Mapping invasive plant with UAV-derived 3D mesh model in mountain area—A case study in Shenzhen Coast, China

Highlights

• We propose a 2D-to-3D strategy for invasive plants mapping in mountain area, using UAV-derived 2D digital ortho map (DOM) and 3D mesh model.

• The texture features overcome the shortcomings of RGB-only image, and helps to better distinguish invasive vegetation.

• A 3D mesh model could better visualize the final results, and show the distribution of invasive plants in the mountains.

Abstract

Invasive plants constitute one of the major causes of biodiversity loss, and its monitoring plays an important role in the management of coastal ecological systems. This study aimed to apply high precision 3D mesh-model and digital orthophoto map (DOM) derived from unmanned aerial vehicle (UAV) multi view images to monitor the invasive plants over coastal mountain region in Shenzhen, China. To overcome the limitations of RGB images, the Gray-level Co-occurrence Matrix (GLCM) features of images were analyzed and combined with spectral features to obtain the 2-dimentional distribution of invasive plants first, using an object-based image analysis technique. A fine analysis was then introduced to obtain a more accurate 3-dimentional distribution of invasive plant by combining 2-dimentional distribution of invasive plant and 3D mesh model. the results have shown that: (1) Although the UAV RGB image has limited spectral information, the low-altitude makes the spatial resolution very high, which can effectively enhance the effectiveness of the texture in mapping invasive plants, and finally achieved an overall accuracy of 93.25%. (2) The use of 3D mesh model, on the one hand, could significantly alleviate the impact of undulatory terrain over mountain area and improve the classification result; on the other hand, it could better visualize the final results, helping us more intuitively understanding the distribution of invasive plants. This study demonstrated the great potential of UAV-derived 3D mesh model in accurate natural resource management over mountain areas.

Introduction

Plant invasion has become a major manifestation of global change, and it may result in habitat degradation, local flora and fauna extinction and ecosystem function change (Daehler, 2003; Pimentel et al., 2000; Richardson et al., 2000a,b). Recent studies indicated that invasive species can not only adapt to local environment, but also change it (Callaway et al., 2010; Richardson et al., 2000a,b). A large number of alien invasive species in the United States cause huge economic losses (Pimentel et al., 2000). The damage caused by invasive plant in China is also quite amazing (Center, 2002). For examples, Mikania micrantha is an extremely fast growing, sprawling vine, it was discovered in Shenzhen City in 1984, and has been widely observed in the Pearl River Delta region since 2008; and in Neilingding Island with 472 ha, about 40–60% of this national nature reserve were infested and some forests had already degraded to wasteland, and 200 of the 600 plant species are suff ;ering from overgrowth of Mikania micrantha (Li et al., 2006; Wang et al., 2003). Thus, the monitoring of invasive plants is quite necessary and inevitable for ecological conservation and economic developments.

Obtaining the location, scope and evolution trend of species are indispensable in monitoring invasive plants (Alvareztaboada et al., 2017). Traditional ground monitoring methods (such as eye estimation and sampling) are less effective and accurate. Quantitative methods (such as field sampling) are more precise; however, they are time-consuming and usually inefficient in accurately mapping the distribution of invasive plants over the areas with complex terrain and dense jungles (Bråkenhielm and Liu, 1995; Godínez-Alvarez et al., 2009). Nowadays, remote sensing techniques have been widely applied in monitoring invasive plants, due to their possibility and convenience for accurate and timely acquisition of vegetation coverage over large area (Asner, 1998). Different satellite imageries with high spectral and spatial resolution, such as AVIRIS, QuickBird, Landsat TM and ETM+, were used to detect various types of invasive plants (Ustin et al., 2002; Wilfong et al., 2008). For example, The Gray-level Co-occurrence Matrix (GLCM) textures, proposed by Haralick (Haralick et al., 1973), were used to detect an invasive plant species (Leucaena leucocephala) in southern Taiwan from high resolution satellite (QuickBird) images (Tsai and Jhong Chou, 2006). Besides the satellite images, some high spatial resolution airborne images were also adopted to analyze the invasive species, such as the early detection of Solidago altissima in moist tall grassland and an invasive bryophyte species on the island of Sylt in Northern Germany (Asner et al., 2008; Ishii and Washitani, 2013; Skowronek et al., 2016).

