Development of a Deep Learning-Based Flooding Region Segmentation Model for Recognizing Urban Flooding Situations

Abstract

Urban flooding has become increasingly frequent due to the rising intensity of rainfall driven by urban development and climate change. Effective prevention measures are crucial to mitigate the significant human and material damages caused by such events. Rapid and accurate pre-detection techniques can help to reduce the impacts of urban flooding. With the advancement of deep learning, deep neural networks (DNNs) have been successfully applied across various domains, including computer vision and speech recognition. In particular, DNNs for computer vision demonstrate high performance with relatively low computational costs. In this paper, we propose a flooding region segmentation model for urban underpasses based on the U-Net architecture. To train and evaluate the model, we collected datasets from the Mannyeon, Oryang, and Daedong underpasses in Daejeon. The proposed method achieved Dice coefficients of 98.8%, 94.03%, and 93.85%, respectively. This model demonstrates high segmentation performance in detecting flooded regions and can be integrated into continuous flood monitoring systems.

1. Introduction

South Korea has been facing increasing flood damage due to extreme rainfall events exacerbated by climate change and rapid urban development [1,2,3]. Notable incidents include severe flooding in critical areas such as Gangnam Station, Gwanak-gu in Seoul, and Daejeon. One significant incident occurred on 16 July 2017, when Cheongju-si in Chungcheongbuk-do experienced torrential rainfall, reaching a maximum of 290.2 mm in a single day [4]. In August 2020, the upper middle watershed of the Geum River experienced an average rainfall of 297.3 mm, resulting in catastrophic damage along its banks. More recently, on 8 August 2022, unprecedented rainfall at a rate of 90 mm per hour flooded the Dorimcheon area located in Seoul, leading to 3 fatalities and 447 victims, with the estimated losses reaching 7.8 billion KRW. In 2023, heavy rainfall resulted in over 220 billion KRW in property damage across nine local governments, including Cheongju and Goesan-gun [5]. On 16 July 2023, in Osong-eup, Cheongju-si, Chungcheongbuk-do, river overflow led to the inundation of an underpass, resulting in nine fatalities and nine injuries [6]. Urban flooding has caused widespread damage, particularly affecting underpasses, underground facilities, and semi-basement dwellings, which are especially vulnerable to such disasters. Research on flood risk and mitigation strategies for these underground spaces is progressing [7,8]. For instance, Busan Metropolitan City has developed an inundation forecast map to improve flood preparedness. Additionally, public notifications are issued to alert citizens when flooding is anticipated.

The Korea Institute of Science and Technology Information (KISTI) generated over 20 scenarios through two-dimensional inundation analysis for all drainage systems in Incheon, utilizing these scenarios for inundation prediction. Urban flooding was simulated using predicted rainfall and the SWMM (Storm Water Management Model) [9]. A strategy was proposed to predict the flooding range of the watershed using the SWMM discharge simulation results [10]. For the results of previous studies, flood prediction strategies have been proposed using a physical model. However, since urban flooding proceeds rapidly, flooding prediction models using a physical model are difficult to use in real time.

Research on urban flood prediction and response has evolved significantly, encompassing both physical model-based and deep learning-based methodologies. Early studies predominantly relied on physical models for flood prediction strategies, utilizing rainfall and runoff simulation data. An approach that predicted urban flooding using forecasted rainfall and the SWMM (Storm Water Management Model) was demonstrated [9]. Similarly, a method for predicting the flooding range of watersheds using discharge simulation results from the SWMM was developed [10]. While these studies effectively leveraged physical models, they highlighted limitations in real-time application due to the rapid nature of urban flooding.

The Korea Institute of Science and Technology Information (KISTI) developed over 20 scenarios using two-dimensional inundation analysis for Incheon’s drainage systems, aiding in flood prediction efforts. These approaches underscore the foundational reliance on physical models for early urban flood predictions, which, while precise, face challenges related to real-time adaptability.

