Ecological Informatics 50 (2019) 43–50
Contents lists available at ScienceDirect
Ecological Informatics journal homepage: www.elsevier.com/locate/ecolinf
Mapping vegetation types in semi-arid riparian regions using random forest and object-based image approach: A case study of the Colorado River Ecosystem, Grand Canyon, Arizona
T
⁎
Uyen Nguyena, , Edward P. Glenna, Thanh Duc Dangb, Lien T.H. Phamc a
The University of Arizona, Department of Soil, Water and Environmental Science, AZ 85721-0038, USA Thuy Loi University, Institute for Water and Environment Research, Ho Chi Minh City, Viet Nam c Vietnam National University Ho Chi Minh City, University of Science, Faculty of Environment, Ho Chi Minh City, Viet Nam b
A R T I C LE I N FO
A B S T R A C T
Keywords: Airborne imagery OBIA Random Forest Grand Canyon Vegetation classification
Riparian regions are essential habitats for wildlife and play a vital role in agricultural production, but they are highly dynamic environments impacted by fluctuations of water levels. Monitoring vegetation types along narrow river corridors is complicated and requires high-resolution imagery and advanced remote sensing techniques due to the mixture of vegetation and other types of land covers. The primary aim of this paper is to develop a framework using airborne imagery, object-based image approach (OBIA), hyper-spectral analysis and Random Forest to classify vegetation along narrow, semi-arid riparian corridors via a case study of the Grand Canyon, the Colorado River. By analyzing hyper-spectral and field data with Random Forest, we found that the bandwidths from 642 to 682 nm and 750 to 870 nm were useful for vegetation classification in this case study. As a result, the red and near-infrared bands of aerial photos were used with ancillary data for species classification, and the overall accuracy (OA) of classification with these images reached up to 94.8% with a Kappa's coefficient of 0.93. The similarity of vegetation phenology caused most of the misclassified cases. Low cost unmanned aerial vehicles should be used to acquire more frequent data, which is essential to understand how vegetation patterns change over time.
1. Introduction Monitoring and mapping vegetation dynamics along riparian corridors are important for detecting ecological responses to human or natural disturbances in river systems (Congalton et al., 2002; Gould, 2000; Klemas, 2014; Villarreal et al., 2012; Weber and Dunno, 2001). Human-induced changes include the introduction of exotic species, diversion of water for human uses, channelization of rivers to protect properties and other land use modifications that can lead to the deterioration of riparian ecosystems (Busch and Smith, 1995; Briggs, 1996). Besides anthropogenic drivers, vegetation patterns in riparian zones are also subject to natural factors such as periodic floods and droughts (Naiman and Decamps, 1997). Recently, climate change is altering riparian corridors through direct effects on vegetation and alterations of the hydrological cycle (Palmer et al., 2008). Accurate and reproducible mapping techniques are necessary to guide management practices to confront the impacts of human intervention and climate change on riverine vegetation.
⁎
Among mapping techniques, remote sensing has advantages over traditional land survey methods such as larger spatial coverage since spaceborne and airborne devices can observe large and/or distant regions with varying levels of regularity. Traditionally, vegetation mapping was carried out by ground surveys, but these methods are laborious, expensive, time-consuming and difficult to be applied to large regions (Dang et al., 2018a; Muller, 1997). In addition to a larger spatial coverage, various remote sensing sources with a wide range of spectral and temporal properties can be employed to map different types of vegetation (Recknagel, 2001; Akasheh et al., 2008; Xie et al., 2008; Rocchini et al., 2015). The choice of remotely sensed imagery depends on the classification goal of each project because each remote sensing platform has its pros and cons. For example, Landsat imagery, which has been freely available with 30 m resolution and 16-day return period from 1972, provides a high frequency of temporal coverage (Baker et al., 2006), but its spatial resolution is generally too coarse to resolve vegetation to the species level (Turner et al., 2003). In Nagler et al. (2005) and Norman
Corresponding author. E-mail address:
[email protected] (U. Nguyen).
https://doi.org/10.1016/j.ecoinf.2018.12.006 Received 16 June 2018; Received in revised form 18 December 2018; Accepted 19 December 2018 Available online 27 December 2018 1574-9541/ © 2018 Published by Elsevier B.V.
