Nature Values Screening Using Object-Based Image Analysis of
Very High Resolution Remote Sensing Data
Aleksi Räsänen*, Anssi Lensu, Markku Kuitunen Environmental Science and Technology
Dept. of Biological and Environmental Science University of Jyväskylä
*[email protected]
GINorden Conference, Turku, June 9, 2011
Outline of the presentation
Background
– General idea
– Different approaches in nature values
screening
Our Approach
– Study areas, data and methods
– Preliminary results – Future work
Photo: Paula Kinanen
Why to screen nature values?
Currently, biodiversity impact assessments are usually
done late in land use planning processes
– Little changes can be made to plans
– Nature values are mostly assessed with rigorous field work which is expensive and time consuming
Nature values can be
screened with GIS methods and remote sensing data
– Provides preliminary information about nature
Photo: Paula Kinanen
Habitats or species as surrogates for biodiversity
– Habitat/biotope classification & ranking
• Manual RS interpretation
• Automated classification
• Habitats ranked, e.g. by the potential number and rarity of species
– Habitat modelling for different species
• Known habitat preferences
• E.g. with LiDAR & other RS data
– Known existence of species -> modelling & extrapolation
Biodiversity / plant-species richness assessment with RS data
– E.g. NDVI, spectral heterogeneity
(list is not exhaustive!)
Nomenclature: VHR and OBIA
Very high spatial resolution (VHR) remote sensing (RS) data
– Pixel size less than 10 m
– e.g. aerial photos, new satellite images (Quickbird, WorldView), laser scanning data
– Pixel based classification would result in salt-and-pepper- effect
Object based image analysis (OBIA)
– Objects instead of pixels are analyzed and classified – Objects are distinguished with segmentation
• Should mimic human perception of objects
– Object oriented and object based at the same time
Our approach
Based on habitat ranking (Rossi & Kuitunen 1996)
– Goal is to update the method to GIS & RS era and to improve it
Two goals
– Automated habitat classification & ranking
• Ranked by potential number (& rarity) of species
– Mapping valuable nature spots
• Possible locations of spots
• May predict more spots than actually exist
Pulmonaria obscura, photo: Tuomo Kuitunen
Study areas
Taka-Keljo-Isolahti in Jyväskylä and Muurame municipalities, Central Finland
– 70 km
2– Diverse, includes old forest conservation areas
Luopioinen, Eastern part of Pälkäne municipality, Tampere Region
– 400 km
2– Diverse area with many herb rich forest patches, lakes and calcareous rocks
– Presence/absence data of 700 plant species from 1 km
2quadrats collected by Tuomo Kuitunen
Taka-Keljo-Isolahti (N 62˚1, E 25˚4)
Image: World View 2, 4.7.2010, ©Digital Globe2010
Luopioinen (N 61˚2, E 24˚4)
Image: IMAGE2000, ©EC JRC
Data
Remote sensing data
– WorldView2 satellite imagery (Taka-Keljo only)
• 8 bands, resolution 2 m, year 2010
– Aerial images (TK & part of Luopioinen)
• 4 bands, resolution 20 cm(TK) or 50 cm (L), 2007 and 2010
– LiDAR (TK & part of L)
– Landsat & Image2000 (25 m resolution) from early 2000
GIS datasets
– NLS Topographic Database – Soil and bedrock maps
– CLC 2006, CLC 2000
– National forest inventory data – Ground water maps
Data for supervised classification & accuracy assessment
– Forest inventory & compartment data
– Own field work data of habitat classes
Methods / analysis steps
1. Segmentation step
• Both region growing & watershed delineation tested
• WV2 / aerial image data used 2. Evaluation of segmentation
3. Classification
• Several variables derived from data
• Different supervised methods tested: classification trees, automated analysis (e.g. random forests)
4. Uncertainty assessment
Not in strict order, e.g. some classification can be
done before segmentation
Segmentation
©Digital Globe 2010
Evaluation of segmentation
• Different methods and parameter choices tested
©Digital Globe 2010
used in classification
Satellite & aerial imagery
– Spectral variables (e.g. mean reflectance, standard deviation, NDVI)
– Textural (GLCM) variables (e.g. entropy, contrast)
– Frequency variables (e.g. local Fourier transform, wavelets)
LiDAR
– (Micro)topography/geomorphometry (e.g. slope, aspect) – Tree stand structural characteristics (e.g. height)
GIS data
– Soil, bedrock
– NLS topographic database classification
– (NFI classification)
(approx. 30 classes)
Forests
– Herb-rich, mesic/moist, dry; all in 4 successional stages – Springs, rocky areas
Mires / peatlands
– Open, pine, spruce; all drained / not drained
Meadows
– Mesic/wet, dry
Riparian habitats, flooded areas, beaches
Waters
– Oligotrophic lakes, eutrophic lakes, streams and rivers
Fields, roadsides, gardens
Classification results
©Digital Globe 2010
Uncertainty assessment
Variation between and within classes assessed
– Gives information about
uncertainty and classification accuracy
– The class of some segments is certain and for some not so certain
Classification accuracy
evaluated also with the help of field work
– Part of the field work data left to this purpose
In image:
– Tone: class
– Dark areas: more certain
– Light areas: less certain
Preliminary results: habitat ranking
Highest rank is given to habitats with high
diversity or high number of rare species
In image: red=low value, green=high
Connectivity and
complementarity are taken into account in future work
©Digital Globe 2010
nature spots
Different possibilities, e.g.
– Small segments classified to different habitat than surrounding segments
– Segments that have some distinct features
Only probable locations
– Must be verified with the help of field work
©Jyväskylä municipality 2007
Problems and issues to consider
How to classify forest habitats in Finland in RS era?
– Cajanderian classification is based on field layer
– Classification only roughly to pine, spruce & deciduous forest?
• Ancillary data is valuable in more precise classification
How to weigh current value vs. potential value?
– Problems with successional stage and human impact
Automated analysis vs. map evaluation by an expert
Mapping on several scales needed
– Both large areas and ecological networks as well as small patches are valuable
– How to compare the value of large areas vs. small hot spots?
Some field work or at least general knowledge of the area is always needed!
– Automated analysis can reduce the amount of field work or can
be used in screening/scoping phase
Issues to consider in future work
Landscape ecological knowledge
– Landscape should be considered as a whole instead of looking into separate valuable spots
– Connectivity analysis, landscape pattern analysis
What are nature values after all and why to conserve nature?
– Ecosystem services, biodiversity and rarity, naturalness, scenic values etc.
– Partly contradictory goals, different areas can be valuable due to different reasons
Linking work to practical and real world land use planning problems
– Societal, ecological, economical aspects
Thank you!
Acknowledgements
– Maj and Tor Nessling Foundation – University of Jyväskylä
– Finnish Doctoral Program in Environmental Science and Technology (EnSTe)
photo: Tuomo Kuitunen