The advent of deep learning has brought transformative changes across various fields, including urban flood management, due to advances in computational power and model sophistication. This evolution has enabled deeper and more complex models to achieve higher accuracy across tasks. For example, an artificial neural network (ANN)-based prediction model was constructed for urban flood thresholds, effectively training the model using extensive, augmented datasets to forecast regional flooding [11].

Advancements in urban flood detection and quantification have been significantly propelled by systems like V-Flood Net, a video segmentation tool that enhances real-time detection and measurement of flood-affected areas. This system, along with real-time water level monitoring using live camera feeds combined with computer vision techniques, exemplifies how modern technology can bridge the gap between data collection and actionable insights. These methodologies facilitate prompt decision making in urban flood response, allowing for more efficient allocation of emergency resources.

Deep learning and computer vision have also been integrated into river flood detection systems, providing automated and highly accurate analysis of water level changes and flood extents. A near-real-time flood detection method that leverages deep learning with SAR (synthetic aperture radar) images has further demonstrated the potential for rapid and precise identification of flooded areas, emphasizing the value of these techniques in minimizing disaster impacts. Enhanced flood detection through advanced deep learning models, which focus on precise water segmentation, shows marked improvements over traditional detection methods, ensuring more reliable urban flood mapping.

Convolutional neural networks (CNNs) have been pivotal in developing automatic flood mapping solutions using data from satellite sources such as NovaSAR-1 and Sentinel-1. The deployment of fully convolutional neural networks (FCNs) for rapid flood segmentation in SAR imagery highlights the models’ ability to process complex datasets and deliver timely outputs, which is crucial for real-time flood monitoring. Near real-time flood mapping supported by weakly supervised machine learning emerges as a flexible approach, requiring less annotated data and facilitating the quick adaptation of models to different flood-prone regions.

Multi-class segmentation of urban floods using deep learning and multispectral imagery underscores the advancements in differentiating between various types of flood-impacted areas. This approach provides more detailed flood assessments, supporting targeted mitigation strategies and comprehensive urban planning. These methodologies collectively illustrate the shift toward leveraging deep learning and computer vision to enhance flood detection and the accuracy, speed, and scalability of prediction models, contributing significantly to urban flood management and resilience [11,12,13,14,15,16,17,18,19,20].

Advancements in flood prediction have been marked by the development of data-driven models based on deep learning techniques, which offer accurate and rapid predictions across temporal and spatial dimensions. These models leverage complex algorithms to analyze large datasets, enhancing the precision and speed of flood forecasting. The application of deep learning extends to river flood predictions, focusing on estimating flood depth and extent to aid in resource allocation and timely response. Urban flood modeling in cities such as Seoul, South Korea, showcases the effectiveness of deep learning approaches in adapting to specific urban topographies and varying rainfall conditions, demonstrating the adaptability and scalability of these techniques.

Progress in urban flood simulation research has shown a significant transition from traditional numerical models to more advanced deep learning methods. These modern approaches overcome many of the limitations of older models, such as computational inefficiency and slower response times. Experiments using deep learning techniques for flood propagation illustrate their capability to accurately simulate water movement in urban areas, supporting better emergency planning and mitigation strategies. Machine learning is also employed to assess surface water flood risks in urban settings, integrating environmental and infrastructural data to produce comprehensive risk assessments. This integration of machine learning enhances the ability to process and analyze complex datasets, leading to more informed urban planning and flood risk management [21,22,23,24,25,26,27].

Developments in flood modeling have introduced frameworks utilizing deep learning with spatial reduction and reconstruction, enhancing the speed and efficiency of inundation simulations. Such models are designed to process complex spatial data rapidly, making them highly suitable for real-time flood prediction and response. Deep learning-based super-resolution techniques have also enabled high-resolution urban flood analysis, transforming lower-resolution inputs into detailed models that support precise urban flood management and mitigation strategies.

