A modern method for building damage evaluation using deep learning approach - Case study: Flash flooding in Derna, Libya

. Year after year, floods become more and more a frequent and destructive force of nature, causing significant infrastructure losses and loss of life. An accurate and rapid assessment is required to determine the degree of contamination. The present study proposes a modern method for building damage assessment using deep learning during the flash flood of Derna, Libya. For this reason, we first exploited SAR satellite data, captured before and after the flood, to accurately determine the flood extent. Next, the footprint of affected buildings within this extent was extracted using a deep learning approach (U-Net model) based on high-resolution satellite imagery (30 cm) from MAXAR. Finally, an additional analysis was carried out using VIIRS VNP46A2 data (500 m spatial resolution) to analyse the night light assessment. The results demonstrate the effectiveness of this method, given that 5877 buildings were submerged by water and 2002 buildings were totally or partially destroyed. Also taking into account the estimated night light, Derna's power supply was reduced by over 90% after the floods. The suggested approach is an effective tool for comprehending the global effects of floods and aiding in relief efforts.


Introduction
The most frequent and dangerous natural disasters is flash floods that cost billions of dollars in damage each year and claim thousands of lives [1] .For crisis response and recovery efforts to be successful, flood damage must be accurately and quickly identified.Traditional methods of damage assessment, however, can be imprecise and require a great deal of time and effort to achieve [2].
The importance of assessing the impact of flood events cannot be overstated.It not only aids in understanding the scale of devastation but is also crucial for supporting effective disaster relief efforts.
Recently, deep learning has become as a mighty tool that help in flood damage assessment.Deep learning models can be trained to extract meaningful features from satellite imagery and other data sources to identify and assess flood damage.
There has been a growing body of research on using deep learning for flood damage assessment in recent years.For example, a study by [3]proposed a deep learning-based flood risk assessment combined with multi-criteria decision analysis in Quang Nam province, Vietnam.Another study by [4] integrated deep learning and machine learning to automatically assess damage in extensive aerial image datasets, particularly targeting the identification of damaged critical infrastructure, which often poses challenges due to its distinct appearance compared to undamaged structures and the lack of training data for damaged infrastructure.
In addition, the use of high-resolution satellite imagery to map building footprints and assess damage has been the subject of previous research such as the study of [5] which has created a partially automated process to build databases of at-risk building elements by identifying building footprints via deep learning.This will help to support regional risk assessment in data-scarce regions.Also the study of [6] that introduced a rapid segmentation process for flooded buildings by integrating multiple sources of satellite imagery into a convolutional neural network.This approach aims to expedite the creation of flood maps from satellite imagery, which is crucial for early response efforts by local authorities.and flood responders during flood events.
In the meantime, using SAR satellite data to determine the extent of flooding has been a topic of interest in some earlier investigations.The effectiveness of SAR data in flood monitoring and damage assessment has been demonstrated by [7], providing a starting point for the use of this technology in the present study.
The method proposed in this article is derived from previous methods, but has a number of strengths.Firstly, it uses a combination of SAR satellite data, MAXAR high-resolution satellite imagery and VIIRS VNP46A2 data, providing more comprehensive information on the flood event.Secondly, it uses a deep learning model (U-Net) trained on basis of high-resolution satellite data.Thirdly, it has been evaluated using a dataset on the damage caused by the flash flood of Derna, Libya, which is a different region from those previously studied, while being considered the deadliest in Libyan In the following sections, we detail the methodology used and the results obtained.Our research not only quantifies the impact of flooding on buildings, but also suggests a wider application for understanding the overall effects of flooding and improving response efforts.

Study area
Derna, a coastal city in north-east Libya [8] , is situated in the Jabal al-Akhdar mountain range (Figure 1) .It is the Derna district's administrative center and home to about 125,000 people.Derna is a historic city that has been inhabited for over 2,000 years.It was once a major center of the Barbary States and played an important role in the American Revolution.Derna experiences hot, dry summers and mild, wet winters thanks to its Mediterranean climate with cool sea breezes moderating the summer heat.Derna experiences relatively mild winters in comparison to other regions of Libya, with average temperatures between 10°C and 18°C.Summers can be sweltering, with average highs of over 30°C.
It is one of the wettest regions in Libya with 265 millimeters (11 inches) of annual rainfall on average.

Method
The study was carried out in four stages, as shown in Figure 2, beginning with data gathering of GRD Sentinel-1 SAR image data of the area of interest before and after the flood, then applying coherent change detection (CCD) to identify the flood extent ,Next, train a deep learning model to extract the footprint of pre-and post-flood buildings within the flood extent and finally, use the trained model to assess the damage level of each building.

Data Used
The key data used in this study are Sentinel-1 SAR GRD, MAXAR high-resolution satellite imagery and VIIRS VNP46A2 data.

Sentinel-1 SAR GRD
The Sentinel-1 satellite constellation gathers data for a particular type of synthetic aperture radar (SAR) called Sentinel-1 SAR GRD (Ground Range Detected).
Radar pulses are transmitted, and the reflected signal is recorded, to gather SAR data.A radar image of the Earth's surface is then produced by processing the reflected signal.
Images from Sentinel-1 SAR GRD are publicly accessible via the Copernicus Open Access Hub.The ICCR'2 images are taken with a 10-meter resolution and revisit time is 6 days (equator), 12 days (poles).
SAR data is considered an excellent tool for flood mapping and damage assessment [9]- [11] , and it was used in this study for flood extent determination (figure 3) .The study's data came from Maxar Open Data, an initiative that makes high-resolution satellite imagery available to the public for free use in disaster relief and humanitarian efforts.

