- Open Access
Overcoming data scarcity in flood hazard assessment using remote sensing and artificial neural network
© The Author(s). 2019
- Received: 30 July 2018
- Accepted: 27 December 2018
- Published: 9 January 2019
Many urban cities in Southeast Asia are vulnerable to climate change. However, these cities are unable to take effective countermeasures to address vulnerabilities and adaptation due to insufficient data for flood analysis. Two important inputs required in flood analysis are high accuracy Digital Elevation Model (DEM), and long term rainfall record. This paper presents an innovative and cost-effective flood hazard assessment using remote sensing technology and Artificial Neural Network (ANN) to overcome such lack of data. Shuttle Radar Topography Mission (SRTM) and multispectral imagery of Sentinel-2 are used to derive a high-accuracy DEM using ANN. The improvement of SRTM’s DEM is significant with a 42.3% of reduction on Root Mean Square Error (RMSE) which allows the flood modelling to proceed with confidence. The Intensity Duration Frequency (IDF) curves that were constructed from precipitation outputs from a Regional Climate Model (RCM) Weather Research and Forecasting (WRF) were used in this study. Design storms, calculated from these IDF curves with different return periods were then applied to numerical flood simulations to identify flood prone areas. The approach is demonstrated in a flood hazard study in Kendal Regency, Indonesia. Flood map scenarios were generated using improved SRTM and design storms of 10-, 50- and 100-year re-turn periods were constructed using the MIKE 21 hydrodynamic model. This novel approach is innovative and cost-effective for flood hazard assessment using remote sensing and ANN to overcome lack of data. The results are useful for policy makers to understand the flood issues and to proceed flood mitigation adaptation/measures in addressing the impacts of climate change.
- Climate model
- MIKE 21
- Sentinel-2 multispectral imagery
Flood modelling is a useful tool for simulating/predicting water-related situations/disasters. In order to achieve a robust and reliable outcome of flood information, good quality data are required, such as a Digital Elevation Model (DEM) and hydrological/hydraulic data. Among those, DEM is crucial in modelling which reflects the actual topographic characteristics of the catchment (Kim et al. 2018). Bathrellos et al. (2016, 2017) conducted research on major factors affecting natural disaster including urban floods using topographic information such as slope, elevation and distance from streams. This quantitative geomorphological analysis was able to verify the past flood events well and this implies that DEM is a critical factor in flood assessment. High resolution and high accuracy of DEM has considerable influence in 2D flood modelling (Abily et al. 2015; Ngoc and Gourbesville 2016). Abily et al. (2016) conducted the sensitivity analysis of high resolution topography data in 2D flood modelling and emphasized that the water depth varies up to 1 m based on different resolutions in the study area. There are several sources to download the free DEMs such as the ASTER (Advanced Space-borne Thermal Emission and Reflection Radiometer), SRTM, and GTOPO30 (Global 30 Arc-Second Elevation). The SRTM at 30 m resolution was chosen for this study to derive finer resolution (20 m) data with improved DEM using the developed technique.
Unavailability of high-accuracy DEM and precipitation data are the major limitations of flood hazard assessment in many developing countries. The design of optimal mitigation measures including efficient stormwater systems request long records of hydrological data in order to generate the relevant rainfall IDF curves and to identify the high-risk areas (Liew et al. 2017). Kendal Regency, Indonesia, facing aforementioned data scarcity, is regularly facing flood issues and warrants a risk analysis. The best option for this analysis is to mobilize a 2D hydraulic model (e.g. MIKE 21) in order to simulate the flood propagation. As the study area suffers from lack of rainfall data, the methodology developed by Liew et al. (2012), which used climate model projected rainfall, has been implemented with rainfall data extracted from a high resolution regional climate model, Weather Research and Forecasting (WRF) driven by the European Reanalysis ERA-40 dataset, referred to in this paper as ‘WRF/ERA40’.
In this research, flood hazard assessment model was developed to construct the flood maps for Kendal in Indonesia. The publicly available SRTM30 DEM was improved using remote sensing imagery incorporated with machine learning technique. The Regional Frequency Analysis (RFA) using the WRF/ERA40 as proxy data method was applied in the study area to generate the IDF curves.