Due to the spatial and temporal requirements of invasive plants monitoring, some critical limitations of satellite remote sensing should be considered, such as optimal spatial resolutions and flexible revisit time (Berni et al., 2009). Recently, tremendously developed unmanned aerial vehicle (UAV) holds great potential in mapping invasive plants. Compared with satellite remote sensing, UAV is much more convenient and cheap, and it can be deployed quickly and repeatedly to obtain high spatial and temporal resolution imageries of interested regions (Lelong et al., 2008; Xiang and Tian, 2011; Feng et al., 2015).A UAV equipped with a hyperspectral camera was applied to capture RGB images for controlling the growth of weed (Hiroshi et al., 2010). Campoy et al. (2015) applied UAV-based images and two classification approaches (pixel-based vs objected-based) to monitor ice plant invasion situation. High-resolution UAV images were used to mapping invasive Phragmites australis in coastal wetlands (Samiappan et al., 2017). UAV was also employed to create accurate distribution maps of yellow flag iris, an invasive plant at two lakes in the central interior of British Columbia (Hill et al., 2017).

Generally, two-dimensional images are employed to map invasive plants over flat terrain. However, for the regions with complex terrain, some objects may be obscured or compressed in the orthophotos of images, and the monitoring results of invasive plants might be less accurate (Gevaert et al., 2017). Thus, the methods with 3D features (such as height) are proposed, and a variety of methods are introduced to generate 3D features from 3D point clouds or digital surface models (Brodu and Lague, 2012; Vosselman and Dijkman, 2001). In recent years, textured 3D mesh model has attracted increasing attentions for accurate 3D object reconstructing and geometric information extraction, and it can be obtained from the sequences of continuous UAV Multi-View Stereo (MVS) images using photogrammetry methods (Remondino et al., 2012; Suveg and Vosselman, 2004; Murtiyoso et al., 2017). However, how to use these 3D models to classify land covers is challenging, and several studies have explored 3D features for land cover mapping. The topological features and multi-scale semantic relations were commonly used for 3D point cloud segmentation over urban scene (Richter and Behrens, 2013; Zhu et al., 2017). For example, after assigning labels to irregularly 3D points, individual trees were separated into individual objects within the labeled 3D points in urban areas (Weinmann et al., 2017). By integrating different data such as 2D orthomosaics, 2.5D Digital Surface Models (DSMs) and 3D point clouds, or combining various methods such as Implicit Shape Models (an object detection technique) and Markov Random Fields, 3D points can be classified with high accuracy (Gevaert et al., 2017; Knopp et al., 2011). For MVS 3D model classification, some methods first process 2D images and then output the result to 3D model (He and Upcroft, 2013). Some 3D meshes were classified into different parts, such as the human body and limbs, by using semi-supervised classifier with structure guiding (Liu et al., 2015). Based on the different 3D features such as elevation、planarity、verticality et al., semantic segmentation methods such as random forest and Markov Random Fields were applied to classify the textured 3D meshes (Rouhani et al., 2017; Blaha et al., 2017). These methods may overcome the limitation that complex terrain cannot be truly reflected in a two-dimensional image due to projection angle.

To obtain an accurate distribution of invasive plants in mountain areas, 3D mesh model should be taken into consideration as a complement to DOM. Existing methods have partly utilized 2D UAV/Satellite images and 3D point clouds for vegetation monitoring, however, these methods cannot be directly used for a textured 3D mesh model. Therefore, the main objectives of this study were: (1) to apply UAV RGB images with limited spectral features to develop an effective invasive plants mapping method and (2) to develop a novel approach to mapping invasive plants using UAV-derived 3D mesh model, which might significantly improve the mapping accuracy of invasive plants over mountainous areas.