Research on flood mapping has highlighted the potential of deep learning methods through a review of current applications and prospective advancements. Flood detection in urban areas using multispectral satellite imagery combined with machine learning algorithms has demonstrated improved accuracy in identifying flood-affected regions. Integrating advanced algorithms and diverse data sources allows for more comprehensive flood assessments and faster response times. Real-time urban flood forecasting and modeling have emerged as state-of-the-art practices, combining multiple data inputs and machine learning for more immediate and reliable flood predictions.

Integrating deep learning with geometric characterization and hydrodynamic modeling provides insights into assessing the impact of sewer defects on urban flooding, as exemplified by case studies conducted in areas like Guangzhou, China. Additionally, urban flood vulnerability and risk mapping, facilitated by combining multi-parametric AHP (analytic hierarchy process) with GIS (geographic information systems), offers a structured approach for assessing and visualizing flood risk. This combination enhances urban planning by integrating various environmental and infrastructural factors, supporting more effective flood mitigation efforts.

Enhancing flood risk assessment in urban areas through the integration of hydrodynamic models and machine learning techniques has proven to be a robust approach for achieving detailed risk analyses. Interpretable machine learning models have been developed to predict urban flash flood hotspots by intertwining land and built-environment features, enabling more transparent and actionable insights for decision makers. The creation of frameworks for modeling flood depth using a hybrid approach of hydraulics and machine learning showcases the potential for more precise and adaptive flood management solutions, bridging the gap between theoretical modeling and practical application in urban environments [28,29,30,31,32,33,34,35,36,37].

Flood detection and monitoring advances have increasingly utilized deep learning models for improved water segmentation. The application of semantic segmentation and multimodal data fusion in flood detection has demonstrated the potential to enhance the accuracy of identifying and analyzing flood-affected regions. Fully convolutional networks (FCNs), which form the basis of many semantic segmentation tasks, have proven effective in parsing complex visual data for quick and reliable flood detection. The U-Net model, originally developed for biomedical image segmentation, has been adapted for environmental monitoring, showcasing its versatility and high performance in detecting flooded areas.

Integrating interpretability into deep semantic segmentation models has emerged as a focus in earth observation research, allowing stakeholders to understand model outputs more effectively. An interpretable approach ensures decision-makers can trust and act on automated flood detection insights. In this context, novel AI-based models that incorporate super-resolution generative adversarial networks (GANs) for real-time flood image recognition have been developed. These models enhance the clarity and precision of flood monitoring tools, enabling timely interventions during flood events.

Flood forecasting and segmentation efforts have also expanded to include real-time and location-specific solutions. For example, water level forecasting using deep learning, as demonstrated in case studies like the Trinity River in Texas, illustrates how predictive modeling can be tailored to specific regions. The development of flood nomographs for urban district inundation forecasting has further contributed to the practical application of these models, helping cities prepare for potential flooding through easily interpretable forecasting tools.

Research on flood management technologies has reviewed various approaches related to image processing and machine learning. Methods for extracting and classifying flood-affected areas based on Markov random fields (MRF) and deep learning showcase how data-driven approaches can enhance the accuracy of flood impact assessment. This combination supports comprehensive flood management by swiftly identifying affected regions and informing effective response strategies [38,39,40,41,42,43,44,45,46,47,48]. Summary of similar previous studies is shown in

Table 1. Summary of similar previous studies.