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
large variety of fauna can be found in low areas (Ralston et al., 2008). Past floods and droughts were the main controls on the distribution of flora in the region (Stromberg, 2001). In the Colorado River basin, the effect of the Glen Canyon Dam on spring floods was described in Schmidt et al. (1998) and Topping et al. (2003). Due to dam regulations, flows had been reduced compared to natural conditions. As a result, riparian vegetation adapted to the modified hydrological regime. Prior to the dam, field surveys reported that Prosopis glandulosa and Celtis reticulata (as in Fig. 1), for example, were dominant vegetation in much of the canyon, and their dense growth affected wildlife habitat quality and geomorphic processes of riverbanks and sandbars (Kearsley et al., 2006; Manner et al., 2014; Schmidt et al., 1998; Sogge et al., 2003). The construction of the dam, however, impacted flow regimes and sedimentation downstream and introduced invasive species such as Pluchea sericea, Baccharis salicifolia and Tamarix ramosissima (Hazel et al., 2006; Kearsley et al., 2006; Melis, 2011).
et al. (2018), Landsat 7 ETM+ and Landsat 8 OLI images were used to monitor vegetation in Arizona, but they only considered riparian forests as a group because of the limited spatial resolution. In some case studies, such as Anderson et al. (2018), Landsat was used to classify vegetation to the species level, but only for large homogeneous areas. While the resolution of Terra MODIS is too coarse for species classification, Sentinel missions were only launched recently (Sentinel 2 in 2015 https://sentinel.esa.int/web/sentinel/missions). Besides optical images, Synthetic Aperture Radar (SAR) was also employed to classify vegetation up to the species level. For example, ALOS-2 PALSAR-2 was applied to distinguish mangrove species (Pham et al., 2018a). Nevertheless, LiDAR and SAR were normally used for the case studies that considered only a few species (Fassnacht et al., 2016). High resolution remotely sensed imagery can be effective for mapping vegetation to the species level in a complex environment (Lane et al., 2014). High-resolution images, however, did not always out-perform coarse images when using the pixel-based approach, since single pixels may only capture partial information of target objects (Fassnacht et al., 2016; Pham et al., 2016, 2018b). To overcome this drawback, we employed the object-based image analysis (OBIA) approach which groups homogeneous and continuous pixels into objects to reduce the spectral variability within a target object before classification (Blaschke, 2010; Pham et al., 2016; Pham et al., 2018b). In terms of vegetation classification, machine learning offers various ways to map land cover types, but non-parametric methods such as Support Vector Machine (SVM) or Random Forest (RF) were preferred because input data did not need to be normally distributed (Faasnacht et al., 2016; Vafaei et al., 2018; Pham et al., 2018b). Random Forest was chosen because it has several advantages over other methods. This classifier does not result in overfitting and can be used for categorical values (Breiman, 2001). Moreover, Random Forest releases robust results even when the number of observations is smaller than the number of variables or the variables are correlated. Random Forest also requires fewer parameters than the other techniques such as Support Vector Machine (SVM) (Pal, 2005; Pham et al., 2016; Pham and Brabyn, 2017). Finally, Random Forest can provide a measure of how important a variable is. While the combination of airborne images and object-based image analysis approach (OBIA) is likely critical for vegetation classification in narrow semi-arid regions with mixed species, the use of OBIA and random forest adds the advantages of a state-of-the-art machine learning technique. Hence, this study aims to investigate the use of airborne images and advanced image analysis techniques including the object-based image analysis approach (OBIA) and the random forest classifier in classifying vegetation along a narrow, semi-arid river corridor to the species level. A part of the Grand Canyon National Park in Arizona, an iconic symbol of the U.S., was chosen as a case study because this complex region is a mosaic of different kinds of vegetation, water and bare lands. This study provides a helpful methodology for obtaining data to support adaptive management in this region in the future. The framework can also be applied to other arid and semi-arid riverine ecosystems undergoing ecological changes over the world.
2.2. Image acquisition Airborne photographs were acquired from the Grand Canyon Monitoring and Research Center of the United States Geological Survey (USGS; see Davis, 2012). The image collections were selected for two main reasons. First, we classified vegetation to the species level, so very high spatial resolution of remotely sensed data was required. The spatial resolution of these aerial photos, from 20 to 44 cm per pixel, was adequate for the aim of this study. Second, the study area is full of shadow, and low sunlight at 10 am in the morning during the overpass window limits the use of the available satellite imagery. A series of 143 images covers the entire Arizona part of the Colorado River Ecosystem. These images were acquired by a Leica ADS-40 camera, which had a spectral panchromatic band (465–680 nm) and four other bands (BLUE 450–510 nm, GREEN 530–576 nm, RED 642–682 nm and NIR 770–814 nm) during the vegetative growing season (see Ralston et al., 2008 for more details). Out of the 143 collected images, only 34 images, which have the highest density vegetation in the canyon and ground survey data, were used in this research for classification. 2.3. Field data collection
2. Materials
Reference data were collected from 451 field locations to provide a detailed spatial distribution of 12 primary vegetation classes in the study areas (listed in Table 1 and shown Fig. 1). The collected field data were divided into two groups: two thirds for training and one third for model validation. The defined vegetation classes were based on previous vegetation classification work of Kearsley and Ayers (1996) and a combination of existing vegetation classes described in Nature Serve (www.natureserve.org). In addition to the above species, field data showed that there were other types of vegetation such as creosote bush, white bursage, catclaw acacia or pinyon pine in the high elevations. However, they were not the target of this research, which is focused on the complex mosaic of vegetation, water and bare lands in the riverine corridor.
