 9.2 (Trimble Inc., Sunnyvale, CA, USA) [50] that involved two steps: image segmentation and object classification. We first applied the multi-resolution segmentation (MRS) which generated objects by merging several pixels together based on relative homogeneity criteria [50]. This criteria is defined by setting values for the scale, shape/color and compactness/smoothness parameters. To generate smaller but spectrally pure objects, MRS was performed using a smaller scale factor (i.e., 50) and the red-edge (RE) band and the NDVI were given higher weights (1.5 and 2 respectively) to focus on photosynthetically active green vegetation. These values were deemed optimal after visually comparing results of several trial iterations. In the first classification, green vegetated objects were distinguished from other objects (bare soil/rock, non-photosynthetic vegetation) using NDVI thresholding where the optimal threshold was visually determined using the 'feature view' tool within eCognition Developer (Figure 4b). The four reflectance bands and NDVI were stacked and multiplied by a scale factor of 10,000 and converted to a 16 bit unsigned integer to reduce data volume. A hierarchical GEOBIA was implemented into a eCognition Developer 9.2 (Trimble Inc., Sunnyvale, CA, USA) [50] that involved two steps: image segmentation and object classification. We first applied the multi-resolution segmentation (MRS) which generated objects by merging several pixels together based on relative homogeneity criteria [50]. This criteria is defined by setting values for the scale, shape/color and compactness/smoothness parameters. To generate smaller but spectrally pure objects, MRS was performed using a smaller scale factor (i.e., 50) and the red-edge (RE) band and the NDVI were given higher weights (1.5 and 2 respectively) to focus on photosynthetically active green vegetation. These values were deemed optimal after visually comparing results of several trial iterations. In the first classification, green vegetated objects were distinguished from other objects (bare soil/rock, nonphotosynthetic vegetation) using NDVI thresholding where the optimal threshold was visually determined using the 'feature view' tool within eCognition Developer (Figure 4b).

Feature Space Selection and Level 2 Classification
Before the detailed classification of vegetated objects, a spectral difference segmentation (SDS) was performed only on the objects classified as green vegetation in level 1 classification. SDS merges neighboring objects if the difference in their spectral mean is below a given threshold (i.e., the objects are spectrally similar) [51]. Again, the optimal SDS parameters were determined by visually comparing results of several iterations. The resulting merged objects better represented the vegetation structure (e.g., a single tree stand represented by less than five objects after SDS compared to more than 20 objects before SDS) (Figure 4c). Training and validation objects for detailed classification at level 2 were selected by (i) intersecting surrounding image objects with specific sample for each class and (ii) including more objects in the neighborhood of field identified samples by visually confirming their cover type. This process resulted in the selection of a total of 386 sample objects, of which 240 (60%) were used for training and the remaining were used for accuracy assessment.