The Artificial Neural Network (ANN) is a useful tool to apply pattern recognition (classification) in engineering and scientific applications such as biology, medical science and remote sensing technology (Nguyen et al. 2018). This study utilizes the strength of pattern recognition in ANN to correct/improve the elevation in SRTM. As stated above, the limitation of scanning on the surface due to coarse resolution and penetration of canopy issue, SRTM has a tendency to overestimate or underestimate the elevation in forest and urban areas. The multispectral imagery supports to classify the different types of areas by its reflectance. The ANN is then trained using SRTM and multispectral imagery, together with the target data (which is the high accuracy DEM) to calculate the weights in different type of areas. Once the results are satisfied in the validation part, the trained ANN can be applied to other areas. The generated DEM through ANN, is then used as an input in the flood model with rainfall information derived from RCM output.
Kendal regency is under Central Java and has borders with Java Sea in the north, Semarang City in the east, Temanggung in the south and Batang Regency in the west. The northern region of the city is dominated by coasts, while in the south, lined by mountains. The study area is classified as Tropical Rainforest ‘Af’ based on Koppen-Geiger climate classification (Kottek et al. 2006). The average annual temperature is 20.2 °C and the yearly average rainfall is 3276 mm. The rainfall is lowest in August and highest in January.
lowland areas in the northern coast;
highland areas in the southern mountains.
In order to utilize the ANN, different types of remote sensing data are required. In this study, TanDEM-X (Wessel 2016), SRTM, surveyed high accuracy DEM and multispectral imagery from Sentinel 2 were used. As the SRTM is attributed with 30 m horizontal resolution in contrast to TanDEM-X (12 m), surveyed DEM (5 m) and Sentinel 2 (10–60 m), all input layers are standardized to 20 m resolution through sampling method.
For the Regional Frequency Analysis (RFA), proxy data from downscaled RCM (30 × 30 km over the study domain) (WRF/ERA40) were used.
The United States (US) National Aeronautics and Space Administration (NASA) released the SRTM data set for the globe (between 60° N and 56° S) at 3 arc-second resolution and 1 arc-second resolution for the US in 2003 (Jarvis et al. 2004). One arc-second resolution for the globe is available after 2015. The SRTM DEM is referenced to mean sea level with the Earth Gravitational Model 1996 (EGM 1996) geoid model. Hence, this study aimed at a higher resolution than its original, resampled to 20 m.
The German Aerospace Center (DLR) has been operating Germany’s first two formations flying Synthetic Aperture Radar (SAR) satellites, TerraSAR-X and TanDEM-X, with the objective to generate an updated global DEM. The DEM has a spatial resolution of 0.4 arc-second (≈ 12 m) with 2–4 m in relative vertical accuracy. The vertical datum of TanDEM-X is WGS84-G1150 ellipsoidal heights (Wessel 2016) and it has changed to geoid system to standardize the datum system with SRTM. As the DEM has higher accuracy and higher resolution than SRTM, it used for training of ANN after the resampling.
Sentinel 2 multispectral imagery
Sentinel 2 spectral bands
Band 1 – Coastal aerosol
Band 2 – Blue
Band 3 – Green
Band 4 – Red
Band 5 – Vegetation Red Edge
Band 6 – Vegetation Red Edge
Band 7 – Vegetation Red Edge
Band 8 – NIR
Band 8A – Narrow NIR
Band 9 – Water vapour
Band 10 – SWIR – Cirrus
Band 11 – SWIR
Band 12 – SWIR
WRF/ERA40 rainfall grid
Generation of improved DEM
The procedure of DEM improvement using ANN is divided into three major parts, which are training, validation, and application. ANN is a nonlinear mathematical structure model which is capable of representing an arbitrarily complex nonlinear process that relates the inputs and outputs of any system (Hsu et al. 1995). The ANN mimics the human’s brain with two primary features: the ability to ‘learn’ and ‘generalize’ from limited information (Hewiston and Crane 1994). Also, ANN is powerful in the analysis of remotely sensed data, particularly in the classification of land use and land cover (Manibhushan et al. 2011).
