Predicting land use effects on flood susceptibility using machine learning and remote sensing in coastal Vietnam

Floods affect infrastructure and the environment, not just the social economy (Khosravi et al. 2019; Islam et al. 2021). Between 1995 and 2015, approximately 150,000 flood events happened worldwide and caused around 157,000 deaths (Islam et al. 2021). Several studies point out that flooding affects approximately 250 million people and cause USD 40 billion in losses each year worldwide (Nguyen 2022a; Nguyen et al. 2022b).

Fluvial, flash, coastal, and urban floods can be classified on the nature of their occurrence. Floods are one of the most devastating natural events, causing damage to human life and infrastructure due to their onset attributes and rapid flow speeds (Islam et al. 2021). In developing countries, flood damage often has more impact than in developing countries due to climate change (Janizadeh et al. 2021). Climate change can lead to weather conditions outside predicted capacity: changes in precipitation and/or heat wave frequency and increases in the number and intensity of extreme weather events such as typhoons or floods (Janizadeh et al. 2021). Flood frequency change may be caused by changes in climatic conditions – e.g., precipitation and/or temperature – and may be apparent in low-lying, riverside, coastal, and urban areas. Flooding occurs not only under extreme climatic conditions but also through the effects of changing ecosystems. In the global scenario, human activities such as deforestation, dam construction, population, and/or urban growth can all change flood intensity (Junger et al. 2022; Pal et al. 2022; Roy et al. 2022). Since the ‘Doi Moi (Innovation)’ renewal policy in 1986, Vietnam has embarked on sustained urban growth. This occurs more rapidly in coastal regions, where it is often affected by typhoons or flooding. Therefore, improving flood risk assessment using state-of-the-art methods can reduce flood risk (Nguyen et al. 2022a).

Four factors, namely topographic, climatic, hydrologic, and human activity cause floods. Topographic factors include elevation, slope, curvature, and aspect, which are relatively stable (Ghosh et al. 2022; Prasad et al. 2022). Climatic and hydrologic factors, precipitation, and drainage networks are dynamic factors because they change every year. Chen et al. (2022) report that climate change, particularly the modification of precipitation, significantly influences the probability of flooding occurring and the speed of flow. At the same time, human activity like land use and land cover change (LULC), road distance, and distance to settlement are considered strong dynamic factors, especially in LULC. Several studies note that human activity can lead to changes in the flow and speed of floods, with the risk of flooding becoming more severe. Deforestation can influence the mechanical disturbance capacity of soils. Changes in the type of land cover have an important effect on the soil's infiltration capacity (Nguyen et al. 2022b; Sugianto et al. 2022). All of this can influence the likelihood of flood occurrence, so assessing the effects of climate change and LULC on flood susceptibility is a critical task in understanding the causes of change in flood susceptibility, which can support decision-makers in proposing appropriate measures to reduce flood damage.

Previous studies using different methods to assess the effects of climate change and LULC on flood susceptibility can be grouped into four main approaches: physical models, remote sensing, and statistical and machine learning models. Physical models include Mike Flood (Mangukiya & Yadav 2022), HEC-RAS (Afzal et al. 2022; Shah et al. 2022). Soil and Water Assessment Tool (SWAT) (Liu et al. 2022) is considered promising as it can simulate floods with details such as flow velocity, flood depth and duration. However, physical models require deep understanding of hydrologic processes to adjust parameters, which poses challenges for researchers in model development. These models also require detailed in situ data, limiting their application to data-rich regions. Over past decades, remote sensing data, including synthetic aperture radar (SAR) data and optical sensor images combined with GIS, have been used to assess floods (Dewan et al. 2006; Alarifi et al. 2022). These methods have been used to build flood occurrence lists. Many different studies use remote sensing databases to assess floods hazards, combined with statistical techniques such as frequency ratio (FR), logistic regression (LR), weights of evidence (WoE model), fuzzy logic (FL) (Nandi et al. 2016; Rahmati et al. 2016; Sepehri et al. 2020; Ullah & Zhang 2020). These techniques can process spatial data to predict the likelihood of flood occurrence with high accuracy.

Several researchers have applied machine learning to develop flood susceptibility models. These include artificial neural networks (ANNs), random forest (RF), support vector machine (SVM), LR, and decision trees (DT) (Tehrany et al. 2014; Al-Juaidi et al. 2018; Chen et al. 2020; Khoirunisa et al. 2021; Wang et al. 2021). However, flood susceptibility model development using machine learning has several challenges, like selecting appropriately from the hundreds of different models and each model providing different results. Therefore, testing the models to assess flood susceptibility with different regions and types of natural hazards is recommended.