Study	Task	Model
Real-time water level monitoring using live cameras and computer vision techniques (2021) [12]	The prediction of water levels from images with the segmentation model	FCN
A Near-Real-Time Flood Detection Method Based on Deep Learning and SAR Images (2023) [14]	Flooding detection with SAR images and a deep learning model	U-Net
Enhanced Flood Detection Through Precise Water Segmentation Using Advanced Deep Learning Models (2024) [15]	Flooding region segmentation	SegNet, U-Net, FCN-32s
A deep-learning-technique-based data-driven model for accurate and rapid flood predictions in temporal and spatial dimensions (2023) [21]	Flooding prediction on simulated flood maps	LSTM
A rapid flood inundation modelling framework using deep learning with spatial reduction and reconstruction (2021) [28]	A rapid flood inundation modeling framework	LSTM
Water Segmentation with Deep Learning Models for Flood Detection and Monitoring (2024) [38]	Pixel-wise flooding region segmentation	SegNet, U-Net, FCN-32s, PSPNet
Flood Detection using Semantic Segmentation and Multimodal Data Fusion (2021) [39]	Real-time flooding detection with semantic segmentation	U-Net

.

In this paper, we propose a strategy to detect flooding regions using a deep neural network and surveillance camera images of underpasses. The dataset to train and test the model is acquired from the surveillance video. A deep neural network for segmentation is proposed as a model based on the U-Net architecture introduced earlier.

The remaining parts of the paper are organized as follows: Section 2 presents the materials and methods in the following subsections: proposed strategy for flood-prone region segmentation (Section 2.1) and dataset acquisition for neural network training (Section 2.2). Section 3 describes the details of the experiment in this study. The results are presented in Section 4. Finally, the concluding remarks of this study are discussed in Section 5.

2. Materials and Methods

This section describes our proposed methodology and dataset for training and testing the deep neural network. Our main objective was to observe and detect flooding regions. Hence, we searched the deep neural network, which performs remarkably in detecting specific areas in images. We also collected the dataset, including the flooding scenes.

2.1. Proposed Strategy for Flood-Prone Region Segmentation

To detect the flooding region caused by water-related accidents, such as heavy precipitation and water overflow in rivers, we needed a strategy to automatically detect the water region in the underpass using water segmentation of the images obtained from videos taken by a surveillance camera placed in the underpass. The most important point in detecting tasks is increasing the accuracy of pixel-level segmentation. We needed the semantic segmentation model to achieve our main objective of detecting the flooding region accurately. Semantic segmentation is a computer vision task that classifies each pixel in the image into a category or object. Semantic segmentation is commonly used to analyze images in computer vision and image processing tasks. The algorithm for semantic segmentation is usually composed of allocating a label to each pixel and acquiring a set of regions in the output.

A lot of segmentation models based on deep learning have been proposed in recent years. In image analysis, the architecture of a deep neural network based on CNN shows outstanding performance. From this concept, a model consisting of only convolutional layers was proposed, called a fully convolutional network (FCN). A CNN model usually has the fully connected layer (FC layer). However, the FC layer has several limitations in improving the performance of segmentation. First, the input size of the model has to be fixed. Since the FC layer is composed of a few dense layers, the number of weights and feature maps is fixed. This problem makes it challenging for the model to be flexible in training various datasets. Second, the FC layer loses the location information of the input data. The main goal of segmentation is categorizing the pixels and separating the foreground from the background. For this point, losing location information has a negative effect on segmentation performance.

The FCN solves these problems by replacing the FC layer with the 1 × 1 convolutional layer. The 1 × 1 convolutional layer makes the network able to take the arbitrary input size and produce the correspondingly sized output, encouraging efficient model training. In addition, by introducing the concept of skip architecture, more accurate and detailed segmentation information is obtained by combining the layers, which are deep, coarse, and contain shallow appearance information. The FCN lays the foundation of the end-to-end semantic segmentation model using the fully convolutional layer and influenced the other models that were proposed afterward.

As FCN is proposed, a lot of models consisting of two sub-networks, an encoder, and a decoder have been proposed. One of the popular models is U-Net. Since U-Net has the encoder-decoder architecture for semantic segmentation, this model operates like an FCN, obtaining the features using convolution, up-sampling the image in the back part, and learning the feature representation. The encoder, also called the contracting path, extracts the features from the image. The decoder, also called the expanding path, restores the down-sampled input image.

