MAPPING OF HIGHLY HETEROGENEOUS URBAN STRUCTURE TYPE FOR FLOOD VULNERABILITY ASSESSMENT

Vulnerability plays an important role in risk assessment. For flood vulnerability assessment, the map and characteristics of elementsat-risk at different scales are strongly required depending on the risk and vulnerability assessment requirements. This study proposes a methodology to classify urban structure type by combining object-based image classification and different high resolution remote sensing data. In this study, a high resolution satellite image and LiDAR have been acquired over Kota Bharu, Kelantan which consists of highly heterogeneous urban structure type (UST) classes. The first stage is data pre-processing that includes orthorectification and pansharpening of Geoeye satellite image, image resampling for normalised Digital Surface Model (nDSM) and followed by image segmentation for creating meaningful objects. The second stage comprises of derivation of image features, generation of training and testing datasets, and classification of UST. The classification was based on three types of machine learning classifiers, i.e. Random Forest (RF), Support Vector Machine (SVM) and Classification and Regression Tree (CART). The results obtained from the classification processes were compared using individual omission and commission error, overcall accuracy and Kappa coefficient. The results show that Random Forest classifier with all image features achieved the highest overall accuracy (93.5%) and Kappa coefficient (0.94). This is followed by CART classifier with overall accuracy of 93.7% and Kappa coefficient of 0.92. Finally, SVM classifier produced the lowest overall accuracy and Kappa coefficient with 88.6% and 0.86, respectively. The UST classification result can be further used to assist detailed building characterisation for large scale flood vulnerability assessment. * Corresponding author The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W9, 2018 International Conference on Geomatics and Geospatial Technology (GGT 2018), 3–5 September 2018, Kuala Lumpur, Malaysia This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-4-W9-229-2018 | © Authors 2018. CC BY 4.0 License. 229


INTRODUCTION 1.1 Introduction
In the context of flood management, flood risk assessment have widely been implemented to replace the traditional flood management.Details about the flood risk assessment can be referred to EU Flood Directive (2007/60/EC).Its aim is to reduce and manage the risks that floods pose to human health, the environment, cultural heritage and economic activity.The Directive requires Member States to first carry out a preliminary assessment by 2011 to identify the river basins and associated coastal areas at risk of flooding.For such zones they would then need to draw up flood risk maps by 2013 and establish flood risk management plans focused on prevention, protection and preparedness by 2015.
According to Penning-Rowsell et al. (2005) and Foudi and Oses-Eraso (2014), the method used to assess flood-risk consists of four steps: (i) hazard assessment; (ii) exposure assessment; (iii) vulnerability assessment; and (iv) risk assessment.Among these four components of flood risk assessment, remote sensing imagery have extensively contributed in hazard assessment to achieve different purposes and scales.For instance, satellite-based rainfall estimates for estimating maximum discharge.A number of studies have been carried out using various satellite-based rainfall estimates (TRMM, CMORPH and PERSIANN) in which results show that estimates to generate streamflow as reliable as using rain gauge data, provided those estimates have to be calibrated (Stisen, S. and Sandholt, I. 2010;Meng et al. 2014;Zubieta et al. 2015;Jiang et al. 2010).In addition to estimate maximum discharge, producing flooded area is of great importance in flood hazard for which different resolution of Digital Elevation Model (DEM) is created using remote sensing technology.LiDAR data with 1 meter high resolution, higher accuracy of flood map can be produced using 1D hydrodynamic model or 2D hydrodynamic model, Such studies have been carried out by Costabile et al. (2015), McDougall andTemple-Watts (2012), andTurner et al.( 2013).
For vulnerability assessment, remote sensing data often to provide some biophysical parameters for mapping of flood vulnerability map in qualitative way using multi-criteria methods (e.g.Analytic hierarchy Process, AHP) (Rimba et al. 2017;Ouma and Tateishi, 2014;Sambah and Miura, 2014) or vulnerability index (Kumar and Kunte, 2012;Islam etal. 2016).In this paper, we adopted the definition of vulnerability by Merz et al. (2007) where vulnerability is composed of two elements, exposure and susceptibility.However, less study focus on exposure assessment using remote sensing data (Gerl et al., 2014;Angela et al. 2013).According to Merz et al. (2007), exposure can be quantified by the number of the value of elements which are at risk.The term elements at risk includes all elements of the human system, the built environment and the natural environment that area at risk of flooding in a given area.In this study, we focus elements at risk on built environment which can be described by land use land cover.
Land use land cover map play an important role in flood risk assessment.Land use land cover maps describe the spatial location of variety type of elements at risk in the flooded area.However, existing coarser land use classes like residential buildings, industry, and agriculture are often obtained from agency and integrated in flood loss modelling (Foudi et al 2015;Gain et al. 2015;Budiyono et al. 2015;Cammerer et al. 2013).Some studies focused on micro-scale, however, most of the adopted land use data is relatively generalized (Velasco et al. 2015).In Malaysia, Department of Irrigation and Drainage (DID) produced a guidelines for assessment of flood damage in which the adopted land use information in this guideline is rather generalized (report: Guidelines and Procedures for the Assessment of Flood Damages in Malaysia, 2003).Therefore, it is timely to update the detailed land use land cover information to assess the current flood risk situation in Malaysia and subsequently it facilitate related-agency to reduce flood damage.With the advent of remote sensing technology with the aid of image processing and classification method, such objective can be easily achieved, referred to (Manandhar et al. 2009;Adam et al. 2013;Yousefi et al. 2015;Mahmon et al. 2015).
A more detailed land use land cover is refer to classification of Urban Structure Type (UST).Urban structure types is defined as spatial units of the built environment at an aggregated neighbourhood scale, which are relatively homogeneous with respect to the overall population in terms of their physical appearance (land cover) and usage (land use) (Wieland et al., 2016).According to Wieland et al. (2016), characterization and analysis of an exposed building on the basis of its susceptibility attributes and combine with the output of hazard analysis is used to compute the expected damage.Several studies exist that use high resolution image with object-based image analysis for the classification of UST where these studies mostly at Germany in which the urban structure type is well planning (voltersen et al. 2014;Puissant et al. 2012;Bochow et al. 2010;Walde et al. 2014).However, Bochow et al. (2010) tried to classify slum areas as one of the urban structure types using Maximum Likelihood as classifier.Result shows highest error in term of commission error and omission error.
Based on previous studies there are still lack of effort to classify highly heterogeneous UST in developing countries using high resolution remote sensing data.This has definitely complicate the process of flood vulnerability and risk assessment, especially in developing countries where detailed elements-atrisk characteristic should be defined prior to vulnerability assessment.The aim of this study is to compare different methods of machine learning classifier in producing urban structure type map using combining high resolution satellite image and airborne LiDAR data.

