Lidar use for wetlands

Report
Lidar use for
wetlands
Annual MN wetlands conference
January 18, 2012
Lian Rampi
Joseph Knight
Agenda

What is Lidar?

Wetland mapping methods

Conclusions
Lidar 101
What is Lidar?

Light Detection and Ranging is an active remote sensing technology that
uses laser light (laser beams up to 150,000 pulses per second)

Measures properties of scattered light to find range and other
information of a distant target

One of the most accurate, suitable and cost-effective ways to capture
wide-area elevation information (vs. ground survey)
Lidar 101
What is Lidar?

Utilize a laser emitter-receiver scanning unit, a GPS, an inertial
measurement unit (IMU) attached to the scanner, on board computer
and a precise clock

Data is directly processed to produce detailed bare earth DEMs at
vertical accuracies of 0.15 meters to 1 meter

Lidar cannot penetrate fully closed canopies, water, rain, snow and
clouds
All available data is
currently accessible via
anonymous ftp at:
•http://www.mngeo.state.
mn.us/chouse/elevation/li
dar.html
•lidar.dnr.state.mn.us
WETLAND MAPPING
METHODS
Wetland mapping methods
Elevation data only
1) DEM resolution for a Compound Topographic Index (CTI)
Data fusion
2) Combination of CTI, NDVI and soils data
3) Random Forest (RF) Classifier
4) Object based classification
Wetland mapping methods
Elevation data only
1) DEM resolution for a Compound Topographic Index
(CTI)
Wetland mapping methods
Elevation data only
1) DEM resolution for a CTI
 What is the CTI:
 Indicator of potential saturated and unsaturated areas within a
catchment area (e.g. a watershed)
 Function of the Natural log (ln) of the Specific Catchment Area
(As) in m² and the Tangent (tan) of the slope (β) in radians
CTI = ln [(As)/ (Tan (β)]
Wetland mapping methods
Elevation data only
1) DEM resolution for a
CTI

Study area
Wetland mapping methods
Elevation data only
1) DEM resolution for a CTI

Goal: assess the CTI to examine how sensitive this index is
to the spatial resolution of several DEMs while predicting
wetlands







3 m Lidar
9 m Lidar
10m *
12m Lidar
24 m Lidar
30m *
33 m Lidar
*DEMs from the 10 m National Elevation Data and 30 m from USGS
Wetland mapping methods
Elevation data only
1) DEM resolution for a
CTI
Results
Wetland mapping methods
Elevation data only
1) DEM resolution source
Accuracy assessment results
CTI (Threshold: CTI>= median + 1/2 sd)
DEM
%Overall Acc % User. Acc
% Prod. Acc
3m lidar
86
68
85
9m lidar
12m lidar
24m lidar
33m lidar
10 m NED
30 m USGS
88
89
90
90
88
84
72
73
76
77
76
74
88
88
87
86
77
69
Accuracy Assessment using a local reference data (wetland size: from 0.1 acres to 788 acres)
Wetland mapping methods
Elevation data only
1) DEM resolution for a CTI
Accuracy assessment results
Omission
Error
Commission Error
Wetland mapping methods
Data fusion
2) Combination of CTI, Normalized Difference Vegetation
Index (NDVI) and soils data
Wetland mapping methods
Data fusion
2) Combination of CTI, NDVI and soils data

Boolean and arithmetic steps using Spatial Analyst tool from
ArcGIS software

Goal: Investigate the effectiveness of combining CTI, NDVI,
and hydric soils for mapping wetland boundaries

Data sets used:



24m CTI (Lidar)
Hydric Soils
NDVI = (NIR band – RED band ) / (NIR band + RED band)*
* NDVI calculated from the NAIP imagery, 2008
Wetland mapping methods
Data fusion
2) Combination of CTI, NDVI and soils data

Assumption behind NDVI
Wetland mapping methods
Data fusion
2) Combination of CTI, NDVI and soils data
Accuracy assessment results
24m
%Overall
Acc
90
% User.
Acc
76
% Prod.
Acc
87
24m
92
82
86
24m
92
82
89
Acres
Combination
DEM
0.1 to 788
CTI
0.1 to 788 CTI + NDVI + Soils
>= to 1
CTI + NDVI + Soils
Wetland mapping methods
Data fusion
2) Combination of CTI, NDVI and soils data
Results
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier

Goal: investigate the use of the RF classifier for mapping
wetlands using different data types

Study area: a small area of the Big Stone lake park subwatershed in Big Stone County, MN
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier: Study area
Elevation
DEM
365
294
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
Data sets used:
 Lidar DEM, Lidar intensity, Spring 2010(leaf off conditions)
 CTI derived from the 3m lidar DEM
 NAIP imagery 2008, Leaf On aerial imagery
 Hydric Soils *
 Organic Matter *
 Slope
*NRCS SSURGO database
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
 Data Used – Lidar
intensity
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
 Data Used – DEM and Slope (Lidar)
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
 Data used – CTI (Lidar)
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
Random Forest results: Top 10
important variables
CTI
Results
Intensity
Green
band
IR band
DEM
Red
band
Slope
Blue
band
Hydric
Soils
OM
Mean Decrease Gini
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier - Results
Partial dependence on Intensity
Partial dependence on Green band
Partial dependence on CTI
Green band
Intensity
CTI
Partial dependence on IR band
Partial dependence on DEM
IR band
DEM
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
Results
UB (Unconsolidated bottom)
CW (Cultivated wetland)
EM (Emergent wetland)
Wetland mapping methods
Data fusion
3) Random Forest (RF) Classifier
Accuracy assessment results
Classification
Random Forest
Classification
% Overall Acc
% User. Acc
% Prod. Acc
91
94
89
NWI
63
78
39
Wetland mapping methods
Data fusion
4) Object based classification
Wetland mapping methods
Data fusion
4) Object based classification

