Cliff vegetation monitoring using close range photogrammetry and UAS: technical issues and practical hints

Cliffs are steep slopes rocky formations, often vertical and sometimes overhanging. They are common on coasts, in mountainous areas, as walls of canyons and glacial valleys or along rivers and are usually formed by rocks that are resistant to erosion and weathering. Even if the total extension of cliff surfaces is not available because steep areas are not fully measurable from common maps, cliffs are widely distributed worldwide. Although cliffs and rocky slopes are quite harsh habitats for plants to grow, they are characterized by a widespread phenomenon: the concentration of a number of endemic and rare plant species larger than in the surrounding flat areas [1] [2] [3]. Although so relevant for biodiversity conservation, studies on plant communities of the cliffs are sporadic, mainly because of site inaccessibility [2]. In such conditions conventional field sampling at the community level (e.g. permanent plots or transects) is challenging even for safety issues and sometimes personnel with rock climbing skills is required [4] [5]. To overcome the inaccessibility of such habitats, remote data collection by means of optical tools (i.e.binoculars, telescopes, telephoto lenses) is often used [6]. Goñi et al. [4] have developed two specific estimate methods to carry out censuses and population demography remote monitoring of rupicolous plant species. The first is based on plant individuals remote counts multiplied by an average Correction Factor (CF = real total number / distant total count); the second is based on scaled panoramic photographs, GIS analysis of total area of occupancy of vegetation patches and estimation of average individual density in a patch. In addition to inaccessibility, verticality of cliff and steep rocky faces poses further problems. As highlighted by Goñi et al. [4] and Gigante et al. [7], the orthogonal projection of vertical surfaces used in conventional maps or orthomosaics brings some shortcomings, like the underestimation of the total area. To overtake such difficulties various attempts are reported.In case of long-term monitoring, the Fixed Point Photography (FPP) or Photo Point Monitoring (PPM) technique was used [8]; it consists in taking multitemporal terrestrial photographs of a site, from a fixed point (x,y,z) and angle, using always the same lens focal length. Moreover, an alternative approach based on oblique aerial photographs, derived from video imagery captured using an helicopter, was proposed by Barron et al. [5]. Within this scenario, the overall aim of our research was to test new protocols and technical solutions for long-term vegetation monitoring on inaccessible cliffs and rocky slopes, taking into account that they all must be realistically practicable in terms of time, staff involved and costs. The herein described study represents a proof of concept showing how close range photogrammetry by means of Unmanned Aerial Systems (UAS), low-cost digital cameras and Structure from Motion (SfM) software can be used to derive high-resolution orthomosaics (orthoplanes) of vertical rocky faces to be used for vegetation and species long-term monitoring. More specifically we report about the effects of various factors on image resolving power and in turn on the usability of such images for vegetation patch mapping and species recognition. Our research was carried out in southern Italy coastal sites classifiable as the EU Habitat 1240 (Vegetated sea cliffs of the Mediterranean coasts with endemic Limonium spp.) and Habitat 8210 (Calcareous rocky slopes with chasmophytic vegetation). Some of the species of these two adiacent habitats are: Crithmum maritimum L., Limonium sp.pl., Limbarda crithmoides (L.) Dumort., Dianthus rupicola Biv., Primula palinuri Petagna, Eokochia saxicola (Guss.) Freitag and G. Kadereit. The time of the year for the aerial survey was choosen taking into account the phenology of the species to be mapped (e.g. flowering or leaf color phenology). We used a DJI "Inspire 1" multicopter with a Zenmuse X3 gimbal camera with a small format Sony sensor (1/2.3" - 12Mpx) and a short focal length lens (f = 3.6 mm). The camera was setup in Auto Exposure mode and ISO 1600. The UAS was manually piloted to evenly spaced hovering points to ensure sufficient side and forward image overlap. Three sets of images were taken parallel to the cliff at about 40m, 20m and 5m; both the first and the second set covered the whole survey area. A fourth set of photos was taken at 20m with the optical axis of the camera no longer perpendicular to the cliff, but facing downwards at about 45° to reduce as much as possible "no-data holes" in the final 3D model and to catch plants growing in rock depressions. Five USAF-1951 Resolving Power targets [9] ranging from group 0 to -5 were evenly distributed at the foot of the cliff to measure the effect of increasing distance on image quality and