Feature Space Selection and Level 2 Classification
Before the detailed classification of vegetated objects, a spectral difference segmentation (SDS) was performed only on the objects classified as green vegetation in level 1 classification. SDS merges neighboring objects if the difference in their spectral mean is below a given threshold (i.e., the objects are spectrally similar) [51]. Again, the optimal SDS parameters were determined by visually comparing results of several iterations. The resulting merged objects better represented the vegetation structure (e.g., a single tree stand represented by less than five objects after SDS compared to more than 20 objects before SDS) (Figure 4c). Training and validation objects for detailed classification at level 2 were selected by (i) intersecting surrounding image objects with specific sample for each class and (ii) including more objects in the neighborhood of field identified samples by visually confirming their cover type. This process resulted in the selection of a total of 386 sample objects, of which 240 (60%) were used for training and the remaining were used for accuracy assessment.
To determine the suitability of different classification features, this study performed three classifications at level 2 with the following three sets of features: only spectral features, spectral features + geometric/shape features and spectral features + geometric/shape features + textural features ( Table 2). For each classification, the best features were selected using the Feature Space Optimization (FSO) and are shown in Table 3. The FSO tool calculates an optimum feature combination based on class samples and is available within eCognition. FSO evaluates the Euclidean distance in the feature space between the samples of all classes and selects a feature combination resulting in the best class separation distance, which is defined as the largest of the minimum distances between the least separable classes [50].
Level 2 classification was achieved using the random trees classifier (also called Random Forest) implemented within eCognition [52]. The random trees classifier builds multiple decision trees from bootstrap samples of the reference data. Each decision tree uses a random subset of training data and a random subset of input predictor variables. Roughly 2/3 of the data are sampled with replacement while 1/3 are withheld from tree construction (also called 'out-of-the-bag' or OOB samples). OOB samples are used to calculate the difference between predicted versus observed samples based on which error matrix is calculated.
Final class prediction is determined by majority voting based on the ensemble of trees. Random trees also calculate the measure of variable importance based on a random permutation of the input variables, and the effect of the permutation is quantified for each variable by the change in the OOB error [52]. Those variables important for separating classes show significant change in OOB error. Variable importance plot was generated and interpreted to determine the relative importance of feature variables. Classification accuracy was accessed using independent sample objects reserved for each class. Error matrices were calculated based on the number of objects per class without considering object size. Accuracy matrices reported include class-wise user's (UA) and producer's accuracy (PA), overall accuracy (OA) and kappa statistic (Tables 4 and 5). Table 2. Definition of the selected features that were used in image classification.

Mean layer value
The mean value represents the mean brightness of an image object within a single band

Standard deviation
The standard deviation of all pixels which form an image object within a band

Maximum difference
Minimum mean value of an object subtracted from its maximum value. The means of all bands belonging to an object are compared with each other and the result is divided by the brightness [50].

Shape
Border index: Describes how jagged an image object is, the more jagged, the higher its border index. Shape Index: a measure of overall shape complexity and it is calculated from the border length feature of the image object divided by four times the square root of its area [50].

Texture after Haralick
GLCM (grey level co-occurrence matrix) calculated after Haralick, Shanmugam, and Dinstein [53] describes how different combinations of pixel values occur within an object; used features (using the mean of all layers). Table 3. Selected optimal feature used for classification; 1 optimal feature space with only spectral features, 1,2 optimal feature space with spectral + shape features and 1,2,3 optimal feature space with spectral + shape+ textural features.

