Overview and motivation
Flood is a natural phenomenon that results in the temporary submerging with water of a land that does not occur under normal conditions (European Commission,
Flooding, a devastating natural hazard, can result from various factors such as weather patterns, geomorphologic and hydrologic conditions, environmental changes, and community development It leads to significant loss of life, property, and resources, particularly in coastal areas where populations, industries, agriculture, and infrastructure are densely concentrated These regions are especially vulnerable to storms and heavy rainfall, which contribute to water accumulation Consequently, coastal flooding has emerged as a critical concern for many countries.
Vietnam, located in Southeast Asia within a tropical monsoon climate, features an elongated shape with a coastline that spans 3,260 km and the Truong Son Mountain Range to the west Central Vietnam, characterized by its narrow coastal plains and steep rivers, experiences frequent flooding exacerbated by extreme weather events such as typhoons, tropical cyclones, and heavy rainfall, which are influenced by the region's topography Consequently, the low coastal areas in Central Vietnam face flood disasters nearly every year.
The coastal lowland region south of Danang City, situated within the Vu Gia - Thu Bon river basin, faces significant and recurring flood hazards This area experiences two distinct seasons: a rainy season from August to December, and a dry season from January to July, with the majority of rainfall (70% - 80% of the annual total) occurring between September and December On average, it is impacted by 1-2 typhoons and 1-2 major flooding events each year.
Understanding the topographical, hydrological, and environmental conditions that contribute to flood risk is essential for effective flood management in any region In the coastal lowlands of Danang City, the interplay of topography and geomorphology plays a critical role in assessing potential flood hazards.
Flood hazards are closely linked to various factors, yet the understanding of these relationships remains limited Numerous studies, including those by Ho et al., have made efforts to evaluate flood risk in these regions.
In recent studies, including those by Ho et al (2012, 2013) and Do et al (2014), a landform classification method was developed for flood hazard mapping utilizing satellite data and SRTM DEM Do et al (2014) specifically assessed the flood hazard potential in the area using ALOS PALSAR data and flow direction information from ASTER GDEM However, these studies faced limitations due to the lack of high-resolution and high-quality DEM, which significantly impacts the accuracy of flood hazard models.
This study focuses on characterizing the topographic surface and assessing flood hazard zonation in a lowland region of Central Vietnam It begins by analyzing Digital Elevation Model (DEM) generation methods, proposing a geomorphology-based approach that fuses optical stereo and Interferometric Synthetic Aperture Radar (InSAR) derived Global DEMs through a weighted averaging algorithm While this fused DEM offers potential applications in areas lacking high-quality elevation data, it still faces limitations in resolution and accuracy, particularly for large-scale flood management In the coastal lowlands south of Danang City, where detailed topographic characterization is essential, high-density field survey elevation data is utilized Consequently, the BS-Horizon DEM generation method is employed to create a high-resolution DEM for this flood-prone area, incorporating remote sensing data from ALOS PALSAR and Landsat.
TM was utilized for mapping flood inundation and classifying land cover The flood hazard potential was assessed through the analysis of Digital Elevation Models (DEM) and their derived parameters, including elevation-based flood inundation (EFI), slope, and the topographic wetness index (TWI), alongside land cover classification The Multi-parametric Analytical Hierarchy Process (AHP) was effectively employed to determine the weight of each parameter and calculate the flood hazard index (FHI) for estimating flood hazard potential Consequently, the flood hazard potential map demonstrated a strong correlation with field survey data from flood pillars and areas inundated as identified by ALOS PALSAR.
Research objectives
This research aims to characterize the topographic surface through the generation of high resolution DEM and evaluation of DEM derived parameters for flood hazard
The coastal lowland of Danang City has significant potential but is also susceptible to frequent flooding events Topography is crucial for effective flood hazard modeling; thus, the initial phase of this study concentrated on generating a high-resolution Digital Elevation Model (DEM) for the area This investigation into flood hazard potential aims to identify regions at high risk of flooding, facilitating improved flood planning for the future.
The objectives of this study can be explained as bellow:
To enhance flood hazard assessment, a high-quality Digital Elevation Model (DEM) was created for Danang City by fusing the Global Digital Elevation Model (GDEM) with the Shuttle Radar Topography Mission (SRTM) data Following this, a high-resolution DEM specifically for the coastal lowland area was developed using the BS-Horizon DEM generation method.
The evaluation of Digital Elevation Model (DEM) derived parameters, including elevation-based inundation (EFI), slope, and topographic wetness index (TWI), is crucial for assessing flood hazard potential Additionally, factors such as land use and proximity to river channels (DIST) were considered to provide a comprehensive analysis of flood risks.
- Evaluating of flood hazard potential by integrating of five factor including EFI, DIST, land use, slope and TWI using multi-parametric Analytical Hierarchy Process (AHP).
Flood situations in Central Vietnam and Danang area
Vietnam, located in low-lying coastal areas, is among the five countries most severely impacted by climate change, with flooding being the most devastating disaster due to its tropical monsoon climate and topographic features Central Vietnam is particularly vulnerable to severe floods, attributed to its unique topography, characterized by high rainfall, narrow coastal plains, and short steep rivers The region experiences an annual rainfall of approximately 3000-4000mm, the highest in the country, with 60-76% occurring during the rainy season from September to December, primarily due to storms and typhoons Major flood events in this area underscore the urgent need for effective disaster management and climate adaptation strategies.
Da Nang, situated on the lowland coastline of Central Vietnam, has its city center along the Han River The region is prone to frequent flooding and experiences annual typhoons According to statistics from the Central Committee for Flood and Storm Control (CCFSC), the disaster history in Da Nang from 1999 to 2009 highlights the city's vulnerability to natural calamities.
Between 1997 and 2009, Danang City experienced an average of two floods annually, resulting in significant economic and human losses During this period, the floods caused injuries to 277 individuals, led to the disappearance and deaths of 219 people, and incurred a staggering financial loss of 6,803 billion VND.
Review of related researches
This section provides an overview of research focused on creating high-quality Digital Elevation Models (DEMs) in Danang City and methodologies for flood hazard zonation It reviews investigations into algorithms for the fusion of global free DEMs and analyzes various studies aimed at generating high-resolution DEMs Additionally, it evaluates research on flood hazard zonation within the Danang area.
Several studies have assessed the accuracy of the Global Digital Elevation Model (GDEM) and the Shuttle Radar Topography Mission (SRTM), comparing the two digital elevation models (DEMs) Notable authors in this field include Li et al (2013), Ravibabu et al (2010), Zhao et al (2011), Suwandana et al (2012), Mukherjee et al (2013a), and Czubski et al (2013) Additionally, Reuter et al (2007) and Mukherjee et al (2013a) have contributed to the comparative evaluation of these DEMs.
In 2013, Fuss assessed the accuracy of global Digital Elevation Models (DEMs) by examining various terrain characteristics, aiming to validate their quality The distinct features and varying influences on the vertical accuracy of optical stereoscopy and InSAR techniques present a valuable opportunity for the fusion of DEMs.
This study introduces a geomorphological method for the fusion of Digital Elevation Models (DEMs), focusing on the accuracy assessment of GDEM and SRTM across various terrains, including mountain slopes, valleys, and flat areas By integrating DEMs from diverse sources with tailored weights, this approach aims to produce a superior fused elevation dataset This innovative technique has the potential to significantly improve the quality of global DEMs, addressing gaps in previous research on DEM fusion.
The fusion of global Digital Elevation Models (DEMs) holds promise for utilizing elevation data in areas lacking high-resolution alternatives Nevertheless, these fused DEMs often exhibit coarse resolution and elevation inaccuracies when compared to field survey data This is particularly critical in flood-prone lowland regions, where high-resolution DEMs are essential for detailed flood condition characterization In this study, a high-resolution DEM for a coastal lowland south of Danang City was generated using the bi-cubic spline (BS) Horizon algorithm and the exterior penalty function method Field survey elevation data served as equality constraints, while inequality elevation constraints were derived from topographic maps to effectively estimate the topographic surface Previous research has evaluated the exterior penalty function method for surface interpolation, but the impact of parameter settings on the resulting topographic surfaces has not been adequately addressed Recent studies by Tran et al (2015, 2016) have proposed suitable parameter estimations for the BS method.
The study evaluates the performance of the BS-Horizon method for generating high-resolution Digital Elevation Models (DEMs), focusing on how parameter settings affect the interpolated surface This advanced DEM generation technique is expected to yield more accurate modeling results compared to existing DEMs in Danang City.
Various methods for flood hazard mapping have evolved, utilizing hydrological, meteorological, and geomorphological approaches (Kenny, 1990; Ballais et al., 2005; Mafreda et al., 2011; Forkuo, 2011) In the Danang area, researchers have attempted to assess flood potential through hydro-geomorphological methods combined with remote sensing data (Ho et al., 2012; Do et al., 2014) However, previous studies have faced limitations due to the unavailability of high-resolution digital elevation models (DEMs) and timely satellite data from near-flood events, which may hinder the effectiveness of flood prevention plans.