Description of Study area
The city of Kota Bharu, Kelantan is located in north eastern part of Peninsular Malaysia where its neighbour country is Thailand.The Kota Bharu is the state capital and has become the central point for Kelantan's administrative and business activities.The landscape of Kota Bharu is different to other cities in Malaysia in which surrounded by a number of Malay villages also known as kampung.According to Fee (1998), the development of kampung does not comply any plan, therefore, the spatial fabric of kampung is highly heterogeneity.The studied area shown in Figure 1 is part of city of Kota Bharu.Traditional houses (Kampung house) at Kota Bharu are mostly elevated wooden houses.Nowadays, most of these traditional houses are converted into different type of houses (Table 1).
The topography of Kota Bharu city is relatively flat with average elevation of 5 m, please refer to Figure 2a.Every year, the Northeast Monsoon from November to January brought heavy rainfall to Kota Bharu which will cause flooding.Due to this reason, the traditional houses at this city are elevated house.

Description of Data
There are two main data used in this study, i.e.GeoEye-1 (GE1) satellite image, and airborne Lidar data.Geoeye-1 satellite image was captured on June 6, 2016.The GE1 data come with four, 2.0 m spatial resolution, multispectral bands (Blue, Green, Red, and NIR) and one, 0.5 m resolution, panchromatic band.The complete LiDAR data (DSM and DEM) with 2 m spatial resolution was obtained in 2008.Both dataset have 8 years difference, however, due to slow development in Kota Bharu, the LiDAR data is still applicable to be used in this study.

Method
In this section, the flowchart of this study is divided into two parts, i.e. data pre-processing and classification of UST with assess its reliability (Figure 3).

Accuracy Assessment
Overall accuracy and Kappa  2010) was used to estimate the suitable scale parameter for multiresolution image segmentation.The scale parameter estimated by ESP is 116 and associated with other segmentation parameters, i.e. shape (0.9) and compactness (0.1) are well used for creating meaningful objects in this study.

UST mapping:
In this step, the classification of UST is important.As there are no compulsory and consistent urban structure types, the classification scheme depends on the study area and the objective of the respective analysis (Banzhaf & Höfer, 2008).Therefore, we defined 8 USTs which shown in Figure 4.  2).In addition, three different machine learning classifiers, i.e.Random Forest (RF), Support Vector Machines (SVM) and Classification and Regression Tree (CART), were used to implement UST classification.However, present study is an preliminary study on the classification of UST, therefore, the parameter of each classifier is set to default.
Both CART and Random Forest are tree-based algorithm.
CART is a single tree of classification model in which we can prune the overgrown trees while RF is an ensemble of number of CART.Also, pruning is not allowed in it thought some tuning function is available in which we can pass parameter like no of trees etc. SVM is a non-parametric algorithm that allow fast and dependable classification with limited amount of data.
Following an object-based approach to image analysis, segments have been used as sample type.In this study, total of 460 segments were generated.Total of 36% of the segments are assigned as training data sets while the rest are testing data sets (295 segments).Training sets data are used for classification purposes while testing sets data are used to evaluate the accuracy of classification result.
Next, accuracy assessment of image classificaiton is carried out by calculation of overall accuracy and Kappa coefficient which are the standard and common method to assess the reliability of the produced thematic map.According to Thomlinson et al.
(1999) accurate a classification when the OA is at least equal to 85% and no class is less than 70%.