usability. The entire data set (40m, 20m, 5m and oblique images) in JPEG format was orthorectified and mosaicked using Agisoft Photoscan, a well known SfM software. Then, the orthomosaics were input in a GIS software (QGIS v.2.18) for visual species identification and vector digitizing the perimeter of individual plants or vegetation patches. Moreover, unrectified images taken at each range were examined for resolution power and species recognition possibility. To discuss the final resolving power of the optical system with and without carrier induced vibrations, we performed a second test in a simplified scenario using a SJCam "SJ6 Legend" action camera (1/2.3" - 16Mpx - f = 3 mm) mounted both on a tripod (terrestrial images) and on a WLToys V393 multicopter with vibration damping supports (aerial images). Two USAF-1951 targets were used to measure the final resolving power (RP) at 3m, 4m, 5m and up to 30m with a 5m step. Further, we tested the effect of using two different ISO settings (200 and 1600). The main factor influencing the resolving power of the unrectified images was the camera-to-subject distance. Until 10 meters it was possible to map vegetation patch perimeter and area, but fine details were blurred and species recognition of small plants was not reliable. Species photointerpretation accuracy increased after using training photographs taken at higher resolution by means of a DSLR camera with a lens of medium focal length (e.g. f = 55 mm). Only at 5 meters from the subject (RP 0.177 lp/mm; min. bar width = 2.83mm) details were sufficient for species recognition without any aid. Other factors of varying degree can influence the resolving power of unrectified images taken using a UAS. To reduce their negative effects we suggest: a) to ensure that the propellers are well balanced and the vibration dumpers are in good condition; b) to set the camera at low ISO values (e.g. 200); c) to prefer a camera with a 1" or possibly APS-C sensor and mounting a lens of relatively long focal length (e.g. f = 8.8; d) to avoid low light conditions. Results showed that the resolution power of the images is further degraded by the treatment with the SfM software due to the difficulty of image matching over vegetated areas and to the resampling procedure needed for orthorectification. Image matching can be even more difficult when images are taken in windy conditions because crown and leaves position changes between overlapping images. The overall results confirm that ground sampling distance (GSD) calculated on the base of sensor size, lens focal length and flying height above the terrain (i.e. camera-to-subject distance) is only a first guess of the dimension of the smallest object visible in the final orthomosaic. After our trial surveys we can highlight some shortcomings of UAS cliff vegetation monitoring and suggest possible solutions. The pilot loses the depth perception when the UAS is just a few tens of meters away from him. Consequently, it is unsafe to flight too close to the cliff face without a second observer nearby it or without some technical aids like First Person View (FPV) or obstacle avoidance sensors. In case of manual UAS piloting, to obtain a full coverage of images with the needed forward and side overlap becomes increasingly difficult as the distance from the cliff decreases. A possible solution is to couple the close image data set (e.g. 5m) with another one at longer distance (about 15 meters), finally using both to create the photogrammetric model in the SfM software. If the morphology of the cliff face is not too complex, an automatic flight/shooting plan could be used, but at present only few "Mission Planning" software offer visual 3D planning functionality like the EPFL TOPO Planner [10]. A pre-determined flight plan is even more desirable in case of a multitemporal monitoring project. Reducing the distance from the cliff, the camera field of view (FOV) decreases too, while the total number of images to be analysed in the SfM software increase. This will impact the analysis time and eventually require a more powerful computer hardware. Consequently, the distance should always been chosen in relation to the required task (i.e. community mapping or species recognition) with the aim to minimize the total number of images. The area and perimeter of a given patch of vegetation can change in relation to the chosen orthoprojection plane. Consequently, in case of a long term vegetation monitoring project the position of such a plane must be described. Finally we suggest involving at least two operators: a UAS pilot and a botanist. The latter will locate the areas to be surveyed and, with the help of an FPV system and a second radio control, will shoot photos of plants details and will support the pilot in maintaining a safe distance from the cliff. For the next future we planned to explore the use of multispectral NIR sensors and semi-automated object-oriented image classification to improve cliff vegetation mapping.