This study develops a flood hazard potential map for the coastal lowland of South Danang City, Vietnam, utilizing a multi-parametric Analytical Hierarchy Process (AHP) approach High-resolution ALOS PALSAR imagery was effectively employed to identify inundated areas during the significant flood event of 2007 The region has a history of severe flooding, making this analysis crucial for disaster management and mitigation efforts.
Thesis outline
Chapter 2 introduces a geomorphological approach for integrating optical stereo and InSAR DEM, utilizing global free DEMs such as ASTER GDEM and SRTM for Danang City A weighted averaging algorithm is employed to fuse these DEMs, taking into account their accuracy across various landform types The resulting fused DEM is then assessed against a reference DEM derived from field survey data and individual global DEMs to evaluate the effectiveness of the fusion method.
In Chapter 3, a high-resolution Digital Elevation Model (DEM) for the lowland regions south of Danang City is created using bi-cubic spline BS-Horizon interpolation The model is based on equality and inequality constrained data derived from spot height field survey elevation points and a topographic map of Danang City from 2010 This constrained data is utilized in a BS-Horizon Fortran program to generate a 5-meter resolution DEM, which subsequently serves as input for
Chapter 3 analyzes the impact of various parameters such as EFI, DIST, TWI, land use, and slope on flood hazard potential The Analytical Hierarchy Process (AHP) is utilized to assign weights to these parameters and calculate the Flood Hazard Index (FHI) This index is then reclassified into four levels of flood hazard potential The resulting flood hazard zonation is validated against field survey data and flood inundation areas derived from ALOS PALSAR, revealing strong alignment with the designated zones and estimated depths Notably, approximately 84 percent of the inundated areas identified in ALOS PALSAR data correspond to high and very high flood hazard zones The significant correlation between the flood hazard zonation map and observed inundation underscores the effectiveness of the AHP-based method in assessing flood hazards.
FUSION OF OPTICAL STEREO AND InSAR DERIVED
Introduction
Digital Elevation Models (DEMs) are crucial for modeling landscapes and natural phenomena like flooding, soil erosion, and landslides High-quality DEMs, created at various detail levels, are essential for terrain-related research and applications These models can be generated from sources such as contour lines, topographic maps, field surveys, photogrammetry, radar interferometry, and laser altimetry Typically, high-resolution aerial photos, satellite data, and LiDAR data serve as inputs for producing detailed DEMs However, collecting surveying data can be both time-consuming and costly, making it challenging to generate DEMs over large areas despite the availability of various remote sensing data.
Global free Digital Elevation Models (DEMs), such as the ASTER Global DEM and Shuttle Radar Topographic Mission DEM, provide extensive coverage and accessible data for various applications, particularly in geomorphology and hydrology However, both GDEM and SRTM exhibit significant height errors that compromise the quality of elevation data Consequently, numerous efforts have been made to develop methodologies aimed at improving the quality of these global free DEMs.
This chapter explores techniques to enhance the resolution and accuracy of freely available Digital Elevation Models (DEMs) through data fusion, evaluating outcomes against high-quality reference DEMs The fusion method relies on assessing the accuracy of each global DEM and the geomorphological features of the study area Additionally, land cover units are utilized to adjust the elevation of GDEM and SRTM in relation to the bare earth surface A weighted averaging method is employed to effectively combine the input DEMs.
The study utilized an 18 landform classification map to apply different weights to GDEM and SRTM based on landform types A denoising algorithm (Sun et al., 2007) was then implemented to enhance the output fused DEM, which demonstrated a remarkable correlation with the reference DEM, achieving a correlation coefficient of R² = 0.9986 The accuracy of the fused DEM improved significantly, reducing the Root Mean Square Error (RMSE) from 14.9m for GDEM and 14.8m for SRTM to just 11.6m Additionally, terrain-related parameters extracted from the fused DEM, including slope, curvature, terrain roughness index, and normal vector of the topographic surface, showed strong comparability to reference data.
Study area
This study was carried out in Danang city, situated in Central Vietnam, covering a test site of 950 square kilometers The area features diverse elevations, ranging from sea level to 1,664 meters above mean sea level Danang city is located along the Eastern Sea coast, extending from 15°55'N to 16°14'N and 107°18'E.
The region at coordinates 108 0 20'E features diverse topography, ranging from flat landscapes to mountainous areas This variation in topography and geomorphology leads to discrepancies in Digital Elevation Model (DEM) data generated by optical stereoscopy and InSAR techniques used in the Shuttle Radar Topography Mission (SRTM) These differences result in inherent anomalies that require detection and minimization for accurate representation.
There are few studies in this area using global free DEMs such as GDEM or SRTM
Ho and Umitsu (2011) along with Ho et al (2012) introduced a landform classification method and flood hazard assessment for the Thu Bon alluvial plain in Central Vietnam, utilizing SRTM as the input DEM source They implemented a bias elimination method to refine surface elevation data to reflect the bare-earth surface However, the low resolution of SRTM (90m) may not adequately capture detailed terrain features Additionally, the InSAR technique used in SRTM can produce unreliable elevation estimates due to issues such as layovers, non-linear distortions, and atmospheric changes, as noted by Karkee et al (2008) While Ho et al (2012) concentrated on low-lying areas by masking high and upland regions, their study did not address methods for enhancing the accuracy of freely available DEMs, particularly in regions with significant topographic relief.
DEM datasets
This study utilizes global free Digital Elevation Models (DEMs), specifically GDEM Version 2 and SRTM Version 4.1 Released in October 2011, GDEM Version 2 offers a resolution of 30 meters and is derived from over 1.2 million scene-based DEMs that cover land surfaces between 83°N and 83°S latitudes The model was created using stereoscopic techniques applied to ASTER optical satellite images, which are collected by the Terra spacecraft's nadir- and aft-looking near-infrared cameras Despite improvements in GDEM Version 2 over its predecessor, including enhanced data processing algorithms, the model still exhibits height errors due to cloud coverage and contains anomalies that must be addressed before local applications (ASTER GDEM Validation Team, 2011).
SRTM Version 4.1, sourced from the Consortium for Spatial Information (CGIAR-CSI), provides publicly available elevation data derived from the Shuttle Radar Topographic Mission, which operated for 11 days in February 2000 This data covers approximately 80% of the Earth's land surface, spanning from 60°N to 56°S (Reuter et al., 2007), and utilizes both X-band and C-band technology for elevation measurements.
Interferometric Synthetic Aperture Radar (InSAR) sensors were first utilized in the Shuttle Radar Topography Mission (SRTM), which released 1-degree DEM tiles in 2003 NASA and the USGS processed this data to provide a 1-arc second resolution (approximately 30m) for the United States and a 3-arc second resolution (approximately 90m) for the rest of the world The Consortium for Spatial Information of CGIAR (CGIAR-CSI) now offers post-processed 3-arc second DEM data globally The original SRTM data underwent multiple processing steps to create a seamless elevation surface, although it initially contained no-data regions over water bodies and areas lacking sufficient textural detail The latest release, SRTM Version 4.1 by CGIAR-CSI, improves upon previous versions by effectively filling void areas.
2http://www.cgiar-csi.org
This study utilized the Shuttle Radar Topography Mission (SRTM) data, which has a resolution of 90 meters While SRTM has a lower resolution compared to the Global Digital Elevation Model (GDEM), it provides reliable coverage under all weather conditions due to its use of InSAR technology However, the limitations in resolution and vertical accuracy in certain regions necessitate editing of SRTM data prior to application Both SRTM and GDEM datasets are based on the geographic coordinate system, employing the World Geodetic System 1984 (WGS84) for horizontal referencing and the Earth Gravitational Model 1996 (EGM96) for vertical referencing.
This study utilizes a Digital Elevation Model (DEM) derived from the 1:10,000 topographic map of Danang city, published in 2010, which includes contour lines at 5m intervals and spot height data collected by the Department of Natural Resource and Environment (DONRE) in Danang, Vietnam The contour lines were created from aerial photographs taken in 2003 and were subsequently surveyed and adjusted in 2009, alongside the collection of spot height data The DEM, referred to as the "reference" DEM, was generated using the Regularized Spline with Tension (RST) algorithm in GRASS GIS 3, known for its effectiveness in interpolating elevation data Given that contour lines do not cover all areas, particularly flat regions below 10m elevation, over 190,000 spot height points were used to apply Inverse Distance Weighting (IDW) interpolation, merging this data with the RST-generated DEM for hilly areas The reference DEM was created at a 30m resolution, with a root mean square error (RMSE) of 1.66m when compared to spot height data Statistical comparisons show that the mean elevation and standard deviation of the reference DEM align closely with those of global DEMs like GDEM and SRTM, although GDEM displays significant discrepancies in maximum elevation values.