2.1. Study area
3. Methodology
The Colorado River Ecosystem, which encompasses 69,750 km2 of land areas, is a portion of the Glen Canyon National Recreation Area, the Grand Canyon National Park and the Hualapai reservation in the U.S. (Ralston et al., 2008). The corridor from the Glen Canyon Dam to Lake Mead extends for about 450 km and is about 75 km wide along each side of the river channel. This region is an arid and semi-arid river system, and vegetation mostly exists along the river corridor (Fig. 1). The lack of water and the severity of climatic conditions prevent the development of various types of vegetation at high elevations, but a
Fig. 2 provides an overview of the methodology. A detailed description and justification of the five main steps are discussed further in the following sections. 3.1. Orthorectification and atmospheric corrections To remove image distortions caused by platform-induced geometry errors, we used a DEM that was extracted from existing topographic maps and ground control points (GCPs) derived from hand-held GPS. 44
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
Fig. 1. A river segment in the Grand Canyon, the Colorado River, Arizona and dominant vegetation species in the region.
of classification and regression tree (CART), was used to determine important bands for classification. Instead of using a single decision tree, this technique uses many decision trees that are generated using different bootstrapped samples from training data (Breiman, 2001). Random Forest includes the following steps as described in Breiman (2001): (1) generating n training datasets (ntree) from a given training dataset of a size k by randomly taking samples. Data points from the original dataset may be used several times or may never be used in this process. (2) an optimal number of predictors (mtry) is determined by a trial and error method to minimize bias. (3) the prediction is derived from majority votes based on ntree. For this study, the values of ntree and mtry were set to 1000 and 5, respectively.
Table 1 12 primary vegetation classes found in the study area. Abbreviation
Common name
Scientific name
ATCA BASL CERE EQFE LATR PHAU PLSE PRGL SAEX SAGO TYDO TARA
Four wing salt-bush Seep willow Hackberry Horsetails Creosote bush Common reed Arrowweed Mesquite Coyote willow Gooddings willow Cattails Tamarisk
Atriplex canescens Baccharis salicifolia Celtis reticulata Equisetum ferrissii Larrea tridentata Phragmites australis Pluchea sericea Prosopis glandulosa Salix exigua Salix gooddingii Typha domingensis Tamarix ramosissima
3.3. Image segmentation In this step, typically, pixels are grouped based on homogeneity thresholds to create separate regions (also called objects) that differ in spectral and shape features from their neighbours. The multiresolution segmentation algorithm in eCognition Developer 9.1 with a homogeneity criterion that is a combination of spectral homogeneity and shape homogeneity was used to generate image objects (Trimble Germany GmbH, 2015). The image segmentation process reduces the number of units for classification significantly and can make the classification easier since the pixels in the image are merged into multiple discrete segments.
The process to perform image orthorectification was run in ArcGIS. Atmospheric correction to remove atmospheric and illumination effects from aerial images was done by using the Atmospheric Correction Algorithm ATCOR-3 which is described in detail in Richter and Schläpfer (2013). 3.2. Hyper-spectral data analysis Ground reflectance for each kind of vegetation was measured by a hand-held hyperspectral device. The values of ground reflectance data for different vegetation types are shown in Fig. 3. Based on hyperspectral data analysis and physiognomic and floristic similarities, we then regrouped the species in Table 1 into five classes. This helps to create a more meaningful set of data and reduces uncertainties in classification caused by the similarity in vegetation structures. The list of the five revised classes and their formation classes are tabulated in Table 2. The random forest technique, a non-parametric classifier and a kind
3.4. Water and soil masks As the study region is a riverine arid and semi-arid region, which includes water and bare soil objects, masking was applied to separate vegetation from other kinds of land covers. The NDWI and NDVI indices, which were the two indicators sensitive to the change in water content and vegetation phenology, were used for masking as in Pham 45
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
Fig. 2. Species classification using the object-based image analysis approach (OBIA), airborne images and random forest classifier.
