The document discusses using LIDAR data and the Normalized Difference Tree Index (NDTI) to extract building footprints from point cloud data. It describes collecting LIDAR data using sensors mounted on aircraft and processing the data to generate nDSM, intensity, and NDTI rasters. Image segmentation and rule-based classification are then used to classify objects in the rasters as buildings or non-buildings. The results achieved an overall accuracy of 94% and kappa value of 0.84, showing high correlation with reference building footprints. Machine learning techniques are recommended for more accurate extraction at smaller scale factors.
2. LIDAR – LIght
Detection And
Ranging
Lidar, which stands for Light
Detection and Ranging, is a
remote sensing method that
uses light in the form of a
pulsed laser to measure ranges
to the Earth.
The data is collected in the form
point cloud data by sensor
scanner mounted on a device.
3. LIDAR components
The main components of
LIDAR system are : -
• Lidar Scanner
• Global Positioning
System
• Inertia Measurement
unit
• Computer System
LIDAR based system –
• Airborne
• Ground based
• Satellite based
5. Objective
The objective of the study is to extract Building Footprints from a LIDAR scanned
data in a complex urban environment using the Normalized Difference Tree
Index(NDTI) .
Accurate building footprint availability is a tough task and for urban planning and
development tasks one needs an accurate building footprint data to better plan
the services in a more realistic multi dimensional scenario.
6. Methodology of the study
Rule based
Object
Oriented
Classification
Lidar Intensity data
nDSM data (FR & LR)
NDTI data
Output
In the study Lidar intensity data, normalized digital surface model(nDSM) of
the first and last returns, and the normalized difference tree index(NDTI)
derived from the two returns are used to extract building footprints from
point cloud data using rule based object oriented classification.
7. Data Collection
The LIDAR data is collected using Optech ALTM 3100 sensor with the help of
Applied Geomatics Research Group.
Data Metrics
1. Altitude – 1000M
2. Scanning Rate – 39Hz
3. Scanning Angle - +-20 Degrees
4. Laser Frequency – 70HZ
5. Site Location – University of Western
Ontario.
Data Acquisition – 15 Flight Lines of width
about 750m and the strip overlap is 50%.
8. Process Flowchart
Raw Lidar
Point Cloud
Data Pre-
processing
Image
Segmentation
Rule Based
Classification
Post
Processing
Building
Footprints
Data Pre-processing
The Digital Surface Model (DSM) which represents the earths surface and includes all objects on it &
Digital Terrain Model(DTM) which represents the elevation of the earths surface with all objects
removed from it.
nDSM – The difference model for DSM and DTM which is a representation of the height of the objects
on the plain surface.
DSM and DTM raster with 0.6m resolution are constructed by interpolation for the 1st and the last
return. nDSM = DSM - DTM
Intensity Raster Layer – Created by Interpolation of Intensity attribute of Lidar data.
NDTI = (nDSM(FR)-nDSM(LR)) / (nDSM(FR)+nDSM(LR)) :-FR – First Return, LR – Last Return
10. Image Segmentation
Segmentation is the method to have image
objects large enough to have meaningful
geometric or spectral values but small
enough so that they do not represent more
than one feature class in the same
classification.
Two Step Approach – Multiresolution
Segmentation & Spectral Difference
Segmentation.
Multiresolution Segmentation – The
method aims to region building within the
data based on a segment scale parameter
which is a threshold for heterogeneity of the
objects. Different sets of nDSM (FR & LR) &
NDTI are used as inputs, different scale
parameters are used to arrive at optimum
result. Scale Parameter 10 is chosen in order
to limit intermixing or excess data.
Comparison of different scale parameters (a)20,
(b)15, (c)10, (d)5.
11. Spectral Difference Segmentation It is
used to merge neighboring objects
on the mean layer intensity values. The
layers used –
nDSM(FR), nDSM(LR), NDTI, Intensity
data – Layer weights 1:1:2:1
(K1,K2,K3,K4)
Normalized layer weights – KN1-
K1/(K1+K2+K3+K4) are calculated.
