Search

Highlights • Structure-from-motion increases accuracy of UAV multispectral data classification. • Overall accuracy at species/genera level was 88.2% using data from a single sensor. • Fusion reduced the dependency on the allocation of ground truth data. Abstract High quality data on plant species occurrence count among the essential data sources for ecological research and conservation purposes. Ecologically valuable small grain mosaics of heterogeneous shrub and herbaceous formations however pose a challenging environment for creating such species occurrence maps. Remote sensing can be useful for such purposes, it however faces several challenges, especially the need of ultra high spatial resolution (centimeters) data and distinguishing between plant species or genera. Unmanned aerial vehicles (UAVs) are capable of producing data with sufficient resolution; their use for identification of plant species is however still largely unexplored. A fusion of spectral data with LiDAR-derived vertical information can improve the classification accuracy, such a solution is however costly. A cheaper alternative of vertical data acquisition can be represented by the use of the structure-from-motion photogrammetry (SfM) utilizing the images taken for (multi/hyper)spectral analysis. We investigated the use of such a fusion of UAV-borne multispectral and SfM-derived vertical information acquired from a single sensor for classification of shrubland vegetation at species level and compared its accuracy with that derived from multispectral information only. Multispectral images were acquired using Tetracam Micro-MCA6 camera in the west of Czechia in a shrubland landscape protected within the NATURA 2000 network. Using (i) multispectral imagery only and (ii) multispectral-SfM fusion, we classified the vegetation into six classes representing four woody plant species and two meadow types. Our results prove that the multispectral-SfM fusion performs significantly better than multispectral only (88.2% overall accuracy, 85.2% mean producer's accuracy and 85.7% mean user's accuracy for fusion instead of 73.3%, 75.1% and 63.7%, respectively, for multispectral). We concluded that the fusion of multispectral and SfM information acquired from a single UAV sensor is a viable method for shrub species mapping. [ABSTRACT FROM AUTHOR]

Unmanned aerial vehicle (UAV) platforms are rapidly becoming popular in many research and industry sectors. Due to their relatively low purchase price and the fact they can be used to monitor areas that are difficult or even unsafe to access, they have been increasingly used in land surveying and mapping of smaller areas. Numerous UAV platforms equipped with various cameras are increasingly available on the market, differing in their suitability for environmental mapping. Surveyors therefore face a question whether to buy or assemble their own UAV. The objective of this study is to assess the performance of two fixed-wing UAV systems for land survey and mapping applications. In particular, we: (1) compared a commercial eBee platform equipped with a Sony Cybershot DSC-WX220 camera with zoom lens and a home assembled EasyStar II equipped with Nikon Coolpix A with a lens of fixed focal length to find out if a home-assembled solution can compete with specialized commercial platform; (2) investigated the utilization of UAV images acquired under leaf-off conditions for digital terrain model (DTM) generation with respect to vegetation cover (steppes and forests); (3) assessed whether an increase in the image quantity can compensate for a lower quality of images; and (4) compared the DTM derived from UAV imagery with the official Czech Republic airborne laser scanning (ALS)-derived DTM. One flight with Easystar II and two perpendicular flights with eBee were performed. From these three flights, four point clouds were derived (one from each flight, and one resulting from a combination of two eBee flights), supplemented with four ground filtered point clouds. The accuracy of point clouds and DTM was assessed through a comparison with a conventional GNSS survey. We successfully identified the bare ground during the leaf-off period in the deciduous forest using images from both platforms. Point densities of point clouds acquired with Easystar II exceeded the densities of those acquired with eBee even after combining images from two eBee flights. Root mean square error of all derived point clouds ranged between 0.11 and 0.19 m, exceeding the accuracy of a nationwide ALS-derived DTM in both forest and open steppe areas. The most accurate point cloud was acquired using Easystar II. This is likely due to a combined effect of the quality of onboard cameras, camera settings and environmental conditions during the flight. For users who prefer to have greater control over their options rather than being dependent on the commercially available kit solution, home-assembled kits utilizing drones capable of carrying any camera available on the market may be an advantage. [ABSTRACT FROM AUTHOR]

Detail plant species classification using very high spatial resolution data is a challenging task. Exploring the potential of imagery acquired by Unmanned Aerial Vehicle (UAV) to identify individual species of plants and assessing values of additional inputs such as height and thermal information into classification process are hot research topics. Our study uses a fusion of visible, multispectral and thermal imagery acquired through the low altitude aerial survey for detail classification of land cover and vegetation types. The study area is located in the central part of the Czech Republic and situated in an environmentally specific area – an arboretum of 2.45 ha. Visible (i.e. RGB), multispectral, and thermal sensors were mounted on a flying fixed-wing Unmanned Aerial System. The imagery was acquired at a very detailed scale with Ground Sampling Distance of 3–18 cm. Besides three mosaics (one from each sensor), normalized Digital Surface Models were built from visible and multispectral sensors. Eight classification models were created – each mosaic (visible/multispectral) was enriched with height data, thermal data, and combined height and thermal information. A classification into a three level system was performed through Geographic Object-based Image Analysis using Support Vector Machine algorithm. In general, Overall Accuracy grew with the amount of information entering the classification process. Accuracy reached 77 – 91 % depending on the level of generalization for the best model based on multispectral data and 67 – 80 % for data from the visible sensor. Both thermal data and height information improved the accuracy; however, the statistical evaluation did not reveal any significant difference between the contribution of height and thermal data. Results also indicate that increasing spectral resolution leads to a significantly better performance of the models than higher spatial resolution. UAVs equipped with a proper sensor provide a convenient technology for detail land cover classification even in areas with many similar plant species. [ABSTRACT FROM AUTHOR]