3.5. Vegetation classification
and Brabyn (2017). Firstly, the Normalized Difference Water Index (NDWI) was used to separate water from other objects. We classified an object as a water body if its averaged NDWI values (of pixels that formed the object) was higher than 0. The remaining objects were classified as soil and vegetation. In the next step, objects with Normalized Difference Vegetation Index (NDVI) values of < 0.3 were considered as soil and were removed. The two indices were calculated as following:
NDWI =
GREEN− NIR GREEN+ NIR
(1)
NDVI =
NIR − RED NIR + RED
(2)
In addition to spectral bands of the image and textures derived from these spectral bands, we used ancillary data such as elevation, slope, and distances to water sources to improve vegetation classification as in Pham et al. (2016). Results from machine learning analysis showed that the bands with wavelengths from 642 to 682 nm and 750 to 870 nm were important predictors for vegetation classification (Fig. 4 shows a comparison between a false color image with an image in natural colors). Most of the previous studies as summarized by Fassnacht et al. (2016) showed that VIS and NIR bands were useful in vegetation classification. Our range was basically in the middle of the ranges proposed in Chuvieco and Huete (2009) (from 640 to 880 nm) and Cho et al. (2010) (from 650 to 1000 nm). These light spectra are related to growth and vigor
where: NIR – near-infrared band; RED – red band; GREEN – green band.
Fig. 3. Spectral sampling analyses from the ground of Cattail, Phragmite, Grass, Equisetum, and 4-wing Saltbush (left) and Baccharis, Mesquite, Coyote willow, Tamarisk and Arrowweed (right). 46
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
Table 2 List of revised vegetation classes for the Colorado River Ecosystem. Group Analysis
Association
Formation name
Formation class
TARA BASL/SAEX PLSE PRGL WTLD
Tamarix ramosissima Baccharis emoryi/Salix exigua Pluchea sericea Prosopis glandulosa Mix species
Temporarily Flooded Micro-phyllous Shrubland Seasonally Flooded Cold-Deciduous Shrubland Seasonally Flood Shrubland Temporarily Flooded Cold-Deciduous Woodland Semi-permanently flooded Temperate or Subpolar Grassland
Shrubland Shrubland Shrubland Woodland Herbaceous Vegetation
4. Results and discussion 4.1. Image segmentation analysis The segmentation process divided the airborne photos into objects that represent features with uniform physiognomy and floristics. In our study, all the four input spectral bands of the aerial photos (RED, GREEN, BLUE, and NIR) were used for the segmentation, with an equal weight of one for all bands. Shape parameters were set at 0.2 because the shape information contributes less to identifying the homogeneous regions than the spectral information. The segmented objects were masked as described above using the NDWI and NDVI indices to remove water and bare soil objects. Segmentation results after masking are shown in Fig. 5 as an example.
Fig. 4. Aerial images in false color and Natural color.
4.2. Classification results
characteristics of plants. Despite the importance of the blue and green wavelength regions reported in Key et al. (2001), Waser et al. (2011) and Pham et al. (2016), these regions were less important in this study. This may be due to different chemical plant components (water content, sugar and carbohydrate content, protein content and aromatics) and the absorption features of vegetation pigments as discussed in Fassnacht et al. (2016) and Xue and Su (2017). The choice of the specific bandwidths helps to select the bands of remotely sensed imagery that should be used for classification. As a result of this analysis, we used only the Near-Infrared and red bands and the ancillary data to delineate different species (Table 3). A digital elevation model (DEM) provided by the United States Geological Service (USGS) was resampled to the resolution of the images using the nearest neighbor resampling. Random Forest was used again to classify vegetation types. Accuracy assessment: The final map products were subjected to accuracy assessments by comparing mapped vegetation to ground survey data (Congalton, 1991). The accuracy of the RF classification algorithm was determined by using a confusion matrix (Congalton, 1991). Overall accuracy (OA), Users' Accuracy (UA), Producer's Accuracy (PA) and Kappa Coefficient, which are well-known indices for thematic map accuracy assessment, were calculated as in Pham et al. (2016; 2017). PA reported the percentage of real features correctly shown on classified maps. UA indicates how often classified objects appeared on the ground.