Fusion of optical stereo and InSAR derived DEM data
The SRTM data was interpolated from a 90m to a 30m resolution for comparison with other Digital Elevation Models (DEMs), while artifacts in the GDEM were removed using the fill and feather method (Dowding et al., 2004) To ensure accurate alignment, both GDEM and SRTM were co-registered to a reference DEM, followed by an evaluation of their vertical and horizontal accuracy under various topographic conditions This evaluation informed the development of a DEM fusion method that addresses factors affecting data quality It's essential to distinguish between Digital Surface Models (DSMs) like GDEM and SRTM, and Digital Terrain Models (DTMs), which represent the bare-earth surface Overestimated and underestimated elevation values in GDEM and SRTM were corrected by comparing them to the reference DEM based on geomorphology and land cover maps The corrected datasets were then used for the DEM fusion process, which involved generating a landform classification map from SRTM to identify areas optimal for different fusion methods The fusion algorithm employed weighted averaging, assigning higher weights to SRTM in flat regions and equal weights in mountainous areas, while GDEM received greater weight in valleys due to SRTM limitations The output fused DEM was then filtered using a denoising algorithm (Sun et al., 2007) and compared to the reference DEM to evaluate the effectiveness of the fusion method.
The data processing described above is shown in Figure 2.2 The data fusion workflow includes four main steps, namely pre-processing, DEM quality assessment, bias elimination and DEM fusion
The SRTM data reveals anomalies in coastal regions and select inland areas, with 377 pixels exhibiting negative values that span approximately 0.34 square kilometers To address these discrepancies, the negative pixels were corrected by averaging the elevation of neighboring pixels in a 3 by 3 grid Additionally, both SRTM and GDEM datasets have been transformed from geographic coordinates for further analysis.
The UTM_WGS84_zone 49N projection was utilized for both the reference DEM and its conversion from VN2000 It's important to note that the vertical datums differ between Global DEMs and the reference DEM; Global DEMs employ the EGM96 vertical datum, whereas the reference DEM utilizes the Vietnamese vertical datum known as Hon Dau_Hai Phong, which is linked to mean sea level at Hon Dau Island in Hai Phong province, Vietnam To align the Global DEMs with the Hon Dau_Hai Phong vertical datum, a downward offset of 1.5 meters was applied.
The SRTM data was enhanced from a 90m resolution to a finer 30m resolution using the RST algorithm available in GRASS GIS through the r.resamp.rst function This RST interpolation not only increases the DEM resolution but also mitigates the staircase effect present in the original SRTM, resulting in a smoother DEM surface Comparative analysis, illustrated in Figures 2.5a and 2.5b, demonstrates the improved profile of the interpolated SRTM against the reference DEM Additionally, the interpolated SRTM exhibits superior RMSE and correlation with the reference DEM compared to the original 90m dataset, as indicated in Table 2.2.
GDEM exhibits artifacts in the western mountainous region of Danang city, primarily due to common cloud coverage in optical satellite data, leading to a high RMSE of 75.6m in the raw GDEM To enhance data quality, these artifacts must be addressed before further processing Various void-filling algorithms have been suggested, including kriging, spline, IDW, moving window, fill and feather, and delta surface fill These algorithms can be classified into three main categories: interpolation, moving window, and fill and feather (F&F), with the F&F method being notably proposed by Dowding et al.
In this study, the F&F approach was utilized to address artifacts in the Global Digital Elevation Model (GDEM) by replacing them with the most accurate available digital elevation source, while removing void-specific perimeter bias (Grohman, 2006) Artifacts were identified by overlaying the slope map of GDEM with the difference elevation map between GDEM and a reference DEM, allowing for the digitization of visible anomalies The Shuttle Radar Topography Mission (SRTM) data was selected to fill these artifacts To ensure a smooth transition, the surface was feathered in the final data processing step using a filtering algorithm As a result, the RMSE error for the GDEM after artifact filling was reduced to just 14.9 meters, and the scatter plot demonstrated a strong correlation with the reference DEM, contrasting with the original GDEM, which displayed multiple outliers (Figure 2.3).
Comparing to original GDEM, it can also be seen that most of the artifacts were eliminated
The horizontal accuracy of global Digital Elevation Models (DEMs) was assessed by analyzing extracted stream networks Comparisons revealed that SRTM exhibited a horizontal discrepancy of approximately 15 meters, while GDEM showed a difference of around 30 meters relative to the reference DEM To enhance alignment prior to the fusion process, GDEM was shifted one pixel east and SRTM was adjusted half a pixel west Post-shifting analysis illustrated that the ridge lines and canyon bottoms in both GDEM and SRTM became more congruent with the reference DEM Additionally, the results in Table 2.3 indicate that the RMSE and correlation of GDEM improved after shifting, demonstrating a more accurate representation compared to its pre-shifted state.
In this study, the Root Mean Square Error (RMSE) for GDEM and SRTM compared to the reference DEM was found to be 14.9m and 14.8m, respectively The correlation coefficients for GDEM and original SRTM were 0.9976 and 0.9979, indicating high accuracy overall However, accuracy varies with topographic conditions; in mountainous regions, both GDEM and SRTM showed similar correlations (0.9966 and 0.9969) Notably, GDEM outperformed SRTM in steep valleys where SRTM struggled due to layover and shadow effects, which distort the representation of topography In contrast, GDEM effectively captured the intricate details of these valley areas In flatter terrains, SRTM demonstrated a correlation coefficient of 0.8504 with the reference DEM, suggesting better performance in such landscapes.
4https://earth.esa.int/applications/data_util/SARDOCS/spaceborne/ Radar_Courses/
The GDEM (R² = 0.5578) exhibits significant degradation in elevation estimates due to the area's low topographic relief As illustrated in Figure 2.7, GDEM displays numerous spikes and unstable elevation values in this flat region, whereas the SRTM data aligns more closely with the reference DEM trends.
The difference elevation maps created by subtracting GDEM and SRTM values from a reference DEM reveal significant vertical errors in mountainous regions, primarily due to forest cover and the limitations of sensing techniques in high-relief areas In contrast, both GDEM and SRTM exhibit lower vertical errors in flat areas The SRTM profile in flat regions aligns closely with the 0m line, while GDEM displays greater variations and spikes, which adversely impact its overall quality.
The differences in topographic height between global Digital Elevation Models (DEMs) and reference DEMs arise primarily from variations in vertical datum and data collection methods Vertical datum significantly impacts elevation discrepancies, while global DEMs like GDEM and SRTM, derived from satellite data, are classified as Digital Surface Models (DSMs) In contrast, reference DEMs are considered bare earth Digital Terrain Models (DTMs), leading to bias offsets that vary based on land cover.
Global Digital Elevation Models (DEMs) were adjusted to align with the Hon Dau_Hai Phong vertical datum According to the Vietnam Land Administration, the global EGM96 model closely matches the Vietnamese vertical datum, with 97% of data exhibiting a height difference of approximately 1.5 meters, while only 3% indicates a greater difference Consequently, an offset of 1.5 meters was applied to the global DEMs to account for the height discrepancy between EGM96 and the Vietnamese vertical datum.
The height offsets of global Digital Elevation Models (DEMs) were determined using a land cover map of Danang city from 2001 This approach was necessary due to the SRTM data being collected in 2000 and GDEM data derived from ASTER imagery between 1999 and 2009 The offsets were calculated by analyzing the elevation differences between GDEM and SRTM, taking land cover into account The r.statistics function in GRASS GIS facilitated this calculation.
The mean elevation differences on each land cover type were calculated, and used as
The assessment of elevation accuracy for GDEM and SRTM reveals significant discrepancies, particularly in water bodies where GDEM shows a maximum error of 4m due to low reflectance values in optical satellite data Additionally, GDEM underestimates elevation in bare land areas by an average of 2m, especially in flat terrains where topographic relief limits the effectiveness of optical stereoscopy In contrast, SRTM exhibits its highest error of 6.3m in forested mountainous regions, where layovers and shadows degrade radar data quality SRTM also encounters a notable error of 3.8m in bare land, attributed to insufficient backscatter for radar imaging Global assessments indicate that voids in SRTM data are prevalent in both mountainous and flat desert areas However, advancements in SRTM V4 have addressed water body inaccuracies through interpolation techniques and void-filling algorithms, resulting in a significantly reduced error of only 0.4m in water bodies.
The elevations for GDEM and SRTM were adjusted based on reference DEM by subtracting the elevation offsets corresponding to each land cover type, as detailed in Table 2.4 This calculation utilized the r.mapcalc function within GRASS GIS software, employing the land cover map as a foundational element The adjusted GDEM and SRTM data were subsequently utilized as input for the DEM fusion processing.
Accuracy assessment for fused DEM
Weighted averaging using a landform classification map has proven to be an effective method for Digital Elevation Model (DEM) fusion The accuracy of the fused DEM is assessed through statistical analyses, including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and linear regression Notably, the fused DEM exhibits significant improvements in accuracy, with the RMSE decreasing from 75.6m in the original Global DEM (GDEM) to 11m in the fused DEM, and from 14.8m in the original Shuttle Radar Topography Mission (SRTM) DEM to 11m in the fused version These results highlight the enhanced performance of the fused DEM compared to existing global DEMs.
The linear regression analysis between the fused Digital Elevation Model (DEM) and the reference DEM reveals a strong correlation, indicated by an R² value of 0.9986 In comparison, the original data correlation coefficients for GDEM and SRTM are 0.9976 and 0.9979, respectively This suggests that the fused DEM demonstrates a superior correlation with the reference DEM.