Input, target and output layers in ANN (example)
Improved DEM (m)
To be calculated
Generation of IDF curves
The RFA has been conducted using proxy rainfall data from WRF/ERA40. Hosking and Wallis (1997) developed a complete algorithm for the RFA method based on the approach of L-moments by pooling the sites with similar statistical characteristics in a homogeneous region instead of a single site in the at-site frequency analysis. The discordancy and heterogeneity measures are the primary indicators for accepting or rejecting the grid points from the study area. A discordancy measure is used to identify the sites with gross errors in the data or the data are grossly discordant with the region as a whole (Liew et al. 2014). Also, the heterogeneity is a statistical test to define the regional heterogeneity which is based on L-moments and on the theory that all stations of the region have the same population of L-moments, and then the studied region can be defined as homogeneous or not (Hosking and Wallis 1997). Figure 5 shows the selected 5 grid points which meet the criteria of RFA.
The selected grid points were used to construct the IDF curves for the study area. Design storms, calculated from the IDF curves of different return periods, were then applied to the numerical flood simulation.
Numerical model setup
The MIKE 21 Flow Model (MIKE 21) is a general numerical modelling system for simulation of water levels and flows (DHI 2017a). This modelling system is applied to coastal, marine engineering, ecological and inland flooding modelling projects. The water levels and flows are calculated on a rectangular grid which covers the area of interest with the information of bathymetry, bed resistance coefficients and hydrographic boundary conditions (Warren and Bach 1992).
A runoff model domain built with MIKE 11
2D hydrodynamic model domain at 20 m resolution due to hefty computational time.
MIKE 21 model setup
1228 km2 (805.4 km2 for runoff model; 422.6 km2 for 2D model)
Hourly tidal level extracted from Global Tide model
IDF Design Storms
20 × 20 m resolution for 2D model domain
Improved SRTM using TanDEM-X + ANN + Sentinel 2
10 h per one simulation
Design storm scenarios
Flood model simulation results
Inundation statistics corresponding to different scenarios for Kendal Regency
0.3 m – 0.5 m
0.5 m – 1 m
> 1 m
The areas of inundation with 0.3–0.5 m water depth are decreasing with higher return period, while the areas with greater than 0.5 m are increasing. 8.3% of the catchment is inundated with 10-year return period and 20% of the area is inundated with 100-year return period. The flooded area of 100-year return period with original SRTM shows less extent than improved SRTM at below 1 m depth due to less connectivity of DEM but it is much higher at above 1 m depth.
The flood hazard assessment has been conducted to address the flooding issues for Kendal Regency where challenges in obtaining high accuracy DEM and good quality with long recorded rainfall data. The improved SRTM was used as the topographic input which generated using ANN with Sentinel 2 multispectral imagery. The ANN was trained in the southern part of Malay Peninsula with TanDEM-X and Sentinel 2. The trained ANN was validated in Singapore and the improved SRTM is marked with a 42.3% RMSE reduction in general. As mentioned in 2.4, ANN was able to classify the land use and cover using the 8 bands of Sentinel 2. Based on the various land characteristics, different weights were calculated to reduce the error between the elevation of SRTM and TanDEM. For example, as shown in (Fig. 8f), the difference between SRTM and Improved SRTM, road areas have relatively larger differences and the areas with buildings have smaller differences. The resolution of DEM was thus able to increase from 30 m to 20 m. This makes the elevation map of the improved SRTM look clearer land shapes than the original SRTM. After the validation, it was applied to Kendal Regency, where surveyed DEM was not available. The improved SRTM and original SRTM for the study area were compared against satellite imagery as reference DEM is not available this area. The shapes of catchment’s surface were clearer in improved SRTM, especially river networks, roads and buildings compared to the original SRTM.
The RCM WRF driven by ERA40 produced the mean daily rainfall data from 1961 to 1990 over Southeast Asia domain and have been compared reasonably well with gridded observation data of CRU. The IDF curves then were developed based on the criteria of RFA. It was found that the derived IDF curves underestimate the existing IDF curves (station in Jakarta) within a range of 38% to 45% thus this study proposed to adjust the range of bias correction to the derived IDF curves.