This study's objective was to test and compare machine learning models in assessing flood sensitivity in the context of climate change and land use in Thua Thien Hue province, Vietnam. This is the first time flooding has been assessed in relation to climate change and LULC in the province, which is one of those most affected by climate change and LULC in Vietnam. The key contributions of this study are: (i) selection of appropriate models to determine flood susceptibility zones; (ii) construction of a flood susceptibility map at different periods and (iii) evaluation of the effects of climate change and LULC on flood susceptibility.

The study area has a dense hydrologic system, comprising the rivers Perfume, Bo, and Ong Lau. The province also has two lagoons, Tam Giang – Cau Hai and Lang Co. The study area's climate has two main seasons: rainy and dry. The rainy season, from September to December, accounts for 70% of annual precipitation. Rainfall during this season is often brief but intense, leading to floods that can cause significant socio-economic and environmental damage.

Thua Thien Hue province is frequently affected by storms. According to the Ministry of Natural Resources and Environment, the study area is impacted by approximately 30% of the storms that occur in the country.

Flood inventory maps are vital in developing flood susceptibility models. They present records of past flood events, and can be used to predict the probability of flood occurrence in future by evaluating the relationships between past events and conditioning factors (Ghosh et al. 2022; Yaseen et al. 2022). In this study a flood inventory map was constructed using flood marks measured in 2020 and 2021. Sentinel 1A images from October 16, 2020 were also used to determine flood points, to ensure the flood susceptibility model's reliability. In total, 986 flood points were collected to build the model. Because the proposed model is classified as binary, non-flooding points must also be collected, for example, in the high elevation and slope regions, never affected by floods.

In the end, 1,864 flood and non-flood points were divided into two groups: 70% as training data to train the models and 30% as validation data to evaluate the models.

Elevation, aspect, curvature, slope, and TWI were calculated using a DEM generated from the 1:50,000 scale topographic map (available from Vietnam Department of Survey, Mapping and Geographic information (https://www.bandovn.vn/en/topographic-maps-7)). NDVI and NDBI in 2017 and 2021 were calculated from Landsat 08 images. Land use in 2017 and 2021 were downloaded from https://www.arcgis.com. Distance to river and distance to road were extracted from the 1:50,000 scale topographic map using Line Density in Arcgis 10.6. Rainfall in 2017 and 2021 were downloaded from https://chrsdata .eng.uci.edu/. (All factors were transformed into raster format at 10 m resolution.)

SVM was developed by Vapnik Research in 1965 but became an algorithm in 1995 (Cortes & Vapnik 1995). SVMs are initially linear and binary, based on maximizing the separation margin between two classes to predict a binary qualitative variable. Kernel functions, however, mean that transformations are now possible for multiclass classification. SVM's basic principle is to divide a dataset into two subsets by a separating hyperplane (Tehrany et al. 2014, 2015). The goal is to determine an optimal hyperplane to maximize the distance to the nearest training data (known as support vectors). This distance is the ‘margin’.

The accuracy of SVM depends on the setting of C and Gamma. The higher the value of C, the more precise the data point separation. On the other hand, if the value of Gamma is significant, only points near the dividing line will be used in calculation.

RF, developed by (Breiman 2001), can solve both classification and regression problems, and is a combination of a bagging algorithm and random subspace. RF works by constructing many DT following three main steps:

• (1)

  Using the bootstrap method to build a data subset of similar size to the original dataset.

• (2)

  Making the DT based on subsets.

• (3)

  Predicting the outcome by majority voting.

RF can handle multidimensional data and can also work well for missing data (Chen et al. 2020). Its performance depends on the number of decision trees (Ntree) and the number of variables used to build the tree (mtry).

Assessing the model's accuracy is a critical step in building a flood susceptibility model. If the model performs well, it can be generalized to assess flood susceptibility elsewhere. In this study, statistical indicators including receiver operating curve (ROC), RMSE, and MAE, were used to assess data learning and validation performance.

The ROC curve shows a model's ability to predict the occurrence of floods correctly and is based on the true positive rate (sensitivity) and false-positive rate (1–specificity). The true positive rate is the proportion of actual positives that are correctly identified as positives by the model, while the false-positive rate is the proportion of actual negatives that are incorrectly identified as positives by the model. The areas under the ROC curve reflect model accuracy.

RMSE and MAE measure the differences between the predicted and observed values (Wang et al. 2019; Nguyen 2022a).

Table 2 gives RMSE and MAE for the flood susceptibility models. Clearly, the RF model performs better than the SVM one in terms of both, with values of 0.21 and 0.06, respectively, compared to the SVM model's 0.22 and 0.08 during training. The same was true in validation, the RF model (RMSE = 0.22, MAE = 0.09) outperformed the SVM model (0.23 and 0.12).