Overall, the classification algorithm performed well with accuracy assessment showing a high overall accuracy of 94.8% and a kappa coefficient of 0.93 (see Table 4). The accuracy indices (UA and PA) for each kind of vegetation are also shown in Table 4. Wetland classes had very high both producer and user accuracies (> 99%), and the producer and user accuracies of other classes were above 80% (Table 4). Saltcedar, mesquites, and arrowweeds, which were larger objects, could be adequately distinguished from each other. On the other hand, common reed, bulrush, horsetail and cattails could not be separated and were placed in a single wetland class. As reported in Table 4, most of the misclassified cases were caused by the similarity in structure and canopy of the species in different stages. Young mesquite and saltcedar trees were difficult to separate from each other and sometimes overlapped with the arrowweed class. Individual emergent marsh species could not be separated from each other, but the general class of marsh or wetland habitat was identifiable. The high level of accuracy supports the applicability of our framework in classifying vegetation to species level. A classification result is shown in Fig. 6 as an example. The high PA and UA values showed that the combination of airborne imagery, the combined Random Forest and OBIA method was highly effective in complex semi-arid environments. The level of accuracy was high compared to similar studies using the OBIA approach such as Pham et al. (2016); Pham and Brabyn (2017) which may be a result of the appropriate spatial resolution of imagery. In a semi-arid environment, tree species are mixed with the surrounding environments, and only pixels that are smaller than tree canopy diameters
Table 3 Variables were used for classification. Categories Spectral
Topographic
Object's feature variables
Number of features
of NIR, RED • Means deviations of NIR, RED • Standard variables of NIR and RED layers: GLCM mean, GLCM standard deviation, GLCM correlation, GLCM homogeneity, GLCM contrast, • Texture GLCM dissimilarity, GLCM entropy, GLDV mean, GLDV contrast, GLDV entropy • DEM • Slope • Distance to the river
47
24
3
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
Fig. 5. Segmented vegetation objects (white polygons) after removing water and bare soil on a background of an airborne image.
the future studies, information such as surrounding context, other vegetation characteristics (density, height or shape), and vegetation indices should be added to improve the classification accuracy (Pham et al., 2016; Pham et al., 2018b; Vafaei et al., 2018). Human infrastructure, climate change and sea level rise have changed the magnitude of flooding, water level's rising and falling rates, and flooding extents in river and floodplains around the world (Angarita et al., 2018; Beyene et al., 2010; Dang et al., 2016, 2018b; Hecht et al., 2019; Marengo et al., 2012; Palmer et al., 2008). These challenges require more regular monitoring efforts. Additionally, we only used the 34 images that covered the high density of vegetation with a total area of 414 ha, so we could not analyze the percentage of vegetation coverage in the canyon in this study. More frequent data acquisition activities with larger spatial coverage by low-cost unmanned aerial vehicles as in Baena et al. (2018) could provide a more comprehensive view of the region in order to detect the changes due to future climate change. It also would be advantageous to develop a fusion approach to combine the benefit of airborne imagery with regular satellite-based images to reduce the high cost of implementing this methodology.
could capture their spatial variability. This conclusion is similar to the case study of Peña et al. (2013) who studied a native forest in Andres, Chile with airborne images. The “close-to-leaf” pixels, however, which contained too much noise, then may also affect the classification of larger species. Pixels in the pixel-based approach contain inadequate data for classification and too much noise in the classification results. The OBIA approach applied in this study was beneficial as it reduced the variability of homogeneous structures of leaves to improve the PA and UA values. Additionally, the application of the random forest classifier in this study had low bias although the technique might sometimes overfit the classification vegetation species at different ages (Fassnacht et al., 2016). Systematic data collection might reduce this form of uncertainty. Yang and Everit (2010) reported that there was no single technique that worked for every region in weed classification. The framework in this study worked well in a narrow semi-arid region with mixed species, and this technique should be assessed with other environments. Machine learning techniques are similar to the decision-making processes in human brains. If we support more, but more appropriate information, the classification result may become more accurate. In this study, we used several ancillary data such as DEM, slope, and distances to water sources to support machine learning analyses. Although the overall accuracy was considerably high, it is also recommended that in
Table 4 Average Percent Producer Accuracy and User Accuracy for all 34 images and explanation of misclassification. Colorado River Ecosystem Vegetation classes
Producer Accuracy
User Accuracy
Confused species explanation
TARA
99.5%
80.4%
PLSE
93%
100%
PRGL
91%
98.8%
BASL/SAEX WTLD
92.8% 99.7%
98.1% 100%
Confused with Pluchea (PLSE) at the low-density areas and with Prosopis (PRGL) at the young, short tree areas Most confuse with other sparse classes, especially TARA. Ground reflection signature may influence of misclassification Confused with Tamarix because they have similar structure and canopy, especially in the young, small, short tree areas Most confused with Pluchea and confused among this group class Confused with Pluchea and Salix spp., Baccharis spp. class especially in the low-density vegetation
Overall accuracy Kappa coefficient
94.8% 0.93
48
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
Fig. 6. Aerial false color composite image (left) and classification results (right) (Note: EQFE & PHAU in the WTLD group).