The approach developed within this project presented an innovative and cost-effective flood hazard assessment. The freely accessible SRTM DEM was improved using remote sensing data and ANN. The developed information is significant and relevant to increase data accuracy allowing use of hydraulic numerical models. RFA was conducted using regional climate model downscaled precipitation data, as proxies, over the study area. The IDF curves with different durations and return periods were constructed as inputs for the numerical model. This method can be applied to areas lacking adequate rainfall records. The methodology and obtained results are significant for smart water management in the areas where data are not sufficient. The paper should be of interest to readers in the areas of remote sensing, artificial intelligence and flood management, especially for the policy makers in proposing relevant flood mitigation measures to the anticipated increasing flood damages, with higher confidence.
We are very grateful to Willis Towers Watson (UK), German Aerospace Center (DLR) and Nice Côte d’Azur Metropolis (France) for providing the data and for making this study possible.
Availability of data and materials
SRTM and Sentinel 2 are publicly accessible through the online. Authors would not share the surveyed DEM and TanDEM-X as those are under the non-disclosure agreement.
Mr. DK; Professor PG and Professor SYL coceived of the main idea for development of improved DEM methodology. DK and SYL desinged the framework and anlaysed the data. DK wrote the manuscript in consulation with PG. All authors read and approved the final manuscript.
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- Abily M, Bertrand N, Delestre O, Gourbesville P, Duluc CM (2016) Spatial global sensitivity analysis of high resolution classified topographic data use in 2D urban flood modelling. Environ Model Softw 77:183–195View ArticleGoogle Scholar
- Abily M, Delestre O, Amossé L, Bertrand N, Richet Y, Duluc CM, Gourbesville P, Navaro P (2015) Uncertainty related to high resolution topographic data use for flood event modeling over urban areas: toward a sensitivity analysis approach. ESAIM 48:385–399MathSciNetView ArticleGoogle Scholar
- Bathrellos GD, Karymbalis E, Skilodimou HD, Gaki-Papanastassiou K, Baltas EA (2017) Suitability estimation for urban development using multi-hazard assessment map. Sci Total Environ 575:119–134, ISSN 0048-9697. https://doi.org/10.1016/j.scitotenv.2016.10.025 View ArticleGoogle Scholar
- Bathrellos GD, Skilodimou HD, Chousianitis K, Youssef AM, Pradhan B (2016) Urban flood hazard assessment in the basin of Athens Metropolitan city, Greece. Environ Earth Sci 75(4):319View ArticleGoogle Scholar
- Coude, R (2018) Background on DEM. Retrieve from http://radiomobile.pe1mew.nl/?Geodata:Background_on_DEM Google Scholar
- DHI (2017a) MIKE 21 flow model and MIKE 21 flood screening tool – hydrodyamic module scientific documentation, published by DHIGoogle Scholar
- DHI (2017b) MIKE 21 toolbox – global tide model tidal prediction, published by DHIGoogle Scholar
- Drusch M, Del Bello U, Carlier S, Colin O, Fernandez V, Gascon F, Hoersch B, Isola C, Laberinti P, Martimort P, Meygret A, Spoto F, Sy O, Marchese F, Bargellini P (2012) Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sens Environ 120:25–36. https://doi.org/10.1016/j.rse.2011.11.026 ISSN 0034-4257. http://www.sciencedirect.com/science/article/pii/S0034425712000636 View ArticleGoogle Scholar
- Heidler LM (2015) Evaluation of different hydrologycal models in data scarce regions on the island of Ceram, Indonesia, Technische Universitat Munchen, GermanyGoogle Scholar
- Hewiston, Crane BC (1994) Precipitation Controls in Southern Mexico, in Neural Nets, Kluwer Academic PublisherGoogle Scholar
