p. scholefield
TRANSCRIPT
Vegetation Structure From a Fixed Wing UAV System Over Peatland Feature Types
Paul Scholefield, Emma Tebbs, Clare Rowland, Dan Morton, Chris Evans, (CEH) Barry Rawlins, Stephen Grebby, Colm Jordan, Martin Hurst (BGS)
Why are peatlands important?
• Blanket bog is a globally rare peatland habitat • 10-15% of the world’s blanket bogs are found in
the UK • Blanket bogs provide valuable ecosystem services
• biodiversity, • carbon storage • hydrological regula4on • poten4al for greenhouse gas sequestra4on
• Despite only covering 3% of the land area of Wales, deep peat soils are estimated to contain approximately 30% of the countries total soil C carbon stock.
Where?
• Welsh blanket bogs • Moorhouse NNR, Pennines, County Durham • Forest of Bowland, Lancashire • Cheshire Meres and Shropshire Wetlands • Lowland bogs • Scottish systems • A significant number of Countryside Survey
sites
Deliverables • A ‘definitive’ map of the full extent of peat (> 40 cm) in Wales • Classification of lowland and upland peat areas into broad
land-use categories • A detailed assessment of upland blanket bog condition using a
combination of lidar and aerial photography (RGB and CIR) data
• Development of UAV approaches for assessing habitat extents and condition.
• An assessment of lowland peat condition based on detailed land-cover data
• An assessment of utility of UAVs for habitat assessment in remote or “difficult” locations
• Use aerial photography to provide an improved condition assessment for lowland fens and raised bogs
Why use UAVs for field survey? • Fields surveyors are
always required for ground truth data.
• But for difficult conditions and for extensive bog habitats, other approaches may have a reduced impact
• For mosaiced habitats features can be missed at ground level due to issues with line of site
• Field work is tiring in wet conditions
UAV systems - software and tools
• Quest UAV 300 • Canon 6D, inbuilt GPS • Canon 450D, NIR • Lumix LX5 • Intervalometer
• Agisoft Photoscan, MicMac, • Grid.Flightmanager
• ESRI ArcGIS
• ENVI/IDL/PCI Geomatica
Image Classification Approach
• Random forests classifier • R Script • Inputs:
• Spectral data
• Continuous or categorical
• DEM
• Slope • Outputs:
• A classified geotiff
• A class probability layer
Abbeystead Fell
• Useful training site. • Grouse moor, and part of the
Grosvenor Estate (Abbeystead Estate).
• First test of system using the LX5.
Abbeystead Fell
• Useful training site. • Grouse moor, and part of the
Grosvenor Estate (Abbeystead Estate).
• First test of system using the LX5.
Abbeystead Fell
• Good image resolution following Photoscan processing.
• Drainage lines visible • 0.07cm per pixel • Successful DEM generation • Should be able to extract
slope characteristics • Possible to yield patch density
metrics
Abbeystead Fell
• Good image resolution following Photoscan processing.
• Drainage lines visible • 0.07cm per pixel • Successful DEM generation • Should be able to extract
slope characteristics • Possible to yield patch density
metrics • Ideal data for the random
forest model.
Moorhouse NNR
• CEH Long term monitoring site, Environmental Change Network
• Moor House-Upper Teesdale, is a nature reserve in the Pennine hills of northern England.
• Large parts of it are upland blanket peatland
Initial Flight
• Initial classification on imagery collected in May.
• Camera trigger failed, but some imagery collected.
• A remote site. • Needed more batteries. • Needed more SD cards
Preliminary classification
• Test area • 7 classes generated • Showed good
matches with ground survey data
Moorhouse – Training data
• Vegetation Survey • 1960s
2 Flights Completed June 2014
Results – Orthorectified Geotiff
• LX5 camera
• 700 images • Approx 1
sq km
DSM – 0.05m resolution
Slope and Aspect
• Slope and aspect were prepared for the random forest classification. Classification processing for this 1 sq km takes 4 hours.
Catchment detail – 7 classes
Catchment detail – 11 classes
Hi-Res PGA data – Infoterra – 0.2m Res.
• Artle Garth Beck
• 0.5 sq km
Test Flight - 0.05m resolution
• Artle Garth Beck
• 0.5 sq km
PCI Geomatica using NIR and 1m LIDAR
GMEP, Wales. BGS have es4mated the magnitude (index) of local peat drainage associated with the ditches, taking account of ditch density, and their orienta4on rela4ve to local slope. Used BGS 1m LIDAR data, NIR and RGB imagery, and the automated linear feature extrac4on.
Lessons learned
Don’t crash into the tops of 100ft ash trees. As a tree surgeon costs £75.
Learn to crash in safe places
Try to avoid golfers
Conclusions and Next Steps
• Image capture of Countryside Survey sites is feasible with UAVs but open and remote areas are best.
• Get qualified to cover a 1 sq km area. • Don’t underestimate how much time image stitching
consumes • Good classification results can be obtained relatively
quickly using open source software. • PCI Geomatica is good for extracting linear features Next Steps : More flights, more habitats, attempt to extract
canopy structure for hedges, trees, and heathlands for improved classification
Don’t always need a UAV…
300 images taken at 2 second intervals using a Canon 6D with GPS Sub-‐cm resolu4on