5. Conclusion
Baena, S., Boyd, D.S., Moat, J., 2018. UAVs in pursuit of plant conservation-real world experiences. Ecol. informatics 47, 2–9. Baker, C., Lawrence, R., Montagne, C., Patten, D., 2006. Mapping wetlands and riparian areas using Landsat ETM+ imagery and decision-tree based models. Wetlands 26, 465–474. Beyene, T., Lettenmaier, D.P., Kabat, P., 2010. Hydrologic impacts of climate change on the Nile River Basin: implications of the 2007 IPCC scenarios. Clim. Chang. 100 (3–4), 433–461. Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J. Photogramm. Remote Sens. (1), 2–16. Breiman, L., 2001. Random Forest. Machine Learning. 45. pp. 5. Briggs, M., 1996. Riparian Recovery in Arid Lands, Strategies and References. University of Arizona Press, Tucson. Busch, D.E., Smith, S.D., 1995. Mechanisms associated with decline of woody species in riparian ecosystems of the southwestern US. Ecol. Monogr. 65 (3), 347–370. Cho, M.A., Debba, P., Mathieu, R., Naidoo, L., Van Aardt, J., Asner, G.P., 2010. Improving discrimination of savanna tree species through a multiple-endmember spectral angle mapper approach: Canopy-level analysis. IEEE Trans. Geosci. Remote Sens. 48 (11), 4133–4142. Chuvieco, E., Huete, A., 2009. Fundamental of Satellite Remote Sensing. Taylor& Francis Ltd, London. Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 37, 36–46. Congalton, R.G., Birch, K., Jones, R., Schriever, J., 2002. Evaluating remotely sensed techniques for mapping riparian vegetation. Comput. Electron. Agric. 37, 113–126. Dang, T.D., Cochrane, T.A., Arias, M.E., Van, P.D.T., Vries, T.T., 2016. Hydrological alterations from water infrastructure development in the Mekong floodplains. Hydrol. Process. 30 (21), 3824–3838. Dang, T.D., Cochrane, T.A., Arias, M.E., 2018a. Quantifying suspended sediment dynamics in mega deltas using remote sensing data: a case study of the Mekong floodplains. Int. J. Appl. Earth Obs. Geoinf. 68, 105–115. Dang, T.D., Cochrane, T.A., Arias, M.E., Tri, V.P.D., 2018b. Future hydrological alterations in the Mekong Delta under the impact of water resources development, land subsidence and sea level rise. J. Hydrol. 15, 119–133. Davis, P.A., 2012. Airborne digital-image data for monitoring the Colorado River corridor below Glen Canyon Dam, Arizona, 2009—Image-mosaic production and comparison with 2002 and 2005 image mosaics. In: U.S. Geological Survey Open-File Report 2012–1139, 82 p. Available at. http://pubs.usgs.gov/of/2012/1139/. Fassnacht, F.E., Latifi, H., Stereńczak, K., Modzelewska, A., Lefsky, M., Waser, L.T., ... Ghosh, A., 2016. Review of studies on tree species classification from remotely sensed data. Remote Sens. Environ. 186, 64–87. Gould, W., 2000. Remote sensing of vegetation, plant species richness, and regional biodiversity Hotspots. Ecol. Appl. 10, 1861–1870. Hazel, J.E., Topping, D.J., Schmidt, J.C., Kaplinski, M., 2006. Influence of a dam on finesediment storage in a canyon river. J. Geophys. Res. Earth Surf. 111 (F1), 1–16. Hecht, J.S., Lacombe, G., Arias, M.E., Dang, T.D., Piman, T., Jan 2019. Hydropower dams of the Mekong River basin: a review of their hydrological impacts. J. Hydrol. 568, 285–300. Kearsley, M.J., Ayers, T.J., 1996. Effects of the 1996 Beach/Habitat building flow on riparian vegetation in Grand Canyon. In: Final Report. U.S. Department of Interior, Bureau of Reclamation Cooperative Agreement CA 1425-96-FC-81-05006 (65 pages). Kearsley, M.J., Cobb, N.S., Yard, H.K., Lightfoot, D.C., Brantley, S.L., Carpenter, G.C., Frey, J., 2006. Inventory and monitoring of terrestrial riparian resources in the Colorado River Corridor of Grand Canyon: an integrative approach. In: Grand Canyon Monitoring and Research Center AZ 86001, (262 pages). Key, T., Warner, T.A., McGraw, J.B., Fajvan, M.A., 2001. A comparison of multispectral and multitemporal information in high spatial resolution imagery for classification of individual tree species in a temperate hardwood forest. Remote Sens. Environ. 75 (1),
The combination of the object-based image approach, random forest, aerial photos, and the use of a hand-held hyper-spectral device as presented in this study takes advantage of both remote sensing data and ancillary data. By focusing on larger target objects, the object-based image approach reduced the variability of the pixels belonging to each object in the photos, which could confuse the classification and therefore improves classification accuracy. Random Forest using hyperspectral data showed that the red and NIR bands were important in classifying vegetation classification in this case study. Random Forest also produced robust results even though the number of observations was small. This is not uncommon in remote sensing studies because it is expensive to collect a large number of data in the field. High overall accuracy and Kappa's coefficient implied that the framework in this study may be applicable for similar narrow