Identification,Monitoring And Loss Prediction Of Flood Event Using Multi-source Observation Data

Both the frequency and intensity of global flood hazard events are increasing under the influence of climate change,posing a serious threat to socio-economic development.With the development of observation technology,big earth observation data have made significant progress in terms of volume,quality and sharing.However,challenges still remain in flood event identification,monitoring and loss prediction related studies.Firstly,existing flood event database are missing the information about the dynamics of the flood impact extent.Second,the flood inundation extent mapping based on synthetic aperture radar(SAR)imagery lacks large-scale comparative validation,and the corresponding flood inundation datasets are rarely published.Finally,the study of flood loss prediction by fusing multi-source big data at the event scale has seldom be reported.To fill the vacancy,this paper using the continental United States(CONUS)as the study area,focus on the national-scale flood event identification,inundation extent monitoring and loss prediction based on big earth observation data using data analysis,fusion,image processing and machine learning methods.The main contents and conclusions of this study are detailed as follows.(1)This study proposes the methods to classify and identify the potential impact range of flood events based on multi-source observation data and form a flood event extent database(FEED).Firstly,the author classifies the flood events into fluvial flood,pluvial flood and coastal flood events,based on the river stage gages,remotely sensed precipitation and coastal water level data.Secondly,the author defines the potential flood zone(PFZ)of each type of flood event using thresholding,watershed segmentation and reverse convergence methods.Finally,the author proposes a spatiotemporal clustering method to identify the flood events by considering the spatial proximity and temporal continuity of the PFZs.Based on this,the author identifies 21,589 flood events over the CONUS area from January 2016 to June 2019 and formed a flood event extent database(FEED).The result indicates that the PFZs of all flood events is mainly distributed in the Mississippi River basin and coastal states,with fluvial flood is the main flood type.The flood events in the western coastal,southwestern,midwestern and eastern coastal regions are most likely happened in the winter,summer,spring and fall/winter seasons,respectively,the pattern of which is generally consistent with the findings of the existing studies.Comparing with the existing databases,FEED,on the one hand,considers more comprehensive drives of flood events and thus more reasonable in terms of PFZs,and on the other hand,can reflect the spatiotemporal dynamics of flood events.(2)This study develops an automated flood inundation extent monitoring system based on Senitnel-1 dual-polarization SAR data and validates the efficiency of the system at a large spatial-temporal domains.Firstly,the author proposes a TD method that can enhance the efficiency of image utilization to improve the RAPID algorithm and form the TD-RAPID flood inundation extent extraction algorithm using SAR data.Secondly,based on TD-RAPID,the author also develops an automated mapping system from the SAR image querying to flood inundation delineation,using FEED as the reference of the spatial and temporal extent of flood events.Finally,the system queries all Sentinel-1 dual-