U-net shows improved performance compared to FCN with a smaller training dataset. U-net introduces the concept of skip connection, in which the feature from the encoder is overlapped with the decoder. This process produces more detailed localization information and high-resolution output. Furthermore, U-Net uses the overlap-tile strategy, which sees the feature map in only a specific area. The computation cost is effectively decreased since the missed area is extrapolated by mirroring. This strategy is appropriate to be applied to large size images without the limitation of resolution. With this strength, U-Net shows reliable results in the segmentation task, which needs more detail, such as cell segmentation in medical images. Considering these advantages of U-Net, we decided to take this network architecture for the flooding region segmentation model.

Figure 1 shows the flooding region segmentation model. The proposed flooding region segmentation model used an architecture based on a conventional U-Net. This model consists of the encoder path, which extracts the features of input data, and the decoder path, like a conventional U-Net.

In the encoder path, each block has two successive convolution layers, subsequently a max-pooling layer. In the decoder path, each block has two successive convolution layers followed by an up-convolutional layer. Each convolution layer (the yellow arrows) uses 3 × 3 kernels and the rectified linear unit (ReLU) for nonlinear mapping. The max-pooling (the blue arrows) uses 2 × 2 kernels to down-sample the input data by a factor of two. We used a padding of 0 to maintain the original input size. The extracted features from each block in the encoder path are copied and concatenated (the sky-blue arrows) to the corresponding blocks in the decoder path. In the decoder path, each up-convolution layer (the red arrows) uses 2 × 2 kernels, followed by ReLU.

As shown in Figure 1, the input data are the images taken from the underpass. The dimensions of the input data are 256 × 256 × 3. The output data correspond to the flooding regions from the input data. The dimensions of the output data are 256 × 256 × 1.

2.2. Dataset Acquisition for Neural Network Training

In order to train the model, we acquired 3 video datasets, which were taken from the underpass of Mannyeon, Oryang, and Daedong in Daejeon on 30 July 2022. On that day, many roads were flooded due to heavy rain across the country. These videos contain scenes of flooding in the roadway and sidewalk of the underpass due to precipitation. Since each underpass has a different structure, the flooding regions are also various. This information on flooding regions makes the segmentation model training more effective. Scenes from each underpass video are shown in Figure 2.

There are several steps required to obtain the training and test datasets. Figure 3 shows the overall process of acquiring an image dataset.

First, since the input dataset for the proposed model is the images, we obtained the images from the original video dataset by randomly capturing each video. The pace at which flooding proceeds varies depending on the topographical characteristics of the underpass or the precipitation in the area. To obtain as many images, including the flooding regions, as possible, we randomly captured them by focusing on the time during which flooding was actively in progress. For that reason, the number of images for each underpass video is different, as shown in

Table 2. The detailed information of the acquired image dataset.

Underpass	Quantity	The Size of Image
Mannyeon	124	1080 × 1920
Oryang	244	1080 × 1920
Daedong	173	1080 × 1920

. Then, the flooding regions in the images were labeled by a human annotator. Each image had the corresponding ground truth that was of a two-dimensional image. The ground truth is the binary matrix comprising the zero pixels and the one pixels. The zeros for pixels represent the background information, and the ones for pixels represent the flooding region information. With this process, the dataset was acquired for the images and the corresponding ground truth. Figure 4 shows the image and ground truth for each underpass. Table 2 shows the details of the dataset.

3. Experiments

We split the dataset into training and testing sets, as shown in

Table 3. The number of images in the training set and test set.

Underpass	Training Set	Test Set
Mannyeon	110	14
Oryang	214	30
Daedong	172	7

. We used the datasets to train the model and evaluate the performance of the model, respectively. The scenes in the images were nearly similar, as the location of the camera in the underpass was fixed. Since the images, including the various flooding situations, were selected as the test set, the number of images per underpass dataset is different.