riverine environments with shadow coverage. The ability to classify the vegetation in the Grand Canyon to the species level could help understand the change in vegetation patterns over time. In the context of global climate change, understanding the vegetation dynamics from the response of the surrounding environment is also important in designing mitigation strategies and maintaining ecological values. Acknowledgement Financial support for this work was provided to Grand Canyon Monitoring and Research Center and the University of Arizona. We would like to thank David J. Wilcox (SAV Ecology, Monitoring & Restoration Program, Virginia Institute of Marine Science, College of William & Mary) for his useful comments on this manuscript. The corresponding author, Uyen Nguyen, would like to express her gratitude to Prof. Edward P Glenn for his support during the whole period of her study. References Akasheh, A., Neale, C.M.U., Jayanthi, H., 2008. Detailed mapping of riparian vegetation in the middle Rio Grande River using high resolution multi-spectral airborne remote sensing. J. Arid Environ. 72, 1734–1744. Anderson, M., Gao, F., Knipper, K., Hain, C., Dulaney, W., Baldocchi, D., Eichelmann, E., Hemes, K., Yang, Y., Medellin-Zuara, J., Kustas, W., 2018. File-scale assessment of land and water use change over the California delta using remote sensing. Remote Sens. 10 (6), 889. Angarita, H., Wickel, A.J., Sieber, J., Chavarro, J., Maldonado-Ocampo, J.A., Herrera-R, G.A., ... Purkey, D., 2018. Basin-scale impacts of hydropower development on the Mompós Depression wetlands, Colombia. Hydrol. Earth Syst. Sci. 22 (5), 2839.
49
Ecological Informatics 50 (2019) 43–50
U. Nguyen et al.
Sens. Appl. Soc. Environ. 13, 298–305. Ralston, B.E., Davis, P.A., Weber, R.M., Rundall, J.M., 2008. A vegetation database for the Colorado river ecosystem from Glen Canyon dam to the western boundary of Grand Canyon National Park, Arizona. In: U.S.G.S. Open-File Report 2008-1216. Recknagel, F., 2001. Applications of machine learning to ecological modelling. Ecol. Model. 146 (1–3), 303–310. Richter, R., Schläpfer, D., 2013. Atmospheric/Topographic Correction for Airborne Imagery. ATCOR-4 User Guide, 565-02. Rocchini, D., Hernández-Stefanoni, J.L., He, K.S., 2015. Advancing species diversity estimate by remotely sensed proxies: a conceptual review. Ecol. Informatics 25, 22–28. Schmidt, J.C., Webb, R.H., Valdez, R.A., Marzolf, G.R., Stevens, L.E., 1998. Science and values in river restoration in the Grand Canyon. Bioscience 48, 735–747. Sogge, M.K., Sferra, S.J., McCarthey, T.D., Williams, S.O., Kus, B.E., 2003. Distribution and characteristics of southwestern willow flycatcher breeding side and territories 1993–2001. Stud. Avian Biol. 26, 5–11. Stromberg, J., 2001. Restoration of riparian vegetation in the south-western United States: importance of flow regimes and fluvial dynamism. J. Arid Environ. 49, 17–34. Topping, D.J., Schmidt, J.C., Vierra, L.E., 2003. Computation and analysis of the instantaneous-discharge record for the Colorado River at Lees Ferry, Arizona: May 8, 1921, through September 30, 2000 Rep. 118 pp.. US Geol. Surv. Prof. Pap. 1677. Trimble Germany GmbH, 2015. Trimble Documentation: eCognition Developer 9.1 Reference Book. Trimble Germany GmbH, Munich, Germany. Turner, W., Spector, S., Gardiner, N., Fladeland, M., Sterling, E., Steininger, M., 2003. Remote sensing for biodiversity science and conservation. Trends Ecol. Evol. 18 (6), 306–314. Vafaei, S., Soosani, J., Adeli, K., Fadaei, H., Naghavi, H., Pham, T.D., Tien, Bui D., 2018. Improving accuracy estimation of forest aboveground biomass based on incorporation of ALOS-2 PALSAR-2 and sentinel-2A imagery and machine learning: a case study of the Hyrcanian forest area (Iran). Remote Sens. 2018 (10), 172. Villarreal, M.L., Van Leeuwen, W.J.D., Romo-Leon, R., 2012. Mapping and monitoring riparian vegetation distribution, structure and composition with regression tree models and post-classification change metrics. Int. J. Remote Sens. 33, 4266–4290. Waser, L.T., Ginzler, C., Kuechler, M., Baltsavias, E., Hurni, L., 2011. Semi-automatic classification of tree species in different forest ecosystems by spectral and geometric variables derived from Airborne Digital Sensor (ADS40) and RC30 data. Remote Sens. Environ. 115 (1), 76–85. Weber, R., Dunno, G., 2001. Riparian vegetation mapping and image processing techniques, Hopi Indian Reservation, AZ. Photogramm. Eng. Remote. Sens. 67, 179–186. Xie, Yichun, Sha, Zongyao, Yu, Mei, 2008. Remote sensing imagery in vegetation mapping: a review. J. Plant Ecol. 1 (1), 9–23. Xue, J., Su, B., 2017. Significant remote sensing vegetation indices: a review of developments and applications. J. Sens. 2017, 17. https://doi.org/10.1155/2017/ 1353691. Article ID 1353691. Yang, C., Everit, J.H., 2010. Mapping three invasive weeds using airborne hyperspectral imagery. Ecol. informatics 5 (5), 429–439.
