Flood risk assessment via innovative geospatial data usage

Lidar and imagery data is being used to enhance urban resilience and emergency preparedness against flooding threats. This includes the integration of crowdsourced geospatial data to augment professionally collected sources.

Flood risk evaluation models require knowledge of the local conditions. Lidar and imagery data can be considered as complementary in providing topographic and radiometric measurements necessary for creating digital elevation models (DEMs) essential for identifying flood-prone areas and supporting hydraulic and hydrologic modelling of flood scenarios. Going beyond that classic use case, this article introduces several innovative uses of Lidar and imagery data to enhance the resilience and emergency preparedness of urban centres and their communities against flooding threats. Additionally, the article explores the integration of crowdsourced geospatial data, a rapidly growing resource that augments traditional geospatial sources collected by professionals.

Both Lidar and photogrammetry technologies can be used to produce digital terrain models (DTMs), the essential topographic input for floodplain delineation, overland flow simulation and prediction of flood extents, depths and velocities. Compared to imagery data, Lidar data is often superior in terms of spatial accuracy and level of detail. Detailed surface features such as street kerbs and drainage systems captured in a Lidar dataset can be useful for fine-scale urban flood modelling, since such features influence how water flows and accumulates. While not providing as much spatial accuracy and resolution as Lidar technologies do, photogrammetry technologies can capture more complete radiometric properties of the target environment. This is a significant advantage compared to Lidar, which is most often monochromatic (i.e. operating on a single wavelength). In the context of urban flood risk assessment, spectral information is valuable for identifying different types of land cover (e.g. impervious and permeable urban surfaces, urban vegetation and water bodies). Such information is useful in assessing surface runoff and infiltration rates. In the context of flood risk assessment, imagery data (e.g. aerial and street-level imagery) is also useful to examine the extent of flooding and resulting damages. Lidar and imagery data are complementary to each other in providing valuable information for flood risk evaluation.

Traditionally, Lidar and photogrammetry data acquisition is conducted systematically by trained personnel using survey-grade equipment to ensure data quality. Since professional surveys are expensive and time-consuming, survey-grade data may not always be available and can be outdated. Some challenges associated with limited survey-grade data can be overcome by exploiting alternative data sources, such as crowdsourced geographic data contributed voluntarily by non-professionals. While crowdsourced data is less reliable due to the limited control of data quality, it may be more widely available and less expensive than professionally acquired data. In many scenarios, crowdsourced data can supplement the limitations of authoritative data. 

Project UrbanARK

In a recent US-Ireland research and development project named UrbanARK, researchers from University College Dublin (Ireland), New York University (USA) and Queen’s University Belfast (Northern Ireland) developed innovative, computationally efficient, geospatial analysis methods for mining information about urban subsurface spaces from airborne Lidar data and crowdsourced street view imagery data. The obtained information is intended to enhance knowledge of flood risks. Information about subsurface spaces – such as building basements, underground car parks and pedestrian underpasses – is crucial for accurately modelling flood scenarios at a fine scale. In addition to affecting flood prediction results, subsurface spaces pose a higher risk during floods, making timely evacuation more challenging. Despite their importance, information about subsurface spaces is scarce and/or hard to access. The overall goal of the UrbanARK project is to enhance flood risk management for urban coastal communities using geospatial science. Most of the analysis methods developed in the project were designed to run on parallel computing infrastructure to ensure they are ready for real-world, large-scale applications.

Fusing Lidar and DTMs to detect open basements

Portions of Dublin, Belfast and Brooklyn (New York) form UrbanARK’s three study areas. Each location has a large number of terraced houses with open basement wells (Figure 1). These basements pose a higher flood risk to occupants and serve as unrecorded ‘sinks’ that may influence flooding but may not be captured in a standard flood model. However, the entryways to these basement wells are commonly larger than one square metre, making it possible for airborne Lidar signals to reach the bottom of the basement floor. At a sampling density of over 100 points/m² (the density level increasingly common in airborne Lidar mapping), each basement well may be captured by hundreds to thousands of sampling points. Those point samples returned from the basement wells can be reliably extracted from airborne Lidar point clouds by referencing the ground elevation present in existing DTMs available in most cities.

While the method is straightforward, the implementation of the method is non-trivial because of the high computational demand required. Comparing elevation values in a point cloud and their corresponding values in a terrain model is essentially a complex spatial join involving two potentially very large datasets. To address the computational load required by the analysis, UrbanARK researchers developed a powerful algorithm which partitions and decouples the input datasets so that different data portions can be analysed in parallel by different computing nodes (i.e. autonomous computers). The strategy effectively makes use of the aggregate computing power of multiple computers to speed up the analysis. The computing strategy is often known as ‘data parallelization’, which is widely used for big data analytics.

