Multiscale Landforms Classification Based on UAV Datasets

The advance uses of Unmanned Aerial Vehicles (UAV) in geosciences by producing very high spatial resolution Digital Surface Models (DSMs), the various UAV flight altitudes led to different scales DSM. In this paper, we analyzed terrain forms using Topographic Position Index (TPI), landforms extracted by Iwahashi and Pike method and morphometric features of three different spatial resolutions DSM processed from different UAV flights height datasets of the same study area. Topographic Position Index (TPI) is an algorithm for measuring topographic slope positions and to automate landform classifications, Iwahashi and Pike had developed an unsupervised method for classification of Landforms and we have used the techniques developed by Peuker and Douglas, a method classifying terrain surfaces into 7 classes. Landforms extracted from the three indices listed above at the three flight heights of 120, 240 and 360 meters and compared with each other to understand the generalization of different scale and to highlight which landforms are more affected by the scale changes.

The factor of scale plays a very important role in Landform classification different levels of measurement (nominal, ordinal, interval and ratio) this paper will discuss the terrain analysis with the applications of Terrain Position Index (TPI), Iwahashi and Pike index and the morphometric features and their effects on generalization and spatial resolutionat different UAV flights altitudes. Pike et al. (2009) remarked that no digital elevation models derived map is definitive, as the generated parameters differs with algorithms and can vary with resolution and scale.
Landform classification stand out with terrain complexity which necessitated specific methods to quantify its shape and subdivide it into more manageable components (Evans, 1990;Gercek, 2010) which constitutes a central research topic ingeomorphometry (Pike, 2002;Rasemann et al., 2004).
An Arc Map Jenness module GIS software for landforms terrain computations was applied on three different spatial resolutions drone based DSM's for the extractions of Topographic Position Index (TPI), Iwahashi and Pike landforms and the morphometric features at different scales.

Study Area
On the western Lebanese mountainous chain our project location lays at an area about 2 hectares in Zaarour region (Figure 1). The chosen non urbanized mountainous area with a slight natural slope, represented by bare lands with elements of anthropogenic relief. The inclusion of anthropogenic micro-relief in the studying area due not only to the requirements of representativeness, but the presence of complicating microform for the experimental modeling of the terrain concave and convex smoothed areas.

Figure 1. Google Earth Spatio-Image of Lebanon Showing the Study Area
A Dji Phantom 3 UAV, caring a camera of 14 megapixels at a focal length of 3.61 mm used to scan the The three UAV missions have the same flight path designed in a mobile autopilot application called Litchi ( Figure 2). The on screen display of the autopilot shows the flight path, the study area and the flight parameters (coordinates, height, time, etc.). All datasets of the three missions of different flight heights was processed in Agisoft photoscan software for the extraction of Digital Surface Models (DSM).

Material and Methods
Throughout the assessment, we comprehensively used this UAV for aerial images acquisition to the generation and interpretation of Digital Surface Models (DSM) by using new photogrammetry technologies.

Topographic Position Index (TPI), the analysis was performed by DSM's simulation to obtain
Topographic Position Index (TPI). The process of formulae (1) calculate the difference between elevation at a specific cell and the average elevation of the neighborhood surrounding cells (Tagil & Jenness, 2008); describing higher and lower areas for the classification of the terrain into different morphological forms (Jenness, 2005).
The simulation required the radius adjustment of neighborhood and its geometric shape based on two different scales or two sizes (Barka et al., 2011). In this study, a radius between 5 m and 25 m was applied to determine the slope positions.  Iwahashi and Pike had developed a Landforms classification unsupervised method based on only three terrain attributes: slope gradient, surface texture and local convexity (Iwahashi & Pike, 2007).
This method restricts a number of landform classes 8, 12 or 16 with a physical meaning of statistical landscape properties.
The unsupervised approach treats topography as a continuous random surface, especially for the three level of details FA-120, FA-240 and FA-360 independent of any spatial or morphological orderliness imposed by fluvial activity and other geomorphic processes.
Morphometric elements, the standard method for the identifcation morphological elements is to establish a mutually position for the central cell in relation to its neighbors (Peucker & Douglas, 1974;Evans, 1979). The classification algorithm can be done by maintaining the continuity of linear elements, which gives advantages over the method of selection on the basis of logical comparison of neighboring cells (Peucker & Douglas, 1974;Jenson, 1985;Bennett & Armstrong, 1989;Skidmore, 1990;Pogorelov & Doumit, 2009).
Morphological elements take the forms of: Planar, pit, channel (thalweg), pass, ridge (division line), and peak. The names of morphological elements may vary in different sources, but they can be uniquely explaining in terms of changes in the three orthogonal components x, y and z (Wood, J., 1996;Pogorelov & Doumit, 2009).

