Hello, everybody. I'm Prativa Sarma, PhD candidate from University of Missouri, Kansas City, and net corresponding and presenting author for this paper on disaster aware global alerting platform for flood events. This paper is the application of the available models in

data and platform rather than waiting for alerting about the flood globally. The platform is to manage the alert is disaster aware. So disaster aware is the cloud-based software maintained by Pacific Disaster Center on Hawaii. It provides multi-hazard warning alerts and situational awareness information like risks and vulnerability,

hazard impact estimation, and critical needs that will aid towards decision making through mobile apps and web-based platform. So far, it has 7K users and 2 million apps have been downloaded globally. Severity of the alerts are categorized to four levels,

namely warning, watch, advisory, and information, and users can too be labeled to be alerted for. However, for the flood, automated global flood identification and alerting component is not

incorporated yet and is done manually based on social media and other flood products from different agencies, thus our paper aims to fill that gap. So towards that end, this work has four components. First is model of models. It integrates flood outputs from

two global flood forecast models, now global flood monitoring system abbreviated as GFMS and global flood awareness system abbreviated as GLOBAS in near real time to forecast flood severity at the global scale. Second component is obtaining inundation output using SAR imagery

at a granular level. This will be obtained for the high severity MAM output for the validation and calibration. Third component is the flood damage assessment in which the deep running approach will be used to get the building footprints within the high severity areas from

the MAM followed by validation by SAR output. Fourth is end-to-end pipeline that is to automate and integrate all the components together for automated alerting using disaster aware. I'll briefly talk about the components of model of models and methodology. This

flowchart shows the workflow of model of model to integrate two flood forecast models. The point value from output from GLOBAS, raster output from GFMS, and global watershed

boundaries developed by World Research Institute that is Polygon are integrated for the hazard identification with hazard score. This hazard score and watershed's riverine risk or coastal risk values will be combined to get the severity at to get the severity of

each watershed and based on this severity the flood alert is created which will be used by disaster aware. So again I'll briefly introduce the WRI watershed GFMS and GLOBAS. So World Research Institute developed the 16,385 static watershed boundaries

that are independent of political boundaries as shown in the figure. In addition to these static boundaries attributes such as riverine flood risk and coastal flood risk is utilized

in the model of model. These values will be updated annually based on inundation caused by river overflow exposure that is population in the flood zone and vulnerability and that is flooded population and existing level of flood protection. Towards the global flood simulation

model, global flood monitoring system GFMS is developed by NASA and University of Maryland using real-time precipitation information from global precipitation mission satellites and implements a hydrologic run-up and routing model for flood detection. It is

functional in quasi-global range at one by eighth degree or 12 kilometer spatula resolution and updated every three hours. Amongst many products of GFMS, flood depth above threshold in millimeter, a raster product as shown in this image is utilized in our project.

JONAT statistics is performed over the watershed using this raster and statistics such as area and percent area in watershed impacted by a flood mean and max flood depth above baseline and duration of flood is used as going factors towards the hazard score. So second flood

forecast model GLOFOS is developed jointly by the European commission and the European center for medium range weather forecast and is also independent of political boundaries. The GLOFOS output is the point data with information on the upstream river conditions.

From the GLOFOS product, the ensemble predictions of events with return period 2, 5 and 20 years alert level, medium, high or severe and peak forecast in days are utilized towards generating the hazard score for each watershed. So from the 10 products,

five from GLOFOS and five from GFMS, score is assigned based on weighing and score criteria as shown in this table. For example, for every flooded 1000 square kilometer flooded watershed area, one point is assigned, the maximum being 10 and so on for other products. So hence total

hazard score with maximum value 100 is thus assigned. The severity value is then the cumulative distribution function of hazard score. The CDF function is a long normal distribution function with scaled river in risk as mean and unit standard deviation. So flat alert for

affected watershed is then created based on the severity score. For example, if the severity is greater than 75%, warning alert is created for the given watershed. Similarly, if severity

is greater than 50% and less than 75%, then watch and if greater than 25% and less than 50%, then advisory and so on. So this map shows the M-M result for the African flood on June 6,

2020. Red, orange, yellow, green watersheds respectively have warning watch advisory and information flood alerts. This year, Central and East Africa, particularly the countries of Kenya, Somalia, South Sudan, South Sudan and the Democratic Republic of the Congo

experienced severe flooding this past spring as greater and more widespread than normal rainfall occurred during the long rain season. And the M-M output seems to be doing a good job identifying the floods. This image here is the centroid of the affected watershed area.

The second component is updating inundated areas from SAR images. This flow chart gives the methodology to get the inundated areas. The C-bands, the C-band SAR images from the Sentinel-1A B satellite is used for the project. First of all, the high severity watersheds

obtained from MOM is used as input subpiles to get the high resolution, ground range detected images from NASA's Alaska satellite facility and distributed active

center. This GRDH product is pre-processed using Sentinel-1 toolbox to the usable form. This pre-processed product is re-sampled to common grid and cropped so they all share the same pixels. This image is then splitted into numerous set of tiles such that each tile

could be analyzed separately. For each tile, maximum normalized between class variance method PCB variance method is applied to get the S-by-S box size which is the maximum number of binodal histograms. Now the automated thresholding algorithm is applied to each box to determine

the threshold for that box. The process is repeated for each box of each tile to take the average to threshold the tile and hence based on this threshold, binary products of the image is captured and finally just the inundated areas. This is the inundation output from the SAR

images targeting the Vadodara India as a case study for the date July 7 and July 20, 2020. The upper images are the pre-processed and re-sampled to common GRDH, common grid GRDH products

and these images below are the final binary products showing the water only. If closely seen, the change in the inundation can be noticed in the later date in the circled areas. So the third component of the project is using deep learning approach for the damage

assessment. Towards this component, the job that is accomplished so far is extraction of building footprint using deep learning approach, you know, using remote sensing images. In this

case study, we have used the Google Earth image. First of all, the image undergoes regional feature aggregation method that is region of interest align, ROI align method as proposed in mask or CNN. It uses bounding boxes to extract partial features within the bounding box. The extracted features will be used to classify if it is a building or background.

The first convolutional layer each ROI align passes is VGG16 network. Here in the images size is reduced to one by 32 of the original size. The second convolutional layer

draws a binary mask that can find the building accurately. The classified layer is the full connection network, FCM, that classify if it is a building or background. Lastly, the classification in the binary mask are combined to get the building footprint.

So this is the same image as shown before for the Johar Somalia. With the inundated extent and depth information from the SAR output, we aim to estimate the number of buildings affected using this building footprint and possibly the damage extent, you know,

towards the situational awareness. So this work is still ongoing and the future works for us are to calibrate the flood forecasted output and severity score from MOM using other global global products such as DFO products, flood observatory products and PDC manual hazards,

global flood detection system output, and some regional models such as rapid infrastructure flood tool developed by Pacific Northwest National Laboratory. These all will be utilized for finding the events missed by MOM and update the severity.

So towards the earth observation and damage assessment, flood depth data will be accessed in addition to the SAR flood extent and finally that depth and flood extent will be

utilized for the flood damage assessment. So finally, all the components will be integrated and automated and the whole process to generate alert based on severity score and generate the situational awareness information as much possible for disaster aware to indicate and emanate. So that was my last slide and I would like to thank all of you for your time

and patience and any questions.