Using 16 computing nodes, each of which has 8 CPU cores, UrbanARK researchers completed the fusion of 1.4 billion Lidar data points and a 0.5m-resolution DTM covering an area of 2.5km² in Dublin city centre within five minutes. That speed and the option to connect additional computers to share the workload make the approach suitable for analysis at the city scale and beyond. Figure 2 shows the results obtained from the algorithm. In addition to open basement wells, the algorithm detects lowered backyards and outdoor staircases to underground floors, among other types of structures lying below the ground level. While airborne Lidar data was used in the study, the algorithm can take any kind of Lidar datasets as input, such as data from terrestrial Lidar or GeoSLAM scanners.

Extracting indicators from imagery using deep learning

Another opportunity to detect the presence of subsurface structures, as recognized by the UrbanARK team, is from photos captured at the street level, i.e. street view imagery (SVI). Due to the favourable range and viewpoint, SVI often contains features such as basement skylights, ventilation windows and basement wells, which are valuable indicators of building basements (see Figure 3). Fortunately, there are many large-scale SVI databases that offer extensive visual documentation of the built environment in many cities and towns worldwide. The prominent SVI data providers include Google Street View, Mapillary, Apple Look Around, and KartaView. Images in those databases often come with rich metadata that can be used to geolocate the images and features of interest in the images. SVI data collected systematically by trained personnel and specialized cameras, such as Google Street View imagery, have high quality but come at a cost and restricted terms of use. Crowdsourced data sources (e.g. Mapillary imagery), on the other hand, are more heterogeneous in terms of data quality but are also more accessible and may have fewer usage restrictions.

Using SVI data from the Mapillary crowdsourced database, UrbanARK researchers trained a deep learning model to automatically identify basement indicators from images and videos taken from the street level. The selected model was You Only Look Once (YOLO), a popular artificial neural network architecture capable of performing object detection and image segmentation. The model was originally trained using over 330,000 images in the Common Objects in Context (COCO) database and was fine-tuned on just 153 SVIs annotated by UrbanARK researchers. Through the original training, the model learned to recognize generic features from images such as corners, edges, textures and colour patterns. The fine-tuning process allowed the model to adapt to the specific task of detecting basement indicators. Figure 4 depicts the basement railings detected by the deep learning model trained by the UrbanARK team. While such a model is not completely accurate and is unlikely to capture every basement, the largely correct information it provides can be useful, particularly when combining the information with other datasets such as aerial images, Lidar data, ground penetrating radar data and building registries. Furthermore, SVI databases such as Mapillary are freely available, and valuable information can be mined from them to enhance knowledge of the urban environment.

Communicating flood risk information

In addition to using geospatial data for information extraction, the UrbanARK team employed Lidar and imagery data to develop a web-based tool for flood risk communication. By integrating multiple sources of Lidar and imagery data, they created a 3D textured point cloud representing the built environment for use as a basis of an immersive virtual reality (VR) environment. Terrestrial Lidar data, including GeoSLAM data, is preferred over airborne alternatives for this particular application where the data density is important to construct a photorealistic model. Imagery data can be sourced from devices such as DSLR cameras, mobile phone cameras or inexpensive cameras mounted on uncrewed aerial vehicles (UAVs or ‘drones’). The point clouds are transformed into an AR environment using open-source Potree and Three.js software libraries. Within this AR environment, the flood level, represented by a polygonal surface, can be adjusted to reflect different flood scenarios.

The UrbanARK tool involves the user wearing an inexpensive or easily constructed VR viewer compatible with a smartphone (see Figure 5). The smartphone renders the 3D environment, which is viewed through a set of lenses to create the 3D effect. To function, the smartphone must be equipped with standard sensors: an accelerometer, a gyroscope and a magnetometer. These sensors track the user’s location and head orientation, simulating head movements within the 3D virtual environment. This system allows participants to explore the environment remotely on a laptop or personal computer, feeding back information for the results. Such a VR tool can enhance flood risk awareness and increase the perception of vulnerability to flood risk among users.

Conclusion

While Lidar and imagery data have provided – and will continue to provide – flood prediction models with essential terrain information, their applications should not stop at the classic use case. As demonstrated in the UrbanARK project, both Lidar and imagery data can be analysed to extract information valuable to enhance flood prediction models, such as information about subsurface spaces in the urban environment. In addition, such data can be used to provide a means to effectively communicate flood risk information.

Acknowledgement

Funding for the UrbanARK project was provided by the National Science Foundation as part of the project ‘UrbanARK: Assessment, Risk Management, Knowledge for Coastal Flood Risk Management in Urban Areas’ NSF Award 1826134, jointly funded with Science Foundation Ireland (SFI - 17/US/3450) and a research grant (USI 137) from the Department for the Economy Northern Ireland under the US-Ireland R&D Partnership Programme. For the purpose of open access, the author has applied a Creative Commons (CC BY) public copyright licence to any author accepted manuscript version arising from this submission.

Further reading

www.urbanark-project.org

Open basement wells: https://artsandculture.google.com/story/RQUx8l4S5n_gKA