Although several different methods have been developed and used to predict the spatial distribution of flood susceptibility, assessment is still considered an obstacle to effective flood risk management (Gharakhanlou & Perez 2023; Wu et al. 2023). Floods are caused mainly by the distribution and intensity of precipitation and are often amplified by topographic conditions and landscape characteristics, including water infiltration capacity. Both climate and LULC are expected to influence flood susceptibility in the future.

The results of this study confirmed that altitude, NDBI, NDVI, distance from the river, and rainfall are the most critical factors for the probability of flood occurrence in Thua Thien Hue province. Altitude is important because it is directly related to water storage capacity. In Thua Thien Hue province, floods occur mainly in the lower altitude, eastern region. The importance of NDBI and NDVI came second and third for flood susceptibility because they influence water infiltration capacity directly. For example, in the study area, the increase in built-up area and decrease in area of vegetation are thought to have changed the spatial distribution of floods from 2017 to 2019. Distance from the river was the fourth factor of importance for flood occurrence, because floods in the study area are fluvial, occurring near the river. Rain was only the fifth factor also because floods in the study area are fluvial, occurring at low altitude and due to LULC. Previous studies justify all this. (Nguyen 2022b) found that elevation and land use were critical factors in flood occurrence in Central Vietnam and Ha Tinh province. Ha et al. (2022) indicated that elevation, rainfall, and distance to river were the most important factors for flood susceptibility in Thua Thien Hue province, Vietnam.

In this study, two popular models, SVM and RF, were proposed to assess the effects of climate and LULC change on flood susceptibility in Thua Thien Hue province, Vietnam. The results show that with its AUC value of 0.98, the RF model outperformed the SVM model (AUC = 0.97) with respect to the effects of climate and LULC change on flood susceptibility. RF is one of the most effective and widely used, existing models, because it is the most easily explained (or interpreted) of the models. In addition, it is more transparent in relation to the use made of training data. RF also has the advantage of using all of its initial data more intelligently, in order to limit its errors (Pal 2005; Uddin & Uddiny 2015). While SVM has the advantage in solving overfitting problems, however, it is not suitable for nonlinear problems (Karamizadeh et al. 2014).

The models proposed in this study use RMSE as an objective function, because of which, the overfitting problem is seen as an important challenge, as it is for machine learning in general. To reduce the problem's effects, several techniques were applied in this study, such as setting a dropout rate and limiting the search range. However, other methods can also be tried, such as collecting training data from different geographical locations (Nguyen et al. 2020).

The study's results also confirmed that the very high flood susceptibility surfaces, concentrated mainly in the east where there were major changes in land use, increased from 2017 to 2021 in the study area. This trend is common worldwide. Nguyen et al. (2022a) show that flood susceptibility will increase in the Nhat Le – Kien Giang watershed (Vietnamese central region) will increase from 2005 to 2050 due to land use and climate change. Hagos et al. (2022) report that flooding will become more frequent due to climate change and land use change in the Awash River basin in central Ethiopia. Nirupama & Simonovic (2007) show the increase in flood risk in the province of Ontario in Canada due to urban growth. This was confirmed by the study by Hettiarachchi et al. (2018), who highlighted the increased flood risk in the South Washington Watershed District (SWWD) in Minnesota, United States, related to climate change and urban growth. This study provides additional evidence for the effects of climate change and change in land use on flood susceptibility.

The objective of this study was to evaluate flood susceptibility to the impact of climate change and LULC, using machine learning and remote sensing models in relation to Thua Thien Hue province, Vietnam.

The proposed models use 1,864 flood locations and 11 condition factors from 2017 to 2021 as input data. The results show that elevation, NDVI, NDBI, distance to river and rainfall were the most important factors for flood occurrence in the study area.

The combination of machine learning techniques (SVM and RF) with remote sensing successfully assessed the impact of climate change and LULC on flooding, with AUC values exceeding 0.95. The method can be generalized to determine flood susceptibility elsewhere, particularly in data-limited areas.

Very high flood susceptibility areas increased by about 30% from 2017 to 2021 in Thua Thien Hue. The changes are related to the increased construction area – the surfaces are concentrated mainly in the east – and precipitation.

This work is supported by national research program from MOST (Vietnam), project number ĐTĐL.CN-93/21.

All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by V.T.V., H.D.N., P.L.V., M.C.H., V.D.B., T.O.N., V.H.H., T.K.H.N. The first draft of the manuscript was written by V.T.V. and H.D.N. All authors read and approved the final manuscript.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.