Results
Results of level 1 classification showed that about 65% of the mapped area was under green vegetation ( Figure 5). The overall accuracy achieved for level 1 was 95.2% (Table 4), confirming that NDVI thresholding of objects by determining the optimal threshold value is simple yet an effective strategy for separating green vegetated objects from others ( Figure 5a); this level of accuracy was observed by previous studies as well [54]. Further visual examination of level 1 classification results confirmed that some green vegetated objects with low vegetation density (dominated by Rhododendron anthopogon) had mean NDVI values below the threshold and resulted in false negative error (when classification erroneously predicted a low vegetation as non-vegetation object).
Based on the feature space selection method explained earlier, three optimized feature spaces were used that produced three classification outputs at level 2. Table 2 provides feature definitions and Table 3 lists all feature types that were used as candidate features and the numbering scheme explains if they were found to be optimal and used in final classification. Optimal features space using only spectral features consisted of 11 features. When both spectral and shape features were utilized, the optimal feature space consisted of 12 features (with only one shape feature) and when all three features (spectral + shape + texture) were utilized, the optimal feature space consisted of 16 features (11 spectral + 1 shape + 4 textural features). Results of level 2 classification using three sets of features show that the overall classification accuracy increased with the number and type of features used for classification. Highest overall accuracy (73%) and kappa (0.68) were obtained when all three feature types (spectral, shape, texture) were used. However, the accuracy with three feature types was not substantially higher compared to using just spectral features (accuracy of 70.77%) ( Figure 5 and Table 5). Per-class accuracy for level 2 results suggested that among green vegetation, Abies spectabilis was mapped with highest accuracy (100% UA), followed by Rhododendron campanulatum (85% UA), Rhododendron anthopogon (66% UA) and Berberis sp. (55% UA). Among non-green vegetated classes, bare soil/rock achieved 78% UA and non-photosynthetic vegetation was mapped at 69% UA. Spatial analysis show that Rhododendron campanulatum was spatially the most dominant cover type (31%) followed by Rhododendron anthopogon (25%) while Berberis sp. and Abies spectabilis were spatially least dominant (covering 3% and 2% area, respectively). Due to their larger spatial coverage, the inaccuracies in the mapping of the two Rhododendron species have higher ecological implications for monitoring treeline vegetation compared to others (e.g., Abies spectabilis) that are mapped at much higher accuracies.
Mapping accuracy was lower for Rhododendron anthopogon than Rhododendron campanulatum, and visual analysis showed that Rhododendron anthopogon was either misclassified as NPV or grass and bare soil was mistakenly classified as Rhododendron anthopogon. A possible explanation for this misclassification is the structural similarity between NPV and Rhododendron anthopogon (dwarf Rhododendron with dark green-to-brown small leaves and small shrubby structure). Although Rhododendron campanulatum was mapped with relatively higher accuracy, it was also misclassified as Rhododendron anthopogon. Visual analysis confirmed that this misclassification occurred in areas where Rhododendron campanulatum was immediately surrounded by dense Rhododendron anthopogon as the understory/below layer. misclassification is the structural similarity between NPV and Rhododendron anthopogon (dwarf Rhododendron with dark green-to-brown small leaves and small shrubby structure). Although Rhododendron campanulatum was mapped with relatively higher accuracy, it was also misclassified as Rhododendron anthopogon. Visual analysis confirmed that this misclassification occurred in areas where Rhododendron campanulatum was immediately surrounded by dense Rhododendron anthopogon as the understory/below layer.  The best accuracy achieving classification for level 2 (Figure 5d) used 16 features. The variable importance plot ( Figure 6) depicts 15 of these features in order of decreasing importance for discriminating vegetation at species level using multi-spectral UAS imagery. Mean NDVI at object level was the most important feature variable for level 2 classification. Furthermore, the presence of 13 spectral features among the top 15 features underscores the finding that spectral features alone provide an acceptable level of mapping accuracy. Adding shape and texture features improves the accuracy marginally. However it should be noted that classification using only spectral features was much faster (~6 min) compared to that using textural features (~50 min).

Discussion
The GEOBIA approach using multispectral UAS data has shown the potential for achieving species-level mapping for the Himalayan treeline ecotone. The exploratory work in finding the optimal segmentation approach (i.e., multi-resolution segmentation followed by spectral difference segmentation) and parameters (e.g., scale, shape and compactness criteria) helped in creating objects that were realistic in representing the spatial variability in the shape and size of the considered classes observed in the field (e.g., relatively larger and smooth canopy of Rhododendron campanulatum versus