To input the images into the model, the images were normalized from 0 to 1 using min–max normalization. The images had 3 dimensions and RGB channels, so the maximum value was 255 and the minimum value was 0. The normalization equation is expressed as follows:

$$min-\max normalization={\displaystyle \frac{X-X_{min}}{X_{max}-X_{min}}}$$

(1)

where $X$ denotes the image pixel value. Then, the dataset, including the ground truth, was resized to 256 × 256, the same as the input data size of the model.

We implemented the proposed model in Pytorch 1.7.1 and trained it for 100 epochs with 16 batch sizes, which took 19 h on a graphic processing unit (NVIDIA GeForce RTX 3080 Ti : NVIDIA, Santa Clara, CA, USA). The training was performed using the ADAM optimizer with a learning rate of 1 × 10⁻⁴. Mean squared error (MSE) was used as the loss function of the network. The networks were trained separately for each underpass dataset, so a total of 3 models were produced.

To evaluate the model, we used the Dice coefficient as the performance metric. The Dice coefficient is usually used in image segmentation tasks. This metric measures the similarity between the two images. The coefficient ranges between 0 and 1, where 1 demonstrates that the two sets are identical, and 0 demonstrates that the two sets have no overlap. The Dice coefficient is given by the following equation:

$$DiceCoefficient={\displaystyle \frac{2\times \left|A\cap B\right|}{\left|A\right|+\left|B\right|}}$$

(2)

where |$A$| denotes the number of elements in set $A$, and |$B$| denotes the number of elements in set $B$. $\left|A\cap B\right|$ denotes the number of elements that exist in both sets. The Dice coefficient is useful for imbalanced datasets where one set is much larger than the other using the weights for the area of of the intersection of two images. An image also represents this equation, as shown in Figure 5.

4. Results

4.1. The Resultants of the Flooding Region Segmentation Model Based on U-Net

To verify that the models were trained adequately, we confirmed the learning curve for each epoch, as shown in Figure 6. The sky blue line indicates the learning curve of the model for the Mannyeon underpass, the pink line indicates the learning curve of the model for the Oryang underpass, and the navy blue line indicates the learning curve of the model for the Daedong underpass. As shown in Figure 6, the loss value decreased as the number of epochs increased. During the training of the Mannyeon model, the loss value was rapidly increased in about 75 epoch, but it was restored again. All loss values were under 0.005 on 100 epochs, and this result demonstrated that all models were trained sufficiently without over-fitting.

We tested the qualified and quantitative performance of the models with all test datasets. We chose representative images to evaluate the qualified performance of the flooding region segmentation model. In the results, we represent the detected segmentation map as a blue-layer map.

Figure 7 shows the results of the flooding region segmentation model for the Mannyeon underpass. In the Mannyeon underpass, the road and sidewalk are separated by a fence. Since the sidewalk is higher than the road, the road is flooded first, and the sidewalk is flooded later. We show the representative test images, segmentation map, and ground truth to visualize the outstanding performances of the segmentation model (Figure 7). The segmentation map and ground truth are overlapped with the input test image. As shown in Figure 7, the segmentation results detect well compared to the ground truth.

Figure 8 shows the results of the flooding region segmentation model for the Oryang underpass. The Oryang underpass has a storm drain, which makes discharging the water smooth. Although a storm drain is present, flooding is unavoidable because of impediments such as garbage and leaves that cover the storm drain. The Oryang underpass does not have a sidewalk; it only has the road. Compared to other underpasses, the Oryang overpass has more cars passing by. As shown in Figure 8, the Oryang model only detected flooding areas without wrong segmentation, such as a car. The model detected that even small flooding regions were relatively similar to the ground truth.