Table 2.6 presents a statistical comparison of the vertical accuracy of GDEM, SRTM, and the fused DEM The results indicate that the fused DEM exhibits lower minimum and maximum errors, as well as improved Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to GDEM and SRTM prior to fusion Although the final fused DEM shows a slight increase in RMSE due to smoothing effects, it effectively reduces surface mismatches and enhances the extraction of topographic parameters based on elevation differences.
The analysis of the fused Digital Elevation Model (DEM) reveals that height errors are more pronounced in mountainous regions, particularly on steep slopes, while minimal errors are found in flatter areas The histogram comparing elevation differences among SRTM, GDEM, and the fused DEM, as shown in Figure 2.15, indicates that the fused DEM achieves a central value of 0m difference, with the highest frequency of cells exhibiting this minimal error This demonstrates a significant enhancement in the quality of global DEMs achieved through the proposed DEM fusion algorithm.
The slope, profile curvature, and tangential curvature maps were derived from GDEM, SRTM, and a fused DEM, with error maps generated for each terrain parameter against a reference DEM The analysis revealed that the fused DEM exhibited smaller Mean Absolute Error (MAE) and Standard Deviation (STD), along with a stronger correlation to the reference DEM compared to GDEM and SRTM Additionally, the derivative maps from the fused DEM, as illustrated in Figure 2.16, showed no significant anomalies or terrace artifacts in the transition zones between different landform classes.
Aspect is measured in circular degrees from 0° to 360°, making quantitative comparisons challenging (Deng et al., 2007) To evaluate the accuracy of aspect and slope, the Unit Normal Vector (NV) of the topographic surface was analyzed The NVs for global and fused Digital Elevation Models (DEMs) were derived from their respective slope and aspect values These NVs were then compared to a reference DEM to assess the angular differences between the two NVs (Figure 2.17).
NV of the terrain surface (T⃗⃗ ) can be calculated as below as suggested by Hodgson and Gaile (1999)
T⃗⃗ = [𝑥, 𝑦, 𝑧] (2.6) where x = sin(aspect)*sin(slope), y = cos(aspect)*sin(slope) and z = cos(slope)
To derive the three-dimensional angular difference between two unit NVs (T⃗⃗ and S⃗ ) pointing away from the same origin, the following formula (Hodgson and Gaile,
1999) was applied: cos(i) = T⃗⃗ ∗ S⃗ = tx*sx + ty*sy + tz*sz (2.7)
The result of angular differences of NV is shown in Table 2.8 As a result, fused
DEM has smaller mean error than GDEM and SRTM, and STD of fused DEM are also comparable with global DEMs
The Topographic Roughness Index (TRI) was utilized to evaluate the quality of the fused Digital Elevation Model (DEM) in this study TRI measures the elevation differences among adjacent DEM cells, as outlined by Mukherjee et al (2013b) By calculating the residuals in elevation between a grid cell and its eight neighboring cells, the Root Mean Square (RMS) of these differences was determined to represent the TRI This analysis was applied to the reference DEM, GDEM, and SRTM, as well as the fused DEM.
DEM show correlation coefficient of 0.71, 0.75 and 0.76 respectively (Table 2.7) The
TRI derived from fused DEM compare well with the reference DEM as compared with
Limitations of fused DEM
Fused Digital Elevation Models (DEMs) significantly outperform original global DEMs, as demonstrated in earlier sections This fusion technique is particularly beneficial for global studies or in regions lacking high-resolution DEM data.
The fused Digital Elevation Model (DEM) presents significant limitations for flat lowland areas, particularly as seen in a small region south of Danang City, which is prone to flooding Remote sensing techniques, including optical stereoscopy used in GDEM and InSAR in SRTM, struggle to capture adequate three-dimensional information of the bare-earth surface due to the lowland topography Additionally, the 30m resolution of the fused DEM fails to accurately represent the detailed topographic features of these regions Field survey data indicates an elevation range from 0 to 7.7m, with slopes between 0 to 2.03 degrees, while the fused DEM shows an overestimated elevation range of -0.6 to 16.3m and slope range of 0 to 6.9 degrees, resulting in a root mean square error (RMSE) of 1.69m compared to actual point elevations Consequently, using this DEM for geomorphological analysis, particularly within the 1.69m elevation range, could yield inaccurate results To effectively characterize flood inundation in such lowland areas, generating a higher resolution DEM is essential.
GENERATION OF HIGH RESOLUTION DEM USING
Introduction
The fusion of global Digital Elevation Models (DEMs) is essential for applications in areas lacking high-quality elevation data In small lowland regions, particularly those requiring detailed topographic analysis, creating a high-resolution and precise DEM is crucial This chapter focuses on generating a high-resolution DEM for the coastal lowlands south of Danang City using the BS-Horizon algorithm, which incorporates bi-cubic spline and exterior penalty functions to achieve a spatial resolution of 5 meters Traditional interpolation methods typically rely on actual data points or contour lines, known as equality constraints In contrast, Shiono et al (2001) introduced the Horizon2000 algorithm, which utilizes field survey observations and incorporates inequality constraints, such as slope information, to determine geological surfaces Noumi (2003) further applied the Horizon2000 method to create topographic surfaces based on inter-contour height relations This approach approximates bi-linear functions within grid cells and uses an exterior penalty function to convert constrained problems into unconstrained ones, minimizing errors through iterative calculations with penalty parameters, thereby providing a close approximation to the optimal surface.
The enhanced BS-Horizon method, introduced by Nonogaki et al (2012), utilizes bi-cubic spline interpolation and an exterior penalty function to generate surfaces while adhering to equality-inequality constraints This advancement allows for the integration of piece-wise bi-cubic spline calculations, facilitating surface generation for larger datasets and achieving finer spatial resolutions Additionally, the effective application of these constraints has proven beneficial in estimating geological surfaces, making this method particularly suitable for determining topographical surfaces from topographic data.
The impact of parameter settings on surface generation using the BS-Horizon method has not been thoroughly explored Previous studies by Tran et al (2015, 2016) examined parameter values for BS-Horizon Digital Elevation Model (DEM) generation This study aims to conduct a comprehensive investigation of these parameter settings and assess their effects on the resulting topographic surfaces, ultimately enhancing the understanding of surface generation utilizing equality-inequality point elevation data.
BS-Horizon theory
The BS-Horizon program, developed by Nonogaki et al (2012), utilizes a bi-cubic spline interpolation algorithm combined with an exterior penalty function method to enhance local approximation results By applying bi-cubic spline techniques within tiles or sub-domains rather than across the entire surface domain, the program achieves improved accuracy in its computations.
In a Cartesian coordinate system, the topographic surface is defined by the equation z = f(x, y) When measuring elevation at a specific point (xi, yi), the elevation data can be categorized into three constraints: the equality constraint f(xi, yi) - zi = 0 indicates that the surface intersects the known elevation, while the inequality constraints f(xi, yi) - zi < 0 and f(xi, yi) - zi > 0 signify that the surface lies below or above the given elevation, respectively To differentiate among these constraints, a parameter l is introduced, where l = 0 corresponds to the equality constraint, l < 0 denotes the surface below the elevation, and l > 0 indicates the surface above the elevation.
Numerous viable solutions can fulfill the observations outlined in (3.1), (3.2a), and (3.2b) (Shiono et al., 2002) By positing that the topographic surface should be the smoothest among these solutions, we can frame the surface determination as a constrained optimization problem The goal is to identify a surface f(x,y) that minimizes a specific objective function.
In the equation (3.3), J(f) is function for evaluating the smoothness of surface f
The smoothness of the surface f(x,y) is highlighted when J(f) is minimized, while a larger J(f) reduces this smoothness The rectangular domain of the surface is represented by Ω in the x-y plane, with S indicating the area of Ω Parameters m1 and m2 are utilized to balance the first and second order partial derivatives of the surface f(x,y) To address the constrained optimization problem, an augmented objective function is introduced, as detailed in equation (3.4).
The exterior penalty function method aims to find a surface f(x, y) that minimizes the value of Q(f, α), defined as Q(f, α) = J(f) + αR(f) In this equation, J(f) assesses the smoothness of the surface, while R(f) serves as an exterior penalty function that evaluates the goodness of fit The parameter α, which must be greater than zero, regulates the balance between the smoothness and the fit quality of the surface.
R(f) indicates the mean square of residuals for data which do not satisfy constraints It is also considered as the degree of violations of constraints R(f) is defined by equation (3.5):
In the equation \( n\sum_{i=1}^{N} E_i^2 \), \( n \) represents the number of height data points that do not meet elevation constraints, while \( N \) denotes the total amount of data If only equality data is considered for interpolation, then \( n \) equals \( N \) The term \( E_i \) signifies the residuals, which are the differences between the interpolated data and the observational data.
As the value of R(f) decreases, the goodness of fit improves, and conversely, an increase in R(f) results in a poorer fit To achieve the minimum value of Q, the exterior penalty method has been utilized, as it is not feasible to analytically determine the minimum Q value due to the constraint conditions involved.
The nonlinear system R(f) utilizes an iterative optimization algorithm that progressively applies a series of penalties, α {α₁, α₂, …, α𝑁 ITR}, to identify the optimal solution (Masoud, 2003) This method incorporates a calculation ratio (r) based on a specific equation, enhancing the efficiency of the optimization process.