Data labeling
Labelling data entails appending labels to unprocessed data.Annotations known as labels give the data context and aid in the comprehension of the data by machine learning models.The building footprints in this study is labelled manually.

Data Preprocessing
The data set is divided into two distinct sets: training (70%) and validation (30%).Training data is provided as input to feed and train the deep learning model, while validation data for evaluation.ICCR'2

Model training
Based on literature, Many deep neural network architectures have been suggested to date as solutions to image segmentation challenges [12].On the basis of numerous studies [13]- [15], we have chosen U-Net as the main architecture for deep convolutional neural network (CNN) in our study .
The most popular deep learning model is the convolutional neural network.The visual cortex of animals served as the model's inspiration.
Convolutional neural network: CNNs operate by taking features out of pictures.Features are identifiable patterns in the image data that help identify the subject of the picture [16].
Convolution is a technique used by CNNs to extract features from images.A mathematical process called convolution is used to find patterns in pictures.In convolution, the input image is dragged over a filter, and the dot product between the filter and the image is computed.
U-Net: Convolutional neural networks (CNNs) with the U-Net architecture have been created for segmentation applications.Since it is a fully convolutional network, no fully connected layers are used.Its ability to handle images of any size makes it a good choice for image segmentation tasks.The encoderdecoder architecture forms the basis of U-Net.The output image is expanded by the decoder path, while the input image is contracted by the encoder path.The decoder path divides the image into segments based on the features that the encoder path has extracted from the input image.U-Net predicts the likelihood that each pixel will be assigned to a specific class, thereby performing per-pixel categorization [17].

Accuracy Assessment
The quality of model training was determining using the intersection over union (IoU) in relation to the intersection area and union area [18].
The degree of overlap of the predicted bounding box with the ground truth bounding box is measured by the IoU score.It is a number in the range of 0 and 1, where 0 denotes no overlap and 1 represents perfect overlap (equation 1) where A : labelled polygon B : predicted polygon

Flood extent
A before and after image of a flooded area in Derna, Libya, is shown in the figure 3. The before image depicts the area prior to flooding, while the after image depicts the area following flooding.By examining both images, coherent change detection was used to determine the extent of the inundation, as shown in red (figure 6).Accurate analysis of SAR satellite data is essential for a number of reasons.Firstly, it enables us to quickly and efficiently assess the extent of flood damage in affected areas.Such information can then be used to establish priorities for rescue operations and allocate resources to regions in greatest need.Secondly, the accuracy of SAR satellite data analysis can be used to validate other flood extent estimates derived from ground surveys or aerial photography.This contributes to the reliability of flood extent estimates, which can then be used for planning and decision-making.

Damage assessment
Figure 7 depicts the footprint of flood-damaged buildings in Derna, Libya, as extracted using a deep learning approach (U-Net model) based on Maxar's high-resolution (30 cm) satellite imagery.The IoU score indicated that the result was 0.6736 accurate.
A score of 0.6736 indicates that the model correctly identifies a significant portion of the affected buildings.However, it also suggests that some of the affected buildings escaped or were misclassified by our model.One possible explanation is that the satellite imagery used to train the model did not represent the full range of building types and damage patterns found in the flood zone.Furthermore, because the U-Net model is a general-purpose deep learning model, it may not have been specifically optimized for identifying flooddamaged buildings.The buildings affected are located around the river (figure 9), as the extent of flooding is concentrated along the river banks.This is because rivers are natural drainage channels, and are more likely to overflow their banks in the event of heavy rain or storm surges.Total buildings affected: 5877, of which 1999 were totally destroyed.
The result is a useful tool for understanding the city's flood risk.It can be used by city planners and emergency managers to develop flood risk mitigation strategies and protect the public from the effects of flooding.

The night light assessment
The figure 8 depicts the nighttime light assessment of the Derna flood in Libya, as captured by VIIRS VNP46A2 data (500 m spatial resolution).The image ICCR'2 depicts a significant decrease in light intensity in the flooded area.This is due to the fact that floods damage infrastructure such as power lines and substations, resulting in power outages.Floodwaters can also obscure street lamps and other light sources, reducing light intensity.Data from the VIIRS VNP46A2 can also be used to track the recovery process after a flood.Light intensity in the affected area increases as the waters recede and infrastructure is repaired.This occurrence can be used to track the recovery process and ensure that all areas recover at the same rate.After a flood, the decrease in light intensity can be used to assess the severity of the flood and identify the area most in need of assistance.Areas with the greatest decrease in light intensity, for example, are likely to be those that have been severely flooded and have the most damaged infrastructure.

Conclusion
Derna, Libya, was flooded severely, causing significant damage to infrastructure and buildings.To assess flood extent, damage, and night light intensity, SAR satellite data, a deep learning approach, and VIIRS VNP46A2 data were used.This provided valuable insights into the event.
Finally, we discover that the extent of flooding can be accurately mapped using SAR satellite data, and that a deep learning model is effective in identifying flooddamaged buildings, but that field validation is required for true accuracy.The evaluation of night lighting is an effective method for analysing light intensity in flooded areas.
This study is useful for understanding the risk of flooding in Derna and supporting disaster relief efforts.These analyses' findings can be used to aid disaster relief efforts, develop flood risk mitigation strategies, and track the recovery process.

Fig. 1 .
Fig. 1.Study area Climate change is having a major influence on Derna.The city is experiencing more extreme weather phenomena, including the most recent and deadly flood which occurred on the night of September 10-11, 2023 and is considered as the deadliest in Libyan history.It