A deep learning technique-based automatic monitoring method for experimental urban road inundation

Reports indicate that high-cost, insecurity, and difficulty in complex environments hinder the traditional urban road inundation monitoring approach. This work proposed an automatic monitoring method for experimental urban road inundation based on the YOLOv2 deep learning framework. The proposed method is an affordable, secure, with high accuracy rates in urban road inundation evaluation. The automatic detection of experimental urban road inundation was carried out under both dry and wet conditions on roads in the study area with a scale of a few m. The validation average accuracy rate of the model was high with 90.1% inundation detection, while its training average accuracy rate was 96.1%. This indicated that the model has effective performance with high detection accuracy and recognition ability. Besides, the inundated water area of the experimental inundation region and the real road inundation region in the images was computed, showing that the relative errors of the measured area and the computed area were less than 20%. The results indicated that the proposed method can provide reliable inundation area evaluation. Therefore, our findings provide an effective guide in the management of urban floods and urban flood-warning, as well as systematic validation data for hydrologic and hydrodynamic models.

Meanwhile, the frequent urban flood inundation causes unavoidable disruption of transport and economic losses (Ruin et al. ; Lv et al. ). Nonetheless, monitoring the urban flood inundation can minimize the damages and losses (Versini ). Monitoring of the urban road inundation plays a key role in the application of urban flood inundation evaluation. Therefore, it is vital to monitor the urban road inundation to avert urban flood disaster.
Conventional urban flood inundation (e.g., road inundation) measurement methods (e.g., manual measurement, auxiliary mark method) have demonstrated several disadvantages under complicated climate and topographic surroundings, including insecurity, time-consuming, and high cost (Nair & Rao ; Zhang et al. ). In contrast with the traditional manual measurement mode, the sensors in modern measurement systems exhibit high precision.
Nevertheless, the sensors might be damaged and buried by frequent flood events (Lin et al. ). Also, the measurement readings could be affected by local electricity supply and Internet access (Amin ). Additionally, deep learning techniques have been effectively used in object recognition.  (Redmon & Farhadi ). These methods were well applied in image automatic recognition and classification problems (Van et al. ). Out of these, YOLO is a state-of-the-art framework for object detection and classification with a very deep layer and special residual net (Redmon et al. ; Zhang et al. ). The object detection can directly be evaluated by image pixels as a single regression problem using YOLO, the bounding boxes, and class probabilities. A single CNN in YOLO simultaneously predicts multiple bounding boxes and their class probabilities (Koirala et al. ). CNN harbors an intelligent learning mechanism, hence easier classification or prediction of objects, and learning the essential features of the object images from a small number of samples (Geng et al. ). The original YOLO network was referred to as YOLOv1. Furthermore, it was enhanced to YOLOv2 based on the YOLOv1 model. Therefore, YOLO v2 is also a state-of-the-art detection framework, with improved detection speed under stable accuracy. The great speed and accuracy of YOLOv2 was an improvement of YOLOv1, it uses a pass-through layer, higher resolution classifier, and anchor boxes (Redmon et al. ). Scholars confirmed that YOLOv2 has an advantage in image object detection (Redmon &  and edge detection techniques to forecast flood inundation, but they primarily focused on identifying the water surface depths of a small river rather than inundated road. So far, no study has systematically detected urban road inundation based on deep learning techniques. Thus, this study aims to provide a novel idea for experimental urban road inundation automatic monitoring using the YOLOv2 deep learning framework based on the collected images dataset. We believe that this could easily produce better performance on urban road inundation accurate detection considering the different scene images, including experimental rainwater collecting tanks inundation and urban road inundation with water.

METHODOLOGY
To automatically identify the urban road inundation, this work applied the YOLOv2-based Darknet-19 network to extract inundation areas. Moreover, the camera technology was applied to support the collection of images.  The learning rate parameter was set as 0.001 while the epochs for training the network were 50. Notably, all the images trained together were collectively called one epoch.
First, it was necessary to collect the experimental urban inundation images dataset before training the model on these images. Also, the object inundation region with water in the collected experimental urban inundation images was labeled before model training, and then these labeled images were used to train the model. Therefore, it was important to make the correct labels of the water object in the images, marking the location and labels for an object within the images, and reshaping the original images into 2D image format. The two consecutive layers of convolution and max-pooling had 3 × 3 convolutions