(3.8) where N ITR is the number of iteration, α 1 is an initial penalty and α 𝑁 𝐼𝑇𝑅 is a final penalty At the initial stage of the repeated calculations, the value of the objective function
In the context of surface calculations, when J is small and R is large, it indicates that the surface does not meet the constraint conditions (Noumi, 2003) By incrementally increasing the sequences of α, the surfaces f(x,y) progressively align with these constraints It is recommended to initiate the process with a small α to define R(f) and subsequently Q(f, α), allowing for an evaluation of the resulting surfaces.
Several parameters such as M x , M y , α min , α max , N ITR , m 1 and m 2 need to be set for BS- Horizon In summary, they can be explained as below;
- M x and M y : the numbers of tiles (sections) that constitute the surface domain (Ω) in which bi-cubic spline function is applied
- α min and α max : the range for penalty parameter α in the exterior penalty function α controls the balance between smoothness of surface and the degree of violation of constraints
- N ITR : number of iterations for a given α range
In the context of spline functions, parameters m1 and m2 play crucial roles in optimizing surface characteristics Specifically, m1 is associated with minimizing the slope of the surface, represented by the first-order partial derivative, while m2 focuses on minimizing the curvature, indicated by the second-order partial derivative, as detailed in Equation (3.3).
Data
The elevation data for a lowland area of 16 square kilometers (4 by 4 km) was obtained from the Department of Natural Resource and Environment in Danang City, Vietnam This test area is characterized by the 1A National Road that runs through the center and is bordered by the Qua Giang and Cai rivers The topography of the region is relatively flat, with elevation ranging from 0 to 7.2 meters, and a total of 9,730 elevation points were collected during the field survey.
2009 The average distance between measured elevation points is around 40m and the observational accuracy is at 10cm
Table 3.1 and Figure 3.2 illustrate the distributed density of point elevation data across various tile sizes (M x, M y) The data presented in Table 3.1 highlights the frequency of point elevation distribution within each tile, quantified by the number of points for different configurations of M settings (M x).
As the size parameter M increases from 50 to 400, the distribution of elevation point data changes significantly At M = 50, data points range from 0 to 12 per tile, with the most common concentration being 4 points per tile However, as M increases, the density of data distribution decreases, particularly at M = 300 and M = 400, where data points are limited to 0, 1, or 2 per tile Additionally, with larger M values, the number of tiles with no data increases due to the smaller tile size, especially noticeable from M = 140 and above, where more than 50 tiles exhibit a lack of data.
Evaluating effects of parameter settings on the BS-Horizon DEM generation
A significant percentage of tiles lack data, leading to the BS-Horizon surface generation relying on elevation information from adjacent tiles for estimates Notably, there are no observational elevation points along the rivers, and surrounding areas exhibit low elevation, which contributes to an underestimation of elevation in the interpolated surface Additionally, the construction of the 1A National Road in this flood-prone study area, which has a higher elevation than nearby points, further complicates the accurate representation of interpolated surfaces.
3.4 Evaluating effect of parameter settings on the BS-Horizon DEM generation
This study evaluates parameter settings using a 5m resolution output Digital Elevation Model (DEM) According to Hengl (2006), the suitable resolution for interpolated surfaces depends on the study area size and the number of elevation points Based on Hengl's recommendations, a 5m spatial resolution is deemed appropriate for DEM generation with the current dataset.
Surface approximation was conducted by varying parameter settings, specifically examining M x and M y within the range of 50 to 400, and selecting α from 1 to 10^12, while m 1 and m 2 were valued between 0 and 1 To assess the impact of these parameters on BS-Horizon surface generation, Digital Elevation Models (DEMs) were initially created using only equality constraints Following this, inequality constraints were incorporated alongside the equality constraints to form a comprehensive set of equality-inequality constraints, which were subsequently input into the BS-Horizon program.
Figure 3.3 illustrates the equality and inequality constraints utilized in this case study, focusing on the topographic map of Danang City (scale 1:25,000, 2010) The study area features an elevation range between 0m and 10m, establishing these parameters as inequality constraints In the figure, point elevations, representing equality constraints, are depicted as circles, while inequality elevations are indicated by triangles Relying solely on point elevation data may result in an estimated surface that closely follows the given points but can lead to overshoots and undershoots due to the lack of elevation control in areas without point data However, incorporating inequality data helps to maintain the interpolated surface within the 0 to 10m elevation range, ensuring that the generated surface elevation does not significantly exceed the study area's defined limits.
The study area is defined by a single constraint ranging from 0m to 10m, leading to the generation of inequality data that includes XY coordinates, Z coordinates, and their relationship to the constrained elevation (l) This data was created at a specified spatial interval, with l set to +1 for elevations above 0m and -1 for elevations below 10m Subsequently, the point elevation data was integrated with these inequality constraints to form a dataset in CSV format that combines equality and inequality constraints This dataset was then utilized as input for the BS-Horizon program, where l equals 0 for the equality point elevation data.
When generating BS-Horizon DEM with only equality constraints, the interpolated surface must adhere to the specified elevation data, rendering iterative calculations ineffective However, incorporating inequality constraints allows for the estimation of an exterior penalty function through iterative calculations with an increasing sequence of penalties, α {α 1, α 2, …, α 𝑁 ITR} This study introduces a ratio (r) based on equation (3.7), which should be small enough to capture trends while remaining computationally efficient; a ratio of r = 1.189 is employed By applying this ratio and a penalty range from 1 to 10^12 in equation (3.8), the optimal number of iterations determined for this case study is 161.
In the initial stages of iterative calculation, the algorithm begins with a rough surface approximation characterized by high smoothness (low J values) and low goodness of fit (high R values) This occurs due to numerous control points failing to meet the constraint conditions of the reference surfaces, indicating potential for improved estimations (Masoud, 2003) Following the first approximation, point data that meet the constraints are excluded from subsequent iterations, as they become equality height data Control points that do not satisfy either the equality or inequality constraint conditions are then identified and utilized to generate a new surface at the next α value, allowing for the calculation of goodness of fit (R) After completing calculations for each α sequence, height data that do not meet the constraint conditions are extracted to evaluate the approximated surfaces This process continues until the maximum iteration limit (N ITR) is reached, while parameters such as M, α, m1, m2, and various inequality constrained intervals are assessed throughout the iterative calculation.
3.4.2.1 M and α settings in case of using only equality constraints
Figure 3.4 illustrates the trends of R(f), J(f), and Q(f) across various M and α settings using equality data, while Table 3.2 presents the calculated values of these functions along with the statistical parameters for each corresponding DEM The results were derived with m1 set to 0 and m2 set to 1 Notably, the initial values of R(f), J(f), and Q(f) are consistent across all M configurations, indicating that at the starting α value of 1.0×10^0, the R, J, and Q values remain similar, as detailed in Table 3.2.
As the parameter α increases, the value of R(f) decreases, suggesting a reduction in the degree of constraint violations and a gradual alignment of the surface with elevation constraints Specifically, R(f) varies from 0.4940 at α = 1.0×10^0 to 0.3361 at α = 1.0×10^8 when M is set to 50.
The value of R decreases from 0.4914 at α = 1.0×10⁰ to 0.1044 at α = 1.0×10⁸, with the rate of reduction varying based on M settings At M = 50, R shows minimal change, dropping from 0.4940 to 0.3361 In contrast, at M = 100, R decreases significantly from 0.4914 to 0.1044, and at M = 400, it is minimized from 0.4908 to 0.0025 While R(f) yields similar values at α = 1.0×10⁰ across different M settings, notable reductions are observed at larger M values.
The smoothness of a surface is characterized by the minimization of the function J(f) As indicated in Table 3.2 and illustrated in Figure 3.4, J(f) increases with higher values of α For M = 50, as shown in Figure 3.4a, J rises from 5.58×10^-2 to 4.43×10^3 when α varies from 1.0×10^0 to 1.0×10^8 Additionally, for M = 100, Figure 3.4b demonstrates a steep increase in J(f).
5.69×10 -2 to 5.44×10 4 ) In cases of M = 200 and M = 400, J slowly increase from 5.71×10 -
The analysis reveals that the increase in J(f) values, ranging from 2 to 9.07×10¹ (M = 200) and from 5.71×10⁻² to 5.32×10¹ (M = 400), is influenced by varying M settings Specifically, a higher M setting results in a slower rate of change in J, indicating that as M increases, the smoothness of the generated surfaces diminishes.
J and R values tend to become stable in the upper range of α (from α = 1.0×10 6 to α
For M values of 150 and above, particularly at M = 200 and M = 400, R is minimized to approximately 0.0025 and remains constant after α reaches 1.0×10^8 This trend is consistent for M = 150 and continues to be observed in larger M settings, as detailed in Table 3.2 and illustrated in Figures 3.4c and 3.4d.