Discussion
The GEOBIA approach using multispectral UAS data has shown the potential for achieving species-level mapping for the Himalayan treeline ecotone. The exploratory work in finding the optimal segmentation approach (i.e., multi-resolution segmentation followed by spectral difference segmentation) and parameters (e.g., scale, shape and compactness criteria) helped in creating objects that were realistic in representing the spatial variability in the shape and size of the considered classes observed in the field (e.g., relatively larger and smooth canopy of Rhododendron campanulatum versus smaller size and compact shape of NPV and soil/rock surface). The aim of species-level mapping was achieved with varying accuracy for different species. Certain classes could be mapped with better accuracy because of their morphological contrast.
When using spectral, spatial and textural features, we show that the spectral features contributed the most to a higher accuracy of species level mapping. Spectral features alone yielded~70% accuracy because the species that were mapped were not only taxonomically different but also structurally different, allowing more accurate classification with spectral features. For example, Rhododendron campanulatum is much taller (mean height 1.5 m) than Rhododendron anthopogon (mean height 0.45 m), and has different architecture of branching and different types of leaves; Abies spectabilis is fine needle leaved, whereas Berberis sp. is broadleaved. Including texture and shape features improves the accuracy marginally, but substantially increases the computational time (nearly 20 times, i.e., from 5 min to about an hour). This can easily scale up to massive computational requirements for mapping over a larger area, as has been reported in previous studies [55]; importantly, inclusion of texture variables has a disproportional effect on computational requirements. It should be noted that the species diversity in our mapping space is not very high and it remains to be tested if variable importance and mapping accuracy will change if either species diversity increases or the study area becomes bigger.
This study opens up possibilities for future ecological research in the Himalaya, based on the invaluable capability of UAS as a tool for providing VHR data that can be useful for ecological monitoring in remote and inaccessible areas. Our results improve upon previous studies that utilized ecological niche models based on the probability of the occurrence approach to map the potential distribution of selected Himalayan species [56,57] instead of the direct classification approach.
As the UAS platform and sensors are becoming cheaper and more sophisticated, future studies should test the suitability of better cameras (e.g., Micasense RedEdge-M and Altum, Gamaya Hyperspectral) [58] as well as UAS platforms that fly for a longer duration to cover larger areas and have vertical take-off and landing capability for data collection mountainous terrain (e.g., FireFLY 6 Pro) [59]. Also, the effects of mission parameters (e.g., flying height, percent overlap) and changing terrain (especially in Himalaya with considerable slope) on resultant map quality needs to be tested to establish a framework for quality UAS imagery collection. Our experience from doing exploratory UAS missions in Langtang Himalaya with varying parameters indicates that the effect of slope and terrain can be minimized by diving the study area into more than one flight and keeping high overlap in images (>80%). However, even with these parameter settings, some edge area of the mission may lack sufficient overlap and would create artifacts which induce inaccuracies in classification output.
Given the potential of UAS data for ecological monitoring, the number of studies using GEOBIA for UAS images analysis is expected to increase. However, for achieving usable accuracies using GEOBIA substantial visual interpretation of objects is required to optimize segmentation parameters, and this can be a challenge in practical applications of GEOBIA over large areas. Therefore, among the possibilities for GEOBIA technique refinement, there is scope for developing algorithms that semi-automatically evaluate the outputs of multiple segmentations and are able to determine the most optimal segmentation parameters for users with limited human interpretation e.g., Reference [60].

Conclusions and Future Work
To the best of our knowledge, this study is the first ever application of multi-spectral UAS imagery for individual tree-level vegetation mapping in the Himalaya with the goal of identifying species. By testing the usefulness of features related to spectral, shape and texture attributes of vegetation in maximizing the accurate identification of species, we present the most optimal and parsimonious model for classification. The Himalayan vegetation, particularly at the treeline ecotone, exhibits a wide range in vegetation composition and structure and is increasingly responsive to environmental impacts and climate change. Understanding the ecological processes and community response requires data collection that should be unbiased and be able to generalize at the landscape level. However, that is a challenging task for the alpine Himalayan communities because of (1) the dramatically heterogeneous topography, environment and community structure on the mountain slopes requiring adequate sampling of vegetation in space, and, (2) the formidable challenges of staying longer in the study site for adequate sampling. We show that data collection using UAS images and field data in areas along the treeline ecotone in the Himalaya has the potential to increase time efficiency and lead to vegetation maps with acceptable accuracy at the species level. This study is, therefore, a step forward in developing valuable future mechanisms to monitor vegetation diversity for a highly understudied yet ecologically important biodiversity hotspots. The classification results of the UAS imagery depend on a number of factors including type of camera, spatial resolution of images, environmental conditions, appropriate pre-processing, vegetation physiognomy, and more. Therefore, future studies in the Himalaya would require adequate ground truthing for calibration and validation of models with semi-automated classification. With all that, a reasonably accurate model of classification of UAS imagery offers an efficient, cost effective method of vegetation species mapping that can be of value for extreme environments like the Himalaya. In summary, UAS and GEOBIA presents us with an unpresented opportunity of scaling up vegetation mapping by many orders of magnitude compared to traditional sampling. This immediately means a more generalizable understanding of ecological processes and responses can emerge out of UAS data.