Figure 9 shows the results of the flooding region segmentation model for the Daedong underpass. The Daedong underpass has a curb in the middle of the road, and several pillars are on the curb. An important factor for training the model with the Daedong underpass dataset is detecting the water on the opposite road between the pillars. Considering this point, we chose representative images, including various flooding regions in the Daedong underpass. As shown in Figure 9, the segmentation model detected the flooding region even on the opposite road. With increasing level of water, the segmentation map size of the opposite road was also increased.

Figure 10 shows each model’s average and standard deviation of the Dice coefficient. This graph shows the quantitative performance of the flooding region segmentation models. The bars denote the average of the Dice coefficient for each model. The points denote the standard deviation of the Dice coefficient for each model. The left vertical axis ranges the average of the Dice coefficient. The right vertical axis ranges the standard deviation of the Dice coefficient. As shown in Figure 10, the average Dice coefficient for each model is greater than 90%. The models achieved average Dice coefficients of 98.8%, 94.03%, and 93.85%, respectively. In particular, the Mannyeon model showed outstanding performance. The Daedong model showed the lowest Dice coefficient, although greater than 90%. The Daedong underpass has little light and is darker than the other underpasses. In addition, the surveillance camera was installed to take the scenes more widely than in the other underpasses. These points make it challenging for the model to detect the water area accurately. Nevertheless, the Daedong underpass model showed a 93.85% Dice coefficient, similar to the Oryang underpass model.

4.2. The Comparison Between Resultants of FCN and U-Net Models

To evaluate the relative accuracy of the U-Net-based model, we trained and tested other segmentation models based on FCNs. The backbone of FCN was VGG19 and pretrained weights were not used. We used the architectures of FCN-16s and FCN-8s. The hyper parameters were same as those used to train U-Net.

Figure 11 shows the results of the flooding region segmentation models based on FCNs and U-Net for all underpasses. Most models based on FCNs showed poor performance compared to U-Net-based models for all underpasses. These models detected only parts of the flooding regions. Especially for the Daedong underpass, which has a very complicated structure, the model based on FCN-8s was not trained properly and could not detect the flooding regions.

Table 4. The evaluation of the Dice coefficients (%) for the flooding region segmentation models.

Underpass	FCN-16s	FCN-8s	U-Net
Mannyeon	41.33%	38.64%	98.80%
Oryang	32.02%	31.64%	94.03%
Daedong	41.33%	Not trained	93.85%

shows the Dice coefficient for each model based on FCNs and U-Net. As shown in Table 4, the models based on U-Net showed the highest Dice coefficient. The models based on FCNs shows lower accuracy than the model based on U-Net. Also, same as Figure 11, the model based on FCN-8s was not trained properly for Daedong underpass.

5. Discussion

During the training process of the Mannyeon underpass model, the loss rate increased rapidly at 75 epochs and then converged again. The model weights are updated every epoch when the model is trained, so the loss rate usually decreases as the number of epochs increases. However, the loss rate may increase rapidly as the weights are updated during learning. For example, if the dataset is shuffled and randomly input during training, the loss rate of the model is different whenever data are input to the model. The loss rate may be the highest if the last input training dataset is not adapted to the weights. If this wrong updating of the weights continues, the loss rate will diverge, and the model will mis-learn. The updating of the weights depends on various components, such as the order of the input dataset, the costs in the cell of the model, and so on. The most important factor during training is whether the loss rate converges again and the weights are updated properly. In this study, an increased loss rate that was not very large increased rapidly to about 0.015, and it showed that the weights were updated correctly from the next epoch, so the training process had no issue.

This study used the vanilla U-Net structure and set the padding to 0, maintaining the size of the resulting map. U-Net is the most representative model among segmentation models and is usually used in various segmentation tasks. In this study, although the construction of each target region was different, we tried to check whether the flooding area detection performed well when the structure of the model was the same. As shown in the quantitative and qualitative accuracy, the quantitative accuracy (Dice coefficient) was more than 90% for all datasets, and the flooding regions were well-detected in most cases. However, when target regions are added in the future, and the same structure of the model is used, the accuracy may differ based on the target. Accordingly, it may require a change of model structure. Also, the available segmentation model, which shows higher accuracy, may differ for the target region.