RESULTS AND DISCUSSION
This section briefly explained on the results of UST classification shown in Figure 5 and discuss the performance of every machine learning classifier, the producer's accuracy and user's accuracy of every class and also the influence of the image features used for classification.Classification of UST was conducted using three different machine learning classifiers, which were CART, RF and SVM.
The classification accuracy statistics are summarized in Figure 4.The results show that classification using RF3 provided the highest overall accuracy (95.2%) and Kappa coefficient (0.94).SVM generated the least accurate classification map with 78.9% overall accuracy and Kappa coefficient of 0.74.Moreover, CART produced relatively high overall accuracy and Kappa coefficient with above than 90 % and 0.85, respectively.The reason SVM produced least overall accuracy compared to other two classifiers is attributed to the parameter of SVM is set to default.
For the RF3, the classes with the highest producer's accuracy were those of water (100%), tree (100%), planned house (100%), impervious (100%), and grassland (100%) followed by unplanned house (97.8%), bare land (94.7%) and buildings (83.6%).User's accuracy was higher for water (100%), Tree (100%), planned house (100%), impervious (100%), and grassland (100%), followed by buildings (98.6%), bare land (96.6%), and unplanned house (79.1%).All classes were easily separable by all classifier algorithms applied.In addition, we divided the 74 sub features into four groups (Table 2) in order to examine their influence on classification.Apparent result of classification using RF classifier depicted that taking more sub-features into consideration, the overall accuracy and Kappa coefficient is increasing.Similar condition is noticeable in using SVM classifier.On the contrary, the overall accuracy and Kappa coefficient of CART classifier is decreasing with increase the numbers of sub-features.

Figure
Figure 1.Study area

1
Single storey concrete house with tile roofing 2 Single storey concrete house with metal roofing 3 Modified single storey conceret house (partial second floor additions -concrete wall) with tite roofing 4 Modified single storey conceret house (partial second floor additions -concrete wall) with metal roofing 5 Modified single storey conceret house (partial second floor additions -wooden wall) with tite roofing 6 Elevated wooden house with metal roof (zinc) 7 Elevated wooden house with tile roofing 8 Elevated wooden and partial concrete house with metal roofing (zinc) 9 Elevated wooden house and partial concrete with metal roof (zinc) 10 Elevated concrete house with tile roofing 11 Elevated concrete house with metal roofing 12 Wooden house with metal roofing (zinc) 13 Double storey conceret house with tile roofing 14 Double storey conceret house with metal roofing 15 Double storey (ground floor: concrete; first floor: wood) with metal roofing 16 Double storey (ground floor: concrete; first floor: metal) with metal roofing 17 Mix single and double storey concrete house with metal/tile roofing 18 Mix elevated wooden and partial single storey concrete house (metal and tile roofing) 19 Mix elevated and partial single storey concrete house with tile roofing 20 Mix elevated wooden and partial single storey concrete house with metal roofing 21 Mix wooden and concrete house with metal roofing

Figure 3 .
Figure 3. Flow chart, red box is data pre-processing, while green box is classification of UST.

Figure 4 .
Figure 4.The urban structure type at study area In this study, we derived a total of 74 image features on the basis of four feature classes, i.e.Spectral, Band indices, Haralick texture, and height information (Table2).In addition, three different machine learning classifiers, i.e.Random Forest (RF), Support Vector Machines (SVM) and Classification and Regression Tree (CART), were used to implement UST classification.However, present study is an preliminary study on the classification of UST, therefore, the parameter of each classifier is set to default.

Figure 5 .
Figure 5.Comparison of every classifier in term of overall accuracy and Kappa coefficient.

Table 1 .
Different type of houses.

Table 3 .
Summary of confusion matrix of each finest classifier.

Table 2 .
The difference of group 1 and group 4 of sub-features is group 4 take height information into consideration.By looking at classification accuracy statistics, it noticed that the overall accuracy of group 4 is lower than group 1.The overall accuracy of RF1 and RF4 are 93.2% and 92.8%, respectively where the overall accuracy of CART1 (93.7) and CART 4 (92.3%).The reason is that the Geoeye-1 satellite image captured at 19.3° off nadir.However such off nadir degree is considerable large compared to LiDAR data, therefore, this is the source of causing misclassification.The main feature class and its associated image features.