Results and Discussions
Landform classifications delineated using the TPI method is shown in Figure 4, TPI values present a powerful way to classify the landscape into morphological classes (Jenness, 2005 concave shape, while "Local Ridges or Hills", "Midslope Ridges, Small Hills in Plains" and "Mountain Tops, High Ridges" all tended to have strongly positive curvature values of a convex shape.  The results of Table 2 shows how the area of some morphological elements is increasing against other elements relating to scale variations. In Table 3  Some morphological elements such as Upland drainage type are not found in any of the three maps and other like Local ridges are disappearing with scales variation and constituting a basic for generalization processes.

Element in the Three DSMs Levels Based on TPI Classification
Area ( Open slopes comprised between 6 and 11% of the total area in all flight altitudes while midslope drainages increasing with the flight heights between 9.75% and 11.46% from the total study area.    The concavity and convexity of the very steep slope with fine texture found only in high spatial resolution models (FA-120), the coarse texture of high convexity increasing with the pixel size.
Step and moderate slopes are not detected in all three models, gentle slope coarse texture high and low convexity are increasing with the flight altitude.
Varying DSM spatial resolution can achieve an elements separation of appropriate scale, without the need of generalization.    Table 4 some morph metric features like pit, pass and peak are not detected in all flight altitudes, otherwise planar areas are detected in FA-120 the lower flight altitude at a very low percentage of area in order of 0.00001%, we cannot judge on this result because the value of this pixel could be a processing artifact. The area of channels is increasing with the flight altitude and the ridge area is decreasing against the channel one.
The dominating land forms of surface specific points channel and ridges of the study area form a comparison models of each flight height with TPI land forms. By splitting channels and Ridges of FA-120, FA-240 and FA-360 and exanimating which TPI land forms are included in each type, Table 5 shows the area percentage of each landform.   The diagram of Figure 7 shows the percentage of TPI land forms area in ridges at different scales, the log curves of 120, 240 and 360 have an intersection point at upper slope this point made a transition of values from low percentage to higher percentage of areas.
The correlation value of R 2 between land forms of FA-120 is 0.6 for FA-240 is 0.9 well correlated because of the proportionality and small percentage interval of areas, for FA-360 the correlation value is 0.6 similar to FA-120.  Fa-120 has a low correlation between landforms R 2 = 0.35 even less than the average. We can conclude from these values that due to cartographic generalization and the transition from flight altitude to other, the degree of similarity for channels landforms areas rising with the flight altitude. Hence for ridges land types the area of canyons and midslope, upper slope local ridge, midslope ridge and mountain tops are increasing with flight altitude, upland drainage, u-shaped valley, plain and open slope area is decreasing with the flight heights.

Conclusion
In this study, drone Digital Surface Models (DSM) at diverse flight heights used as input data. By using forms are more affected to generalization than other forms.
The landform classes obtained for the three scales differentiate dynamic terrain characteristics of the study area. Landform classifications extracted form drone DSM and GIS fast the presented results and discussion by integrating the geospatial multiscale approach of terrain analysis.
The result shows that TPI provided a powerful tool for describing topographic attributes of a study area and there is a relationship between landform map and spatial resolution. By deep understanding of the terrain characteristics, potential and specific constraints of cartographic generalization. Information and methods discussed in this paper are valuable results for cartographicmultiscale studies and analysis.
Landforms are dissolving with scales against each other's, some of them gaining areas and some disappeared. This paper analyzed the generalization at three different scales (flight altitude), for future researches we are planning to examine and monitor changes of landforms at micro, local and global scales.