For comparison, we trained models based on FCNs. The results of the models based on FCNs showed lower accuracy than the U-Net-based model. U-Net is a more advanced model than FCNs and all of the layers from the encoder and decoder have skip connections. Compared to FCNs, the predicted segmentation map is denser and more detailed, so the quantitative results (Dice coefficient) are reliable. For FCNs, FCN-8s usually shows most detailed and more accurate results among FCNs models. However, it was not trained properly for the Daedong underpass. As expected, when the feature maps from each layer were concatenated in the skip architecture, the coarse maps and dense maps were too unmatched, giving poor results and large errors. We figured that the structure of the Daedong underpass was too complex to use the model having only a few skip architectures. Likely U-Net, with an all skip connection architecture, is more proper for the flooding segmentation model.

Since the goal of this study was the segmentation task, a relatively complex labeling process was required. We labeled the flooding regions within all captured images using an image labeling tool. The label dataset had the form of COCO annotation. This process was time consuming and complex, but it was essential to improve the model’s accuracy.

Various hyperparameters were used during training to optimize the model. We trained the model by changing some parameters, such as the batch size, the optimizer, the learning rate, and so on. The maximum batch size was 16 due to the limited hardware environment. We tried lower values of 8 and 4, but the performance was not improved. For the optimizer, we used RMSprop (root mean square propagation) and ADAM, and ADAM showed better performance. The learning rate did not significantly affect the performance of the model. Therefore, the current hyperparameters were defined to train the model with the highest performance.

6. Conclusions

In this study, we constructed the dataset using surveillance camera videos of the Mannyeon, Oryang, and Daedong underpasses in Daejeon Metropolitan City. We proposed a flooding region segmentation model using a deep neural network. First, we obtained image data from video data to train and test the segmentation model. Considering the architecture of each underpass, images were captured to include various flooding situations. To train the model, the ground truth indicating the flood area was acquired by hand-crafted labeling. The flooding region segmentation model was implemented based on the architecture of the conventional U-Net model, which showed outstanding performance in semantic segmentation. The input and output sizes of the model were set to 256 × 256 × 3 and 256 × 256 × 1, respectively. Since the models were trained with each underpass dataset separately, 3 flooding region segmentation models were developed in the results. The parameters used to train the model were MSE loss function, ADAM optimizer with a learning rate of 1 × 10⁻⁴, and the models were trained on 100 epochs. The Dice coefficient was used for the evaluation metric. In the results, the models showed average Dice coefficients of 98.8%, 94.03%, and 93.85% for Mannyeon, Oryang, and Daedong, respectively. With these results, we made the following conclusions.

First, current flooding detection needs several elements, such as precipitation, road overland flow, water level sensors, and so on. However, this is difficult due to the absence of human resources, budget, and control systems. Like in our proposed study, if underpass surveillance camera images and deep learning technology are combined, a practical strategy can be suggested to automatically detect the status of flooding and flooding regions using only existing surveillance camera images.

Second, with surveillance camera images having a 10 min or 20 min range, the deep neural network can be trained to predict the flooding occurrence time in advance. The model will contribute to securing the golden time to respond to flooding situations.

Third, we used a video dataset only taken at a certain time in the underpasses. If datasets taken for a long time are acquired as training data, the amount of training data will increase, and the trained model will perform better.

In this study, an image dataset taken from only one site of the underpasses was used. However, if the model was trained on various surveillance camera datasets taken from the entrance, middle, and exit points of the underpass, the performance of the model would be improved. The result of this study, a flooding region segmentation model based on deep learning, can improve disaster response efficiency in areas at risk of flooding by integrating various technologies, such as AI-based real-time flooding monitoring, provision of spatial big data-based information, and e-SOP.