J also slowly changes and becomes stable at higher values of α for M = 150 and above
The values of the augmented objective function Q(f) = J(f) + αR(f) fluctuate with changes in α, resulting in an exponential increase in Q as α rises Although R is minimized with increasing α, both J and α escalate exponentially, leading to significant growth in Q values For instance, when M = 50, Q rises from 0.549 to 33.6 million, and when M = 400, it increases from 0.548 to 250 thousand, as illustrated in Table 3.2.
Figure 3.5 illustrates various Digital Elevation Models (DEMs) at different M and α settings, while Table 3.2 presents the statistical results of the corresponding surfaces The black lines in Figure 3.5 represent contours extracted from DEMs at 0.2m intervals, with a minimum elevation of -4m and a maximum of 8m Notably, closely spaced contours indicate areas of overestimation or underestimation of the surface Contours exceeding 8m or below -4m are omitted from Figure 3.5, as they represent anomalies in the generated surface.
Discussion
3.5.1 Comparing BS-Horizon DEM generation from equality and equality- inequality constrained data
When comparing surface behaviors using only equality data versus equality-inequality data, it is evident that the Digital Elevation Model (DEM) derived from equality data immediately meets elevation constraints, exhibiting small RMS errors (R approximately 0.4900 at α = 1.0×10^0) The elevation range for this DEM aligns closely with the study area's elevation range (refer to Table 3.2) Conversely, the DEM generated from equality-inequality constraints typically fails to satisfy the initial elevation constraints, as indicated by R values around 2.6200 and J values of approximately 2.70×10^-5 at α values of 1.0×10^0 across all M settings, resulting in mostly flat surfaces with elevation ranges from -0.0017m to 0.1190m (see Table 3.3) However, as α increases, the surfaces gradually begin to meet the constrained conditions.
DEMs derived solely from equality data typically meet most elevation constraints, resulting in small RMS errors at the initial alpha setting However, these surfaces exhibit significant oscillations due to the limited number of elevation constraints In regions lacking point elevation data, the estimated elevations can greatly exceed the actual range of the study area For instance, as shown in Table 3.2, the DEM generated from equality data has a minimum elevation of -773.71m and a maximum of 924.85m when M = 50 and α = 1.0×10^8, indicating a considerable elevation range.
When M = 50 and α = 1.0×10^12, the elevation range is between 23,111.86m and 56,683.90m In cases where the Digital Elevation Model (DEM) is derived from equality-inequality data, the presence of inequality constraints limits the estimated elevations, resulting in less variation compared to using only equality constraints Specifically, when M = 200 and α = 1.0×10^12, the maximum elevation range for the DEM generated from equality-inequality constraints is from -664.65m to 576.45m (see Table 3.3).
Digital Elevation Models (DEMs) generated solely from equality data often exhibit significant elevation undershoots, particularly in areas lacking field data, such as rivers and transitions between different land uses, like the boundaries of paddy fields and man-made structures These undershoots occur due to substantial elevation differences between adjacent regions By integrating inequality data into the BS-Horizon model, the number of elevation constraints increases, leading to a reduction in these undershoots However, even DEMs that incorporate both equality and inequality constraints may still display undershooting, especially with higher alpha settings It is important to note that spline algorithms struggle to accurately represent discrete transitions, frequently resulting in undershooting at the edges of floodplains and other slope breaks (Hengl and Reuter, 2008).
3.5.2 Selection of parameters for BS-Horizon DEM generation
Experiments with various parameter settings have shown that the BS-Horizon program effectively generates surfaces that meet elevation constraints The use of equality-inequality data has proven to be a superior input for Digital Elevation Model (DEM) generation compared to relying solely on equality constraints In this case study, a surface with R(f) ≤ 0.25 is deemed to adequately satisfy the elevation requirements; however, the M = 50 setting fails to produce any compliant DEMs While the M = 100 setting can generate a surface with R ≤ 0.25, larger M settings yield improved statistical parameters Notably, the surfaces produced with M = 200 and M = 400 are similar across all statistics, making M = 200 the optimal choice for BS-Horizon DEM generation in this study, as it meets the R ≤ 0.25 criterion while maintaining acceptable processing times.
The α parameter plays a crucial role in balancing surface smoothness (J) and goodness of fit (R) in Digital Elevation Models (DEMs) When α is set to 1.0×10^0, the resulting surfaces are the smoothest, indicated by the lowest J values However, as α increases, R decreases while J and α rise exponentially, leading to a significant increase in the augmented objective function Q(f, α) According to BS-Horizon theory, the ideal surface minimizes Q values, effectively balancing smoothness and fit Nonetheless, increasing α compromises surface smoothness and introduces contour artifacts, impacting the overall quality of the generated surfaces.
Based on the analysis of surfaces generated at various alpha values in Figures 3.5 and 3.7, it is recommended that for the current dataset, the optimal DEM occurs at α = 1.20×10² and M = 200 when utilizing only equality constraints Conversely, the parameters for DEM derived from both equality and inequality constraints are identified at α = 1.58×10⁴ and M = 200.
Parameters m 1 and m 2 also should be appropriately selected All of R(f), J(f)and
The parameters m1 and m2 influence the quality of surface generation, with an increase in m1 and a decrease in m2 leading to reduced Q(f, α) values Surfaces typically exhibit flat characteristics in regions with point elevation and become steep at the transitions between different topographic areas To balance the minimization of constraint violations, maximize surface smoothness, and accurately represent topographical features, this study has chosen m1 = 0.5 and m2 = 0.5 for optimal surface generation settings.
In this case study, the BS-Horizon Digital Elevation Model (DEM) was created using equality-inequality constrained data at a 5-meter resolution, with parameters set to M x = M y = 200, α = 1.58×10^4, m1 = 0.5, and m2 = 0.5 It is important to note that these parameters may vary for different datasets based on the elevation criteria and topographic features of the study area Therefore, a careful examination of suitable parameters is recommended, taking into account the specific characteristics of the area being studied.
This study is the first to comprehensively evaluate the impact of various parameters in the bi-cubic spline algorithm used in the BS-Horizon program The bi-cubic spline function in BS-Horizon uniquely integrates equality and inequality constraints, enhancing topographic surface generation To optimize the use of the BS-Horizon program, it is essential to first analyze parameter M according to the specified elevation criteria Next, selecting the appropriate α is crucial to achieve a balance between J(f) and R(f) Finally, setting m1 and m2 correctly is vital for improved representation of the generated surface.
BS-Horizon DEM assessment
The 5m resolution BS-Horizon Digital Elevation Model (DEM) was created using equality-inequality constrained data with parameters set at α = 1.58×10^4, Mx = My = 200, m1 = 0.5, and m2 = 0.5 Due to the absence of reference height data, the accuracy of the DEM cannot be directly assessed (Masoud, 2003) Instead, the accuracy of the estimated surface is evaluated using elevation data from field survey points that contributed to the DEM generation, revealing insights into the model's reliability.
The Mean Square Error (RMSE) for the current dataset is 0.4963m, significantly lower than the RMSE of 1.98m for the fused Digital Elevation Model (DEM) derived from GDEM and SRTM data This indicates that the BS-Horizon DEM provides improved resolution and vertical accuracy, particularly beneficial for flood-prone lowland areas Subsequently, the parameter settings for BS-Horizon were utilized to create a high-resolution 5m DEM, facilitating a comprehensive assessment of flood hazards across larger regions in Central Vietnam.
FLOOD HAZARD ZONATION USING MULTI-PARAMETRIC
Introduction
This study investigates the generation of topographical surfaces for flood hazard zonation in a lowland area south of Danang City, Vietnam, which has a history of frequent flooding Flooding is a prevalent natural hazard globally, influenced by geomorphic factors such as elevation, slope, topographic wetness index, and proximity to river channels Understanding the impact of these factors and integrating their effects is crucial for effective flood management Central Vietnam, particularly Danang and Quang Nam provinces, has faced several severe floods, making hazard zonation essential for evaluating flood risk in the region.
Various hydrological, meteorological, and geomorphological methods have been developed for flood hazard mapping (Kenny, 1990; Ballais et al., 2005; Mafreda et al., 2011; Forkuo, 2011) In the Danang area, researchers have attempted to assess flood potential using hydro-geomorphological methods combined with remotely sensed data (Ho et al., 2012; Do et al., 2014) However, these studies face challenges due to the unavailability of high-resolution digital elevation models (DEMs) and satellite data collected during flood events, potentially compromising the accuracy of results essential for effective flood mitigation planning.
This chapter presents a hierarchical model for flood hazard zonation in Central Vietnam's coastal lowlands, utilizing a multi-parametric approach Key factors considered include elevation-based flood inundation, proximity to river channels, land use, slope, and the topographic wetness index The Analytical Hierarchy Process (AHP) method is employed to assign weights through pair-wise comparisons of these parameters, ensuring a systematic evaluation of flood risk.
The Flood Hazard Index was calculated using 50 data points, resulting in the classification of flood hazard susceptibility into four distinct zones Validation of this zonation was conducted by comparing it with field survey flood pillar points and a flood inundation map derived from the 2007 ALOS PALSAR image The analysis revealed a strong correlation between the demarcated zones and the estimated flood depths, with approximately 84 percent of the inundated areas identified in the ALOS PALSAR data located within high and very high flood hazard zones This excellent alignment between the flood hazard zonation map and observed inundation underscores the effectiveness of the AHP-based method in assessing flood hazards.
