UAV based hyperspectral grassland monitoring in an alpine shallow erosion area : lessons learnt from classifying vegetation indicating shallow erosion risk

Abstract

UAV based hyperspectral grassland monitoring in an alpine shallow
erosion area
Recent research in the Alps found that a reduction in grassland management is correlated to
an increase in a certain type of shallow erosion areas called blaiken. This change also entails
changes to the dominant grassland vegetation. New developments in hyperspectral technology have produced cameras which are sufficiently light for Unmanned Aerial Vehicles
(UAVs). This thesis explores whether UAV mounted hyperspectral cameras (RIKOLA
(SENOP Optronics, Lievestuore, Finnland)) can detect the spectral signatures of managed
vs. blaiken-related unmanaged grasslands accurately, and whether their signatures are so
distinct as to enable grassland classification on an aerial image. Accurate mapping of Alpine
grasslands can guide erosion mitigation measures to manage grassland types that are more
susceptible to erosion.
A field study was undertaken at a blaiken hotspot at the Schluter lodge, located in the Dolomite ¨
mountain range within the Italian Alps, in order to evaluate the accuracy of the RIKOLA camera.
Its spectral signatures taken on the ground were compared to those of a high precision spectroradiometer (Spectra Vista Corporation, Poughkeepsie, USA). The study also evaluates whether the
spectral signatures of different grassland types were sufficiently unique and the quality of the
Rikola images taken from the UAV system was high enough to permit the characterization of
grassland types by a maximum likelihood classification algorithm.
The spectral separability analysis demonstrated that the spectral signatures of designated grassland
classes separate in Rikola images taken on the ground. However, an orthomosaic created from aerial UAV-Rikola images displayed low precision and accuracy. These errors stem from unstable
lighting conditions throughout the flight and the NIR sensor malfunction. The malfunction caused
the NIR bands not to be recorded, a wavelength range which was shown to be highly relevant when
separating grassland communities. A first classification of the grassland classes within the orthomosaic has a low accuracy, which is in part due to botanists’ class definitions, that were unsuitable
for spectral classification, as well as changing light conditions and the sensor failure. Nevertheless,
when inspecting the spectrometer data and the orthomosaic image bands, the grassland surrounding the blaiken was found to have a distinct species composition and a distinct spectral signature,
as shown by the spectral signature evaluation and spectral separability test. A better classification
for identifying grassland with higher blaiken risk was found to be obtained by using a different selection of classes. The low accuracy of that map suggested that the Rikola UAV data does not fit the
requirements to classify grasslands from a botanical point of view, despite the quality limitations
of the UAV data aforementioned. Thus, the grassland classes should be based on prior spectral
image analysis. Better weather conditions (overcast sky) during the flight should also enhance the
classification as it reduces the influence of shadow artefacts in the image. Furthermore, species
richness and NDRE (Normalised Difference Red Edge Index) values were found to be potentially
connected to blaiken and erosion risk.

Popular Abstract

Hyperspectral grassland and erosion monitoring by UAV in the
Italian Alps
Recent research in the Alps finds that decreased management of grasslands is correlated to
increased soil erosion. These developments also cause changes to the dominant grassland
vegetation. Different types of vegetation have been shown to be distinguishable by hyperspectral cameras. However, these cameras, mounted to aeroplanes and satellites, are expensive, inflexible in their use and have a relatively low ground resolution. Newly developed,
cheaper, miniaturized hyperspectral cameras, have a more detailed ground resolution and
can be deployed more flexible when mounted to Unmanned Aerial Vehicles (UAVs). This
thesis explores whether these, UAV based, cameras can be used to detect differences between
highly similar grassland vegetation of managed and unmanaged grasslands. Classification
maps derived from hyperspectral images can be used for the mapping of alpine grasslands
types and can support erosion mitigation projects by locating areas with vegetation which is
more susceptible to erosion.
A field study was undertaken at an erosion hotspot near the Schluter lodge, located in the Dolomite ¨
mountain range within the Italian Alps, in order to evaluate the accuracy of the UAV based hyperspectral camera. Measurements taken on the ground were compared to those of a high precision
reference measuring device (Spectrometer). The study also evaluated whether the image quality
of such an approach is high enough, and the grassland types are separable when using a image
classification algorithm. Analysis of the hyperspectral data shows that grassland types could be
distinguished by the new lightweight hyperspectral camera. However, the aerial image created
by stitching together UAV images (orthomosaic) displayed unsatisfactory precision and accuracy.
These errors stem from a partial sensor malfunction and unstable lighting conditions throughout
the flight. The malfunction caused the Near Infra Red NIR wavelength to be recorded incorrectly,
which is highly relevant when analysing vegetation. The resulting classification of the grassland
classes was of low accuracy. This is partly due to the low precision orthomosaic as well as due to
class definitions, that were not suited for image classification. Nevertheless, when inspecting the
image data, the grassland vegetation surrounding the erosion areas, was found to display a distinct
pattern. This indicates that a better classification can be obtained by using a different approach to
selecting classes. Identifying grassland with a higher erosion is possible by selecting classes based
on a prior analysis of the orthomosaic. Furthermore, it was found that the vegetation around the
erosion spots displays less vegetation species diversity compared to areas without erosion. This
potential connection between species diversity and erosion risk requires further research. This research could be aided by aerial images, which have previously been used to detect different levels
of species diversity. The research shows the potential of using UAV based camera classification
methods for mapping grasslands in the Alps. The method is complicated and not without challenges but it is cheaper and faster than manual surveying of alpine areas. The increase in number
and intensity of extreme weather events due to climate change will likely also increase the amount
of erosion and landslides in the Alps. Therefore, it is important to reduce, limit and prevent alpine
erosion, as it causes damage to the surrounding landscape, settlements and also puts the local
population at risk. Locating grassland areas with high erosion risk helps to focus prevention and
mitigation efforts on the most relevant areas and is therefore of great value.

• author: Peitz, John
• supervisor: Virginia Garcia; Abraham Mejia-Aguilar
• organization: Dept of Physical Geography and Ecosystem Science
• course: NGEM01 20182
• year: 2019
• type: H2 - Master's Degree (Two Years)
• subject: Earth and Environmental Sciences
• keywords: Physical Geography, Ecosystem Analysis, Shallow Erosion, Blaiken, UAV, Hyperspectral Imaging, Alps, Grassland, Spectral Signatures, Geomatics
• publication/series: Student thesis series INES
• report number: 473
• language: English
• additional info: External supervisor: Dr. Abraham Mejia Aguilar, EURAC, Scientific Research Institution in Bolzano, Italy
• id: 8978657
• date added to LUP: 2019-06-13 14:07:22
• date last changed: 2019-06-13 14:07:22