100–112. Klemas, V., 2014. Remote sensing of riparian and wetland buffers: an overview. J. Coast. Res. 30, 869–880. Lane, C.R., Liu, H., Autrey, B.C., Anenkhonov, O.A., Chepinoga, V.V., Wu, Q., 2014. Improved wetland classification using eight-band high resolution satellite imagery and a hybrid approach. Remote Sens. 6 (12), 12187–12216. Manner, R.B., Schmidt, J.C., Scott, M.L., 2014. Mechanisms of vegetation-induced channel narrowing of an unregulated canyon river: results from a natural field-scale experiment. Geomorphology 211, 100–115. Marengo, J.A., Chou, S.C., Kay, G., Alves, L.M., Pesquero, J.F., Soares, W.R., Santos, D.C., Lyra, A.A., Sueiro, G., Betts, R., Chagas, D.J., 2012. Development of regional future climate change scenarios in South America using the Eta CPTEC/HadCM3 climate change projections: climatology and regional analyses for the Amazon, São Francisco and the Paraná River basins. Clim. Dyn. 38 (9–10), 1829–1848. Melis, T.S. (Ed.), 2011. Effects of three high-flow experiments on the Colorado River ecosystem downstream from Glen Canyon Dam. Arizona Rep. 1366 147 pp., U.S. Geological Survey Circular. Muller, E., 1997. Mapping riparian vegetation along rivers: old concepts and new methods. Aquat. Bot. 58, 411–437. Nagler, P.L., Glenn, E., Hursh, K., Huete, A., 2005. Vegetation mapping for change detection on an arid-zone river. Environ. Model. Assess. 109, 255–274. Naiman, R.J., Decamps, H., 1997. The ecology of interfaces: riparian zones. Annu. Rev. Ecol. Syst. 28, 621–658. Norman, L.M., Middleton, B.R., Wilson, N.R., 2018. Remote sensing analysis of vegetation at the San Carlos Apache reservation, Arizona and surrounding area. J. Appl. Rem. Sens. 12 (2), 026017. Pal, M., 2005. Random forest classifier for remote sensing classification. Int. J. Remote Sens. 26 (1), 217–222. Palmer, A., Liermann, C.A., Nilsson, C., Florke, M., Alcamo, J., Lake, P.S., Bond, N., 2008. Climate change and the world's river basins: anticipating management options. Front. Ecol. Environ. 6, 81–89. Peña, M.A., Cruz, P., Roig, M., 2013. The effect of spectral and spatial degradation of hyperspectral imagery for the Sclerophyll tree species classification. Int. J. Remote Sens. 34 (20), 7113–7130. Pham, L.T., Brabyn, L., 2017. Monitoring mangrove biomass change in Vietnam using SPOT images and an object-based approach combined with machine learning algorithms. ISPRS J. Photogramm. Remote Sens. 128, 86–97. Pham, L.T., Brabyn, L., Ashraf, S., 2016. Combining QuickBird, LiDAR, and GIS topography indices to identify a single native tree species in a complex landscape using an object-based classification approach. Int. J. Appl. Earth Obs. Geoinf. 50, 187–197. Pham, T.D., Yoshino, K., Le, N.N., Bui, D.T., 2018a. Estimating aboveground biomass of a mangrove plantation on the Northern coast of Vietnam using machine learning techniques with an integration of ALOS-2 PALSAR-2 and Sentinel-2A data. Int. J. Remote Sens. 1–28. Pham, L.T., Vo, T.Q., Dang, T.D., Uyen, T.N., 2018b. Monitoring mangrove association changes in the can Gio biosphere reserve and implications for management. Rem.
50