Study area and data used
The study area spans 98 square kilometers of lowland in the southern part of Danang city, Vietnam, encompassing rural regions of both Danang and the northern part of Dien Ban district in Quang Nam province Bordered by the Cam Le and Yen rivers to the north and west, the Yen river merges with the Cai and Vinh Dien rivers at the center of the area, all of which flow into the Han river, part of the Vu Gia river system The eastern boundary is defined by the Co Co river, which has been disconnected from the Thu Bon river, the largest river in Central Vietnam, due to channel filling The topography is predominantly flat, with elevations ranging from 0 to 10 meters This region experiences two distinct seasons: a rainy season from August to December and a dry season from January to July, with the majority of rainfall occurring between September and December On average, the area is impacted by 1-2 typhoons and serious flooding events annually (Do et al., 2014).
The in-situ spot height data surveyed by Department of Natural Resource and
In 2009, the Department of Natural Resources and Environment (DONRE) in Danang City utilized 79,600 elevation points to generate a 5-meter Digital Elevation Model (DEM) for the study area To assess land use related to flood hazard zonation, Landsat TM satellite imagery from 2007 was analyzed Additionally, RapidEye imagery with a 5-meter resolution was employed to detect river channels Flood pillar data collected during a field survey in March 2015 served as reference data for evaluating the flood hazard zonation map, further enhancing the understanding of flood inundation patterns.
51 areas extracted from ALOS PALSAR in 2007 was also used for validation of the flood hazard zones demarcated from this study The data used is described in Table 4.1.
Methodology
The flood hazard zonation for the study area was developed using a multi-parametric approach that considers five key factors: EFI, DIST, Landuse (LU), Slope (S), and TWI, evaluated through the Analytic Hierarchy Process (AHP) This method employs a paired comparison matrix to assess the relative importance of each factor and assigns weights accordingly The resulting Flood Hazard Index (FHI) classifies areas based on their susceptibility to flooding, following the methodology outlined by Kazakis et al (2015) The proposed flood hazard zonation method is illustrated in the flow chart in Figure 4.2.
Following the calculation of the weights, the FHI can be calculated using the equation (4.1):
(Equation 4.1) where n is number of criteria, r is the rating values of parameters assigned based on their relations to flood, w is the weight of each parameter
The present study utilizes parameters such as EFI, slope, and TWI derived from the Digital Elevation Model (DEM), which is crucial for characterizing the geomorphological and hydrological conditions of the target area High-quality DEM generation is essential for achieving accurate flood hazard zonation results For this purpose, in-situ spot height data comprising 79,600 elevation points was incorporated into the BS-Horizon DEM generation process (Nonogaki et al.).
The BS-Horizon method, developed by Nonogaki et al (2012), utilizes a bi-cubic spline algorithm and an exterior penalty function within a FORTRAN program for interpolation This technique allows for the integration of elevation data as equality-inequality constraints on the surface, ensuring that the resulting Digital Elevation Model (DEM) adheres closely to these elevation constraints.
In their 2016 study, Tran et al investigated parameters for generating the BS-Horizon digital elevation model (DEM) surface, emphasizing the importance of surface smoothness for flood hazard zonation mapping The study utilized a DEM to identify homogeneous flood-prone areas, applying parameters α = 1.0×10^3, Mx = 348, My = 242, m1 = 0.5, and m2 = 0.5 as outlined by Nonogaki et al (2012) and Tran et al (2016) This resulted in a 5m resolution DEM for the study area, with elevations ranging from 0m to 11.3m The mean error of the DEM surface generated by the BS-Horizon program was 0.89m, while the Root Mean Square Error (RMSE) compared to field survey points was only 0.48m Notably, the BS-Horizon DEM outperformed global free DEMs, such as GDEM and SRTM, in terms of RMSE, indicating its effectiveness for improved flood hazard zonation.
4.3.1.2 Flood inundation map from satellite image
The flood inundation map is crucial for assessing and validating flood hazard zonation results This study utilized Advanced Land Observing Satellite (ALOS) Phased Array L-band Synthetic Aperture Radar (PALSAR) imagery, specifically from 31st October 2007, to analyze the initial phase of a significant flood event that occurred between 28th October and 9th November 2007 The ALOS PALSAR data, recorded in dual polarization (HH and HV) with a spatial resolution of 12.5m, indicated that HH polarization is more effective for flood inundation mapping, as noted by Twele and Martinis (2009) Therefore, HH band data was chosen to extract the extent of flooding, employing a method that distinguishes between water and non-water objects Initially, a Lee filter was applied to minimize speckle noise in the PALSAR data, followed by the calculation of the normalized backscattering coefficient (σ0) in decibels (dB) from the Digital Number (DN) values, as outlined by Shimada et al (2006).
Backscatter coefficient obtained from ALOS PALSAR indicates the strength of
Microwave irradiation emitted from the antenna and reflected back after scattering off a surface allows for the analysis of the backscattering coefficient derived from PALSAR images, which is crucial for estimating surface water areas (Do et al., 2014) The calculation of the backscattering coefficient for the PALSAR standard product follows the methodology outlined by Shimada et al (2006), expressed in Equation 4.2: σ0 = 10*log10(DN²) + CF.
The Normalized Backscattering Coefficient (NBC), denoted as σ 0, is calculated using the radar amplitude (DN) represented as a digital number, along with a calibration coefficient (CF) for PALSAR standard products, which is set at -83 dB according to Shimada et al (2006) In this study, the NBC was determined using the PALSAR level 4.1 product.
The backscatter coefficient indicates variations in brightness, with smooth water surfaces appearing dark due to weak backscatter (Do et al., 2014) PALSAR imagery from October 31, 2007, reveals extensive dark areas corresponding to water coverage To differentiate between water and non-water objects, a threshold of 10dB was established based on training sample histograms, though some water pixels were still identified in non-flooded regions To improve the accuracy of the flood inundation map, a 5m Digital Elevation Model (DEM) was incorporated, as all flood pillar points recording inundation depths were below this elevation Consequently, a threshold of 5m was applied to eliminate inundated pixels situated above this height The resulting flood inundation map aligns well with water objects identified in the PALSAR imagery from the same date (Figure 4.3).
The Analytic Hierarchy Process (AHP) is a process that converts multidimensional complexity into an integrated single dimension scale of priorities (Choosumrong et al.,
The Analytic Hierarchy Process (AHP), developed by Saaty in 1990 and refined in 2008, is recognized as a powerful method for addressing complex decision-making scenarios Its application in weighting has been extensively documented across various studies, highlighting its effectiveness and versatility in research contexts (Mishra, 2013; Choosumrong et al., 2012; Ouma and Tateishi, 2014; Kazakis et al.).
The Analytic Hierarchy Process (AHP), developed by Saaty in 1990, is a decision-making tool designed to resolve conflicts and analyze judgments by evaluating the relative importance of various criteria This method employs a pair-wise comparison approach using a 9-point scale to facilitate a structured assessment of these criteria (Danumah et al., 2015; Danumah et al., 2016).
The Analytic Hierarchy Process (AHP) consists of four key steps (Ouma and Tateshi, 2014) First, the problem is broken down into simpler components organized in a decision hierarchy Next, priorities among these decision elements are established, utilizing a pair-wise comparison matrix to evaluate the significance of each criterion in relation to the overarching goal Values ranging from 1 to 9, as outlined by Saaty (1990), indicate the importance levels of the criteria In the third step, the relative weights of the decision elements are computed by calculating normalized values for each criterion and determining the principal eigenvectors This involves dividing each cell's value by its column total and averaging the normalized values Finally, the consistency of the subjective evaluations is assessed by calculating the Consistency Index (CI) based on the maximum Eigenvalue (λmax), which is derived from the product of the relative weights and the original comparison matrix totals, as detailed in Table 4.2 (Vasgas, 2010).
The Consistency Index (CI) is calculated using the maximum Eigenvalue (λmax) of the comparison matrix and the number of evaluated criteria (n) To assess the adequacy of the CI, Saaty (1990) recommends using the Consistency Ratio (CR), which is the ratio of the CI to the Random Consistency Index (RI) The formula for calculating the CR is provided in the relevant equation.
The consistency ratio (CR) is a crucial measure in evaluating the consistency of a comparison matrix, calculated using the consistency index (CI) and the random consistency index (RI) According to Danumah et al (2016), the values of RI vary based on the number of criteria (n) A comparison matrix is deemed consistent when the CR is below 0.1, or 10%, as established by Saaty in 1990 If the CR exceeds this threshold, it indicates a lack of consistency in the comparisons made.
55 necessary to revise the comparison matrix and re-calculate the weights for better weighting scheme
In susceptibility mapping, various conditioning factors serve as independent variables (Liu and De Smedt, 2005) Researchers have created flood hazard models using a multi-parametric approach (Mishra, 2013; Ouma and Tateishi, 2014; Kazakis et al., 2015; Tehrany et al., 2015) An effective flood model minimizes the number of independent data while maintaining high accuracy (Tehrany et al., 2014) This study integrates five key factors into the AHP flood hazard model: elevation-based flood inundation (EFI), distance from the river channel (DIST), land use (LU), slope, and topographic wetness index (TWI) Figure 4.4 illustrates the spatial distribution of the normalized parameter ratings on a 0-10 scale for the study area.
4.3.3.1 Elevation-based Flood Inundation (EFI)
Topography significantly influences the spatial variation of hydrological conditions, with Digital Elevation Models (DEMs) being crucial in assessing flood hazards, as lower elevations are more prone to flooding Previous studies have primarily utilized elevation data from DEMs for flood susceptibility analysis This research evaluates the impact of elevation on flood hazards using a water filling algorithm implemented in the r.lake.series module of GRASS GIS software This model simulates flood inundation scenarios by filling areas based on surrounding elevations, generating different flood inundation scenarios linked to specified water levels The Elevation-based Flood Inundation (EFI) model utilizes a 5m resolution DEM and river channel data from RapidEye imagery to identify areas at higher risk of inundation The analysis normalizes EFI data into rating values that correlate with flood hazard potential, highlighting regions susceptible to flooding when water levels rise above 1 meter.
Results
4.1 Determining the weights for parameters of flood hazard
Using the Analytic Hierarchy Process (AHP) method to determine the relative weights of flood hazard parameters, a pair-wise comparison was conducted through a 5x5 matrix, as detailed in Table 4.2 Each row in the pair-wise comparison (PC) matrix indicates the relative importance of two parameters, with the first row highlighting the significance of the Elevation Flood Index (EFI) compared to other factors EFI was deemed more critical than slope, receiving a value of 5, while slope's value in relation to EFI was assigned the inverse (1/5) This approach was consistently applied across all parameters listed in Table 4.2 EFI emerged as the most significant parameter, as it integrates elevation data with river channel relationships, crucial for flood scenario modeling The distance from the river channel was identified as the second most important factor, given that flooding typically occurs when water overflows from the river and impacts surrounding areas, making locations nearer to drainage networks more vulnerable The Topographic Wetness Index (TWI) ranked third in importance, aligning with findings from relevant studies (Tehrany et al., 2015) Conversely, land use and slope were assessed as having lower significance in the pair-wise comparison.
The pair-wise comparison yielded normalized values (NV) and weights (w i) for flood hazard parameters, as detailed in Table 4.2 The Effective Flood Intensity (EFI) emerged as the most critical parameter, assigned the highest weight of 0.42, while the weights for Distance (DIST), Topographic Wetness Index (TWI), Land Use (LU), and slope were 0.26, 0.16, 0.10, and 0.06, respectively The calculated λmax was 5.061, with five criteria considered (n = 5) and a random index (RI) of 1.12, as referenced by Danumah et al (2016) The consistency index (CI) was determined to be 0.015, leading to a consistency ratio (CR) of 0.014, which indicates a satisfactory level of consistency since the CR value is below the acceptable threshold.
0.1, the weighting scheme used in this study is considered to be appropriate
4.2 Flood hazard index (FHI) and flood hazard zonation
FHI was calculated based on Equation (1) using the weighting scheme determined by AHP method The input parameters were normalized into rating values (from r = 0 to r
The Flood Hazard Index (FHI) was calculated by combining rating values with corresponding weights, resulting in a range from 1.28 to 10.00 A flood hazard zonation map was created by categorizing the FHI into four classes: low, moderate, high, and very high, based on an analysis of the FHI data histogram to optimize inter-class variance Additionally, the 2007 flood inundation map from ALOS PALSAR was utilized to refine the thresholds for identifying high and very high flood hazard areas Notably, river channels were designated as areas with no flood hazard, as the typical water levels in permanent bodies do not pose a flooding risk.
The study reveals that 59.9% of the area is classified as having high to very high flood hazard levels, primarily found in low-elevation regions (0 to 5m) that are predominantly used for agriculture (58.9%) and settlements (34.6%) Additionally, 17.2% of the area falls within a moderate flood hazard zone, while 16.5% is categorized as low flood hazard Notably, the remaining 6.4% of the study area consists of river channels with no associated flood hazard.
Characterizing the topography and geomorphology is crucial for assessing flood hazards in the coastal lowlands of Danang City, Vietnam Despite previous efforts to evaluate flood hazard potential in this region, the reliance on global free DEM data, such as SRTM or GDEM, has limitations that result in low accuracy for flood models High-quality elevation data is often unaffordable in developing countries, making it essential to explore evaluations of global DEMs and their fusion to enhance elevation data for regional flood hazard assessments In areas with available field survey elevation points, the generation of high-resolution DEMs using the BS-Horizon method has proven to be an effective solution for achieving high accuracy in flood hazard zoning.
Global Digital Elevation Models (DEMs) derived from remote sensing data often contain vertical and horizontal inaccuracies, making it crucial to assess their quality and validate their accuracy prior to application This study evaluates the accuracy of the GDEM and SRTM by analyzing height differences against a reference DEM To address extreme elevation artifacts in GDEM, SRTM data was utilized as auxiliary information Additionally, river networks extracted from both DEMs facilitated the detection and correction of horizontal errors, enhancing co-registration The study also quantified the bias effects caused by tree-top canopies and buildings by comparing Digital Surface Models (DSMs) with elevation data from the reference DEM.
In 2001, a land cover map of Danang city was utilized to assess the height discrepancies between GDEM and SRTM across various land cover types After identifying the bias offsets, adjustments were made to correct the elevation data of these Digital Elevation Models (DEMs) in relation to the bare land surface.
A global assessment of Digital Elevation Models (DEMs) in Danang city reveals that the accuracy of GDEM and SRTM differs based on the area's geomorphological features GDEM demonstrates superior accuracy compared to SRTM in mountainous regions, particularly in steep valleys However, in flatter terrains, GDEM exhibits numerous spikes and unstable elevation values, whereas SRTM aligns more closely with the reference DEM The fusion of these two models can enhance elevation data reliability.
The integration of 61 global Digital Elevation Models (DEMs) using a geomorphological approach significantly enhances the quality of free DEMs for Danang City, Vietnam By employing a data fusion technique that utilizes weighted averaging of GDEM and SRTM based on topographic context, the method assigns the highest weight to GDEM in valleys and to SRTM in flat areas, with local adjustments made according to landform types Comparative analysis with reference DEMs reveals that the fused DEM offers improved accuracy over individual global DEMs, effectively eliminating most artifacts This innovative method provides a valuable solution for regions lacking high-quality DEMs.
The fusion of GDEM and SRTM offers an opportunity to access free Digital Elevation Model (DEM) data with acceptable quality; however, the resolution of the fused DEM limits its effectiveness for detailed topographic characterization This study explores the generation of high-resolution DEMs using the BS-Horizon method, focusing on the impact of various parameter settings The evaluation revealed that parameters α, M, m1, and m2 significantly influence surface generation Additionally, the implementation of equality and inequality elevation constraints has proven effective in creating topographic surfaces, showcasing varied responses to different parameter configurations.
This study marks the first comprehensive effort to generate a Digital Elevation Model (DEM) for the lowland areas of Danang City, Vietnam Utilizing elevation data from field surveys and the bi-cubic spline algorithm within the BS-Horizon program, this research effectively incorporates equality-inequality constraints for topographic surface interpolation The resulting 5m resolution DEM is a reliable resource for applications like flood hazard mapping Additionally, the evaluation of parameter settings in the BS-Horizon model has enhanced the understanding of how parameters such as M, α, m1, and m2 influence surface generation These findings are crucial for optimizing the use of available data to accurately represent the topographic landscape in accordance with local conditions.
To effectively utilize the BS-Horizon program, it is essential to first analyze parameter M according to the established criteria Next, the selection of α should focus on achieving a balance between J(f) and R(f) Lastly, the values for m1 and m2 must be determined to complete the process.
To achieve optimal representation of the topographic surface, it is essential to appropriately set the parameters in BS-Horizon These parameters may vary for different datasets based on the data distribution and topographic features of the study area Therefore, it is crucial to thoroughly evaluate the suitable parameters while taking into account the specific characteristics of the area being studied.
Future research should focus on exploring the impact of parameter settings in regions with moderate to high relief, which will enhance our understanding of how local topography influences these settings This investigation will aid in developing adaptive parameter settings tailored to specific landscapes Additionally, the findings from this study will contribute to the creation of high-resolution Digital Elevation Models (DEMs) for assessing flood-prone areas across broader regions of Central Vietnam.
The BS-Horizon DEM at a 5m resolution has been utilized to extract geomorphological parameters for flood hazard zonation, focusing on elevation-based flood inundation (EFI) models that simulate flooding scenarios based on the area's elevation and proximity to river channels This study considers five key factors: EFI, distance to river channels, land use, slope, and the Topographic Wetness Index (TWI) to analyze flood hazard probabilities The Analytical Hierarchy Process (AHP) is employed to assign weights to each parameter and compute the Flood Hazard Index (FHI) through pairwise comparisons on a 9-point scale A consistency ratio (CR) is used to ensure the reliability of the assessments The flood hazard potential is classified into four categories—low, moderate, high, and very high—based on the FHI results.