University of Twente Student Theses


Using statistics to optimise the detection of collapsed building from laser scanner data

Fathi, Seyed Abdol Majid (2011) Using statistics to optimise the detection of collapsed building from laser scanner data.

[img] PDF
Abstract:In the present study, a method for the detection of collapsed buildings from post event Lidar data is presented. Strong earthquakes require extensive and immediate field investigation to record damage patterns. The investigations of the collapsed building and its spatial distribution after an earthquake are of primary importance for planning the rescue activities and for evaluating the level of damages in affected area. Effective disaster management requires real-time data to various decision makers. Airborne Lidar scanner (ALS) as a state- of -the art technique is capable of delivering large amounts and very accurate point clouds of our interested area in a relatively short time. ALS data is a suitable technique as a basis for damage analysis because it can be acquired directly after a disaster, independent of weather conditions and during days and nights. However, considering the amount of captured data, the automatic detection and interpretation of ALS data remains a challenge to several scientists in the field. Up to the present, a wide variety of algorithms for processing of ALS data has been already introduced and developed and nowadays, the extraction of classes such as buildings, vegetation, etc. is interesting for many applications in Geomatics. One approach for classifying ALS data is to employ machine learning techniques, for which many statistical methods and tools are applicable. This research has been conducted to assess the capability of maximum entropy (Maxent) approach to automate the detection of collapsed buildings from ALS point clouds after an earthquake. Maxent can be considered as a new method for one-class classification. The output of Maxent is the probability distributions of the introduced class. Post event Lidar data of Haiti with a density of 2 points per square meter was used after segmentation of the planar surface in a region growing algorithm, to calculate some features as input predictors for our classification. In this study, 281 collapsed building records have been used to train and evaluate the classifier. The classification for 8516 points was done using Maxent. The importance and contribution of each and every variable was calculated by Jacknife test. The model was evaluated using one threshold independent technique, the area under receiver operating characteristic curve (AUC), and two threshold-dependent techniques, Kappa and true scale statistics (TSS). According to the results, the most important variable is the number of points per segment, which suggests the size of a segment contains useful information. The results also showed that the ratio of unsegmented points to the segmented points, and the distance to DTM were the second and third important variables, respectively. The average behaviour of Maxent in 30 bootstrap simulations using all features revealed that some features (e.g. Density of points in 2D, density of points in 3D and residuals to planarity) had the least predictive power. The evaluation of Maxent suggests that this technique can be considered as a fairly accurate model to detect the collapsed building in a one-class classification problem. Keywords: Maximum entropy, Maxent, classification, Lidar, collapsed building
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page