Goal: Evaluate the performance of an object based classification
for identifying wetlands

Data sets used
 2003, 2008 NAIP leaf on imagery
 2005 NAIP leaf off imagery
 NDVI leaf off 2005 and leaf on 2008
 3 m DEM
 Slope
 CTI 3m
 Thematic lake layer
Wetland mapping methods
Data fusion
4) Object based classification

Pilot study area

The Northeast and Central East area of the city of
Chanhassen

Good representation of the variety of wetland types in
the entire city
Wetland mapping methods
Data fusion
4) Object based classification
 Methodology
1. Image segmentation
2. Hierarchical object-based classification
These objects were classified either as wetlands or
uplands/others :





Urban areas: residential areas, buildings and roads
Lakes
Tree canopy
Agricultural fields
Grasses and bare soils
Wetland mapping methods
Data fusion
4) Object based classification
 Methodology
2) Hierarchical object-based classification based on the following attributes:




Shape
Color
Texture
Object features :



NDVI values
Imagery brightness values
Infrared band & red band mean values reflectance from optical
imagery
Wetland mapping methods
Data fusion
4) Object based classification

Methodology
Main algorithms used:
Image classification
Image object fusion
Morphology operations
Geographic Information System (GIS)-post processing to generalize
objects
Wetland mapping methods
Data fusion
4) Object based classification
Results
OBIA wetland
polygons
Wetland mapping methods
Data fusion
4) Object based classification - Results
North East area, Chanhassen City
Central East area, Chanhassen City
OBIA wetland polygons
Wetland mapping methods
Data fusion
4) Object based classification - Results
North East area, Chanhassen City
Central East area, Chanhassen City
OBIA wetland polygons
Reference data wetlands polygons
Wetland mapping methods
Data fusion
4) Object based classification
Accuracy assessment results
Combinations
CTI
Object-based
Classification
%Overall Acc
89
% User. Acc
75
% Prod. Acc
84
95
87
91
Wetland mapping methods
brief review
Accuracy assessment
Combination
%Overall Acc
% User. Acc
% Prod. Acc
CTI 24 m
90
76
87
92
82
89
Random Forest
Classification
91
94
89
Object-based
Classification
95
87
91
CTI + NDVI + Soils
Boolean and arithmetic
classification
Pros and cons of each method
Pros
CTI
Combination
CTI + Soils +
NDVI
Requires Elevation data
only
Help to solve the
problem of wetlands
topographically suitable
for wetlands because of
the low elevation
Lidar is available for most
part of MN
Open Source program
available for CTI calculation:
Whitebox GAT
Free extensions and
toolbox (TauDEM,
ArcHydro) for ArcGIS 9.3
Cons
Does not work well for
every area in the
landscape with low
elevation
Technical knowledge to
process Lidar data
Soil data and NAIP
aerial imagery (1 m )
available to the public
(no charge)
Combination bring
all layers together
and increase
accuracy of wetland
identification
Require ESRI extension
(Spatial analyst: raster
calculator, reclassify)
Require manual
reclassification steps
Random
Forest
Free Software package
Output graphs of key
variables, Gini index,
confidence maps,
and land
classification
GUI interface of
Random Forest
required same size
resolution and grid
alignment for land cover
classification map
output
Necessary statistical
knowledge and
ability to interpret
results
OBIA with
eCognition
Developer
Allow data fusion of
different type of data
and spatial resolution
Classification of
objects shapes
(groups of
homogeneous pixels)
Allows to add more
elements of image
interpretation beside
spectral characteristics
for classification of
objects
Software
requirement
expensive
CPU storage
requirements for faster
processing
CONCLUSIONS
Conclusion
1)
DEM quality is important for the development of terrain
indices used for mapping wetlands.
2)
LIDAR DEM outperforms 10 m NED & 30 m USGS in
accuracy assessment.
3)
Random forest helped to determine key input variables for
wetland mapping classification and resulted in higher accuracy
for wetland mapping.
Conclusion
4)
Combination of lidar DEM, CTI, aerial imagery and NDVI for
an object based classification performs better with higher
overall accuracy compared to the CTI method.
5) Several factors to keep in mind to decide which method is the
best for wetland mapping.
Acknowledgments
 David Mulla and his research group (UMN)
 Paul Bolstad (UMN)
 Remote Sensing and Geospatial Analysis Laboratory (UMN):
•
Jennifer Corcoran
•
Bryan Tolcser
 Steve Kloiber (MN, DNR)
 Tim Loesch (MN, DNR)
 Carver County
Acknowledgments
 Funding for this project was provided by the Minnesota the
Environment and Natural Resources Trust Fund through the
Department of Natural Resources (MN DNR)
Thank you
for your attention!

similar documents