Expression – 𝑖=1
4
KNi∗(Ki1−Ki2)
Ki1-Ki2 = Difference of the mean
intensity values.
Higher weight for NDTI data is used
separate trees from buildings.
The loop is reiterated multiple times.
nDSM of the last return as a visual reference and
segmentation output.
12. Rule based classification
The classification has three steps -
Assign basic building objects – This step classifies the basic
buildings based on shape and small height difference
between first and last return. NDTI <=0.001 & nDSM(LR)
>=2.5 or nDSM(LR) >=30 Further check for roundness
<=0.41 & area <=190.
Assign Adjacent Building objects – It is used to identify
objects adjacent to buildings. Rel. Border to building is
calculated using the border length shared between two
adjacent buildings and is used to classify them.
Assign building edge objects – The last step is to identify the
building edge objects. Rel. border to building >=0.4.
At each step if the query is true it proceeds as a building
candidate if it is false it puts the object into unclassified
category and furthers to next step.
13. Rule based classification
Fig(a) – In Step(1) the major buildings are classified and
the unclassified buildings in this step are sent for step
2.The highlighted areas are the classified buildings.
Fig(b) – In step(2) the adjacent building based on the
shared boundary criteria of >=0.6 with respect to the
parent object is used and the buildings classified in this
step are darkened further.
Fig(c) – In step(3) the edge building which share even
lower boundary with the parent object >=0.4 and further
the roundness >= 2.8 and nDSM(LR) >=6 is filtered. Due
to penetration of light at the edges of the building
therefore some small building at the edges may be missed
hence a further check with 3.5<=nDSM<=5.3 and rel.
border >=0.5 is used to collect them.
The remaining objects are assigned to the unclassified
class.
14. Post processing & Accuracy assessment
In post processing geometric check using area and
rectangular parameter fit is used to further filter the data and
arrive at the final result.
Accuracy assessment of the data is done using Confusion
matrix algorithm.
A confusion matrix is a table that is often used to describe
the performance of a classification model (or “classifier”) on a
set of test data for which the true values are known.
Kappa Value – It measures the agreement between the
classification and the true values.
Classified
Data/Reference
data
Building Non-Building
Building (d) (c)
Non-Building (e) (f)
User accuracy – d/(c+d), e/(e+f)
Producers accuracy- d/(d+e), c/(c+f)
Kappa -(d + c)*(d + f)–((d + c)*(d + e) + (e + f)*(c +
f))/((c + d + e + f)^2–(d + c)*(d + e)*(e + f)*(c + f))
15. Result
Overall accuracy – 94% ,
Commission error – 6.3%, Kappa-
0.84.
This means that there is a high
correlation with the predicted
data and the actual data.
Fig(1)
Samples
from the
reference
building
footprints,
(2) – Results
from the
anlysis.
Summary
We have a clear sight of moving
ahead of the traditional optical
images format.
The NDTI index has proved to be a
game changer in the analysis and can
be used in various ways to extract
meaningful data.
Different sets of algorithms and
techniques can be utilized to access
the output and object based
classification proved to be a winner.
16. Recommendations
We can make use of machine learning technique for object classification and
segmentation using which we can go down to the scale factor of 5 making the data
more accurate and not compromising on time taken. This can be done using
automated feature extraction from point clouds and is an advance machine
learning technique.
17. References
Congalton, R. G. 1991. “A Review of Assessing the Accuracy of Classifications of Remotely
Sensed Data.” Remote Sensing of Environment 37: 35–46.
Definiens. 2010. Definiens Developer 8.0.1 Reference Book. München: Definiens AG.
El-Ashmawy, N., A. Shaker, and W. Y. Yan. 2011. “Pixel vs Object-Based Image Classification
Techniques for LiDAR Intensity Data.” ISPRS Workshop on Laser Scanning, Elsevier, XXXIII,
Calgary, August 29–31.
Frauman, E., and E. Wolff. 2005. “Segmentation of Very High Spatial Resolutions Satellite
Images in Urban Areas for Segments-Based Classification.” Proceedings of the 3rd
International Symposium Remote Sensing and Data Fusion Over Urban Areas, Tempe, AZ,
March 14–16.