- Hosking JRM, Wallis JR (1997) Regional frequency analysis: an approach based on L-moments. Cambridge University Press, Cambridge; New York, xiii, p 224View ArticleGoogle Scholar
- Hsu K, Gupta HV, Sorooshian S (1995) Artificial neural network modeling of the rainfall-runoff process. Water Resour Res 31(10):2517–2530. https://doi.org/10.1029/95WR01955 View ArticleGoogle Scholar
- Jarvis A, Rubiano J, Nelson A, Farrow A, Mulligan M (2004) Practical Use of SRTM Data in the Tropics: Comparisons with Digital Elevation Models Generated from Cartographic Data. Working Document 198. International Center for Tropical Agriculture: Cali, ColombiaGoogle Scholar
- Kim D, Sun Y, Wendi D, Jiang Z, Liong SY, Gourbesville P (2018) Flood modelling framework for Kuching City, Malaysia: overcoming the lack of data. In: Gourbesville P, Cunge J, Caignaert G (eds) Advances in Hydroinformatics. Springer Water. Springer, SingaporeGoogle Scholar
- Kottek M, Grieser J, Beck C, Rudolf B, Rubel F (2006) World map of the Köppen-Geiger climate classification updated. Meteorol Z 15:259–263. https://doi.org/10.1127/0941-2948/2006/0130 View ArticleGoogle Scholar
- Liew SC, Liu J, Vu MT, Raghavan SV, Nguyen NS, Liong SY (2017) Constructing IDF curves using proxy data from downscaled reanalysis: Kendal regency (Indonesia), Willis Research NetworkGoogle Scholar
- Liew SC, Raghavan SV, Liong SY (2014) Development of intensity-duration-frequency curves at ungauged sites: risk management under changing climate. Res Lett Geosci Lett 1:8View ArticleGoogle Scholar
- Liew SC, Raghavan SV, Liong SY (2012) A novel approach, using regional climate model, to derive present and future IDF curves for data scarce sites. IAHR-APD 2012Google Scholar
- Manibhushan, Patel N, Sahoo G (2011) Artificial neural network: a tool for classification of land use and land covers using satellite images. Int J Comput Sci Inf Technol 4(2):55–59Google Scholar
- Ngoc DV, Gourbesville P (2016) Model uncertainty in flood modelling. Case study at Vu Gia Thu Bon Catchment-Vietnam, Procedia Engineering, vol 154, pp 450–458Google Scholar
- Nguyen K, Stewart R, Zhang H, Sahin O (2018) An adaptive model for the autonomous monitoring and management of water end use. Smart Water 3. https://doi.org/10.1186/s40713-018-0012-7
- Rodriguez E, Morris CS, Belz JE, Chapin EC, Martin JM, Daffer W, Hensley S (2005) An assessment of the SRTM topographic products, technical report JPL D-31639, Jet Propulsion Laboratory, Pasadena, California, p 143Google Scholar
- Uppala SM, Kållberg PW, Simmons AJ, Andrae U, da Costa Bechtold V, Fiorino M, Gibson JK, Haseler J, Hernandez A, Kelly GA, Li X, Onogi K, Saarinen S, Sokka N, Allan RP, Andersson E, Arpe K, Balmaseda MA, Beljaars ACM, van de Berg L, Bidlot J, Bormann N, Caires S, Chevallier F, Dethof A, Dragosavac M, Fisher M, Fuentes M, Hagemann S, Hólm E, Hoskins BJ, Isaksen L, Janssen PAEM, Jenne R, McNally AP, Mahfouf JF, Morcrette JJ, Rayner NA, Saunders RW, Simon P, Sterl A, Trenberth KE, Untch A, Vasiljevic D, Viterbo P, Woollen J (2005) The ERA-40 reanalysis. Quart JR Meteorol Soc 131:2961–3012View ArticleGoogle Scholar
- USGS, 2000, US GeoData digital elevation models, US geological survey, Reston, VA. Available online at: http://erg.usgs.gov/isb/pubs/factsheets/fs04000.html. Accessed 2 Feb 2009Google Scholar
- Warren IR, Bach HK (1992) MIKE 21: a modelling system for estuaries, coastal waters and seas. Environ Softw 7(4):229–240, ISSN 0266-9838. https://doi.org/10.1016/0266-9838(92)90006-P View ArticleGoogle Scholar
- Wendi D, Liong SY, Sun Y, Doan CD (2016) An innovative approach to improve SRTM DEM using multispectral imagery and artificial neural network. J Adv Model Earth Syst 8. https://doi.org/10.1002/2015MS000536
- Wessel, B (2016) TanDEM-X ground segment – DEM products specification document, EOC, DLR, Oberpfaffenhofen, Germany, Public Document TD-GS-PS-0021, Issue 3.1. https://tandemx-science.dlr.de/ Google Scholar