Inundation area computation approach
As mentioned before, the proposed novel idea was used to obtain information for the experimental urban road inundation. The inundation area was computed for the images captured by the cameras in the vertical orientation.
Additionally, if the images were collected from different angles, the object images were converted into an overlooked perspective as new test images using the inverse perspective transformation technique (Kim ), and then transformed images were used to estimate the inundation area. The inverse perspective transformation method transforms a two-dimensional image into a three-dimensional real-world space image, generating a bird-view (or top-view) image.
The spatial calibration was performed using this image processing method to correct the barrel distortion of images; it is because some images are taken by a surveillance camera which is equipped with a wide-angle lens. Given that a different inundation region size appeared in the images, the proposed different assessment methods were used to calculate the inundation area under two scenarios. Figure 3 shows the flowchart of the inundation area computation process. Moreover, the CNN algorithm was applied to accurately predict the number of more detection anchor boxes within images (Koirala et al. ). For the first scenario, if the inundation region with water was an irregular area, the inundation area was obtained from the accumulation area of smaller predicted detection anchor boxes that covered the entire water region in the images. The formulation of the inundation area is as follows: the inundation region with water was nearly covered by a predicted larger anchor box, the inundation area was evaluated by the area of this box coverage water region, reflecting the inundation region in the actual scene. Thus, the inundation area is also calculated using the formula: Moreover, to verify the performance of the method, the actual raining urban road inundation images were selected for the model test ( Figure 6). The following section describes the image preprocessing.

Image preprocessing
To enhance the performance of the model training and transformation, image rotation, and salt and pepper noise removal methods were applied to make an expansion processing on images numbers from 1,000 to 3,000. Among them, image rotation and salt and pepper noise removal methods were applied to make an expansion processing on images numbers from 700 to 2,000. The images

EXPERIMENTAL RESULTS AND DISCUSSION
The detection results of experimental urban road inundation based on YOLOv2 are described in the 'Experimental recognition accuracy evaluation' section; besides, we provided the recognition accuracy rate of experimental object detection evaluation. The new method for evaluating the experimental inundation area considering two scenarios is introduced in the 'Evaluation of the inundated area' section.

Experimental recognition accuracy evaluation
The accuracy rates in recognition of the model training and validation with an apparent change pattern are shown in Figure 8. As shown, the recognition accuracy rates increased at the beginning then reached a steady state for both the validation and train curves after 30 epochs. It was apparent that both model training and validation were higher than 0.9 and 0.8, respectively, and more than 10 for the model training epoch. As summarized in Table 1, the optimal training and    the lake surface water change in the global change detection, hence, confirming that the result of surface water detection effectively managed flood monitoring and warning. However, smaller-scale surface water was not considered.
Therefore, these findings guide the monitoring of urban road inundation based on our automatic detection analysis.

Scenario 1
A reliable result of the high performance of the established model for the detection of the inundated region is provided in the previous section. In this subsection, the novel proposed method was used for computing inundation areas covered by water via adjusting the size of predicted anchor boxes when the inundation region with water had an irregular area scenario. For example, Figure 13 shows the measured water area and recognition result with 10 anchor boxes. Besides, a green mark above these boxes shows the confidence score and classification information for model output detection result. As shown in Figure 13

Scenario 2
For the second scenario, the inundation area was obtained to compute the predicted box region area covered by water when the inundation region was nearly covered by a predicted larger anchor box. Fifty cases were tested in this section, and the experimental 20 images with the larger inundated region area were collected using a surveillance camera with high resolution. For instance, the example of four typical cases is shown in Figure 14, there was a better accuracy rate in the inundation automatic detection. The images were captured under a similar experiment site and taken in the vertical orientation (Figure 14(a) and 14(b)).
Also, after the object images were converted into an overlooked perspective, the larger inundated region area was accurately computed as shown in Equation (2) ( Figure 14(c) and 14(d)). Therefore, the total measured area in the real scene for these images was directly measured with the image input pixel of 416 × 416. Moreover, the predicted values of (x min , y min ) and (x max , y max ) coordinate of the anchor box for two images are listed in Table 2, which calculated the area proportion of the detection box coverage area to the total measured area, with the origin coordinates (0, 0) being at the upper-left corner of the image region.
Thereafter, the inundation area in the images is computed by Equation (2). Unlike the measured and computed areas for the inundation region with water, the relative • Moreover, the results of inundation recognition accuracy rates showed that the model training and validation accuracy rates were high with 96.1 and 90.1%, respectively.
Moreover, further validation confirmed that it harbors a higher accuracy rate for the actual road inundation detection. Therefore, the results indicated the impact of high accuracy and reliability for road inundation automatic detection.
• Furthermore, by comparing the measured inundation area and computation inundation area for two scenarios, the area relative errors of the test cases were less than 20%, with the average relative error percentage of 7.7%.
The findings indicated an effective performance in the assessment of the inundation area.