UAV-based vegetation mapping in a Namibian savanna: do remotely-sensed
structural parameters contain indications of the prevailing disturbance regime?
M. J. Mayr 1, S. Malß ², H. Joß ², E. Ofner 1, C. Samimi 1,3
1 Dept. of Geography, ² Dept. of Geosciences, ³ BayCEER, University of Bayreuth, Germany
- accuracy: high
- pair preselection: Reference
- key/tie point limit: 40000/10000
- posterior optimization
Dense cloud creation
- camera trigger
position/attitude
- quality: high
- depth filter: moderate
- telemetry
- position (L1 GPS)
- attitude (IMU)
Digital Elevation
Model (DEM)
- point cloud classification
- max angle =10
- max distance: 1
- cell size = 10
Lens calibration
Mesh creation
Orthophotos
- focal length = 10 mm
- sensor dimension
- distortion coeff.
- surface type: height
- source data: dense cloud
- face count: high
- interpolation: enabled
- posterior hole filling
- clipped to sites
- resolution: 0.1 m
y = 0.74 * x + 1.45
R² = 0.72, adj. R² = 0.7
RMSE = 0.91 m
a)
15
Rule-based classification
- inputs:
Orthophoto, CHM
- scale = [30, 60]
- merge = [15, 60]
- avg CHM > 1m
- GRVI* > 0.0
- roundness = 0.9 - 1.1
d)
4
5
6
7
8
9
max tree height by last fire
c)
5
6
7
measured height [m]
8
9
0
0
20
40
60
80
estimated cover [%]
Fig. 2: Linear models and Root-Mean-Square-Errors (RMSE) of UAV-derived vs. in situ
parameters per site. Whereas height parameters (a-c)) show sophisticated agreement with in
situ parameters, woody cover (d)) is strongly underestimated with the imagery used and the
classification procedure applied.
12
10
8
20
14
max tree height by disturbance regime
6
4
>15 yrs (n=7)
4
3
UAV-derived max tree height [m]
2
Woody cover [%]
per site
* GRVI: Green Red Vegetation Index ((Green-Red)/(Green+Red+eps)).
Not applicable to YellowGB imagery. According to Mothoka et al.
(2010): GRVI > 0.0 discriminates green vegetation.
b)
mean tree height by grazing
8
12
10
<15 yrs (n=12)
1:1 line
linear m.
30
40
y = 0.35 * x - 7.05
R² = 0.41, adj. R² = 0.38
RMSE = 39.19 %
10
UAV−derived cover [%]
7
6
5
4
UAV−derived height [m]
8
9
4
woody cover per site (n=19)
3
1:1 line
linear model
(single class)
14
3
measured height [m]
y = 0.71 * x + 1.55
R² = 0.71, adj. R² = 0.69
RMSE = 1.02 m
2
Based on expert interviews with landowners, residents and
forestry staff, in-field recognition and NASA’s FIRMS (Fire
Information for Resource Management System) database,
the prevailing fire and grazing regimes were assessed.
2
median height per site (n=19)
Image
segmentation
UAV-derived mean tree height [m]
c)
- Structural vegetation parameters:
- Disturbance regime:
1:1 line
linear model
measured height [m]
Fig. 1: The Soleon Coanda x12 on duty (Image: C. Samimi).
Tree and shrub heights (> 1.5m) were measured using a laser
distance meter (Leica DISTO D510) along a regular point
pattern (30m grid). Height measurements (n = 322) were
conducted for all woody individuals covered by an
upward-facing full-frame fisheye photograph (taken at 1m
height). Furthermore, we estimated vegetation type, cover
fractions and strata contribution for each of the 19 sites (size
range: 0.65 - 1.85ha).
10
(ENVI EX 5.1)
8
3
5
- maximum
- mean
- median
Woody cover delineation
6
9
8
mean height per site (n=19)
2
1:1 line
linear model
Tree height parameters per site
UAV-derived tree height parameters (cf. Fig. 3 a-c)) showed good agreement with in
situ measured heights and could thus be used as a surrogate for those in assessing
the influence of disturbances on per-site tree heights. In contrast to heights, woody
covers (cf. Fig. 4 d)) were severly underestimated in the UAV-derived approach.
Maximum tree heights were significantly lower on sites that were burned within the
last 15 yrs and decreased as a function of an intensified disturbance regime (cf. Fig.
3 a+c)). Frequently grazed sites had sigificantly lower mean tree heights (cf. Fig. 4 b)).
7
UAV−derived height [m]
10
15
y = 0.67 * x + 2.61
R² = 0.7, adj. R² = 0.68
RMSE = 1.86 m
5
UAV−derived height [m]
b)
maximum height per site (n=19)
(individual tree identification)
- CHM > 1.5m
- low-pass filter
- seed to saddle > 2m
- min. area > 1m²
(= DSM - DEM)
4. Results
a)
Watershed
segmentation
Canopy Height
Model (CHM, nDSM)
low
(n=4)
moderate
(n=8)
high
(n=7)
7
Image reference
extraction
(R (sp, raster, rgdal, rgeos), QGIS + SAGA GIS plugin))
- clipped to sites
- resolution: 0.1 m
UAV-derived max tree height [m]
Flight logs
Digital Surface
Model (DSM)
Image alignment
- TIFF conversion
- manual image
selection
Tree height delineation
6
- RGB (n = 11)
- YellowGB (n = 8)
- NIR (n = 19;
not used here)
Product creation
5
Image
preprocessing
Imagery
Mosaicked output
4
(n = 19)
- Flight campaign:
19 flights with a Soleon Coanda x12 (cf. Fig.1) were carried
out at the end of the dry season 2015. The UAV was
equipped with two Nikon 1 V3 cameras (VIS and NIR)
mounted on a two-axis gimbal-stabilized platform. The
flights were preferentially undertaken around mid-day and
flown in autonomous waypoint-mode with a 50 % sideward
and forward overlap (two consecutive images via
IR-triggering). A flight altitude of approximately 70m
yielded a ground sampling distance < 2cm.
Structure for Motion
(SfM) (Agisoft Photoscan v1.2.4)
Preprocessing
3
Flight campaign
6
2. Data
We used a Structure for Motion (SfM) approach to generate mosaicked orthophotos, DSMs as well as DEMs derived from point cloud classification. Direct
georeferencing relied on the on-board single-frequency GPS and barometric sensor. Canopy Height Models (CHM), from DSMs and DEMs, were further processed
using watershed segmentation to yield per-site tree heights. To assess woody covers, we applied a rule-based feature extraction approach.
5
Disturbances such as fire and grazing herbivores are one
major feature that drive the coexistence of grass and woody
species in savanna ecosystems (Scholes & Archer (1997),
Wagenseil & Samimi (2007) - among many others). In Namibia‘s
Northern Otjozondjupa Region, land use and tenure
(commerical and communal rangelands, state-protected
forest) cause diverse disturbance regimes, while other
environmental factors affecting plant communities (e.g.
rainfall) remain relatively homogeneous.
Unmanned Aerial Vehicles (UAV) equipped with
consumer-grade cameras are nowadays regarded as serious
remote-sensing systems. We hypothesize that such a system
can reliably assess structural vegetation parameters (height
and cover), which we investigate for their disturbance
impacts.
5. Discussion
3. Methods
4
1. Motivation
regular (n=8)
irregular (n=11)
Fig. 3: Boxplots illustrating the significant effects
of disturbances on site-scale tree height
parameters derived from UAV-imagery. a)
Maximum tree height and last fire occurrence.
Group differences (mean) were significant at the
0.1 level from a Welch Two Sample t-test. b) Mean
tree height and grazing. Group differences were
significant at the 0.1 level from a Welch Two
Sample t-test. c) Maximum tree height and the
combined disturbance regime. A group difference
(mean) was significant at the 0.05 level from an
one-way ANOVA. Disturbance classes correspond
to ‚low‘ (no fires within 15 yrs AND irregular
grazing), ‚moderate‘ (fire within 15 yrs AND/OR
regular grazing) and ‚high‘ (several fires within 15
yrs AND regular grazing).
Despite the absence of high-accuracy reference data
(Ground Control Points/dGPS, high-resolution survey-grade
DEM), the approach applied here was able to produce data
of sufficient geographic accuracy for the aims of this study.
The derived site-scale tree heights were fairly consistent
with in situ measurements. A comparison on individual tree
scale was not possible here, but is hypothesized to yield
lower accuracies. This is due to the point clouds relying on
few matches at the end of the dry season (many species not
yet or only partly with leaf-on canopies), which might partly
explain the retrieved underestimation of tree heights.
Hence, the accuracies can be expected to increase under
fully leaf-on conditions. This also holds true for the
derivation of woody covers, which definitely need further
refinement.
Our study assumes an exclusive direct feedback of the
disturbances fire and grazing on tree/shrub structural
parameters. Clearly, a such is a simplification of the real
world. Soils and nutrient availability, species differences,
but also other factors of the disturbance regime (browsing,
drought, frost, pests) determine site-scale structural tree
parameters. These need to be better accounted for, where
possible, in future studies.
6. Conclusions
Dry-season remote sensing of vegetation with optical
sensors is generally a challenging task. However, the
approach and results presented here are promising. We
found a high agreement (R² ~ 0.7) but also slight
underestimation (RMSE < 1.8m) for each of the site-scale
UAV-derived tree height parameters with their in situ
correspondents. The results for woody cover are not
sufficient in their current state (strong underestimation;
RMSE = 39.19 %; R² = 0.41). However, it needs to be
mentioned that also reliable in-field estimations of woody
cover during dry season are difficult.
Although all sites under investigation were considered
quasi-homogeneous regarding their environmental
framework (flat terrain, rainfall, sandy soils), small-scale
variability of environmental factors cannot be neglected.
Furthermore, species differences an further disturbance
factors were not considered here. This clearly hampers the
generality of our study and the results retrieved.
Contact
manuel.mayr@uni-bayreuth.de
klimatologie.uni-bayreuth.de
References
- Mothoka, T., Nasahara, K.N., Oguma, H., Tsuchida (2010): Applicability of Green-Red Vegetation Index for
Remote Sensing of Vegetation Phenology. Remote Sensing 2, 2369–2387.
- Scholes, R.J., Archer, S.R. (1997): Tree-grass interactions in Savannas. Annual Review of Ecology and
Systematics 28, 517–544.
- Wagenseil, H., Samimi, C. (2007): Woody vegetation cover in Namibian savannahs. A modelling approach
based on remote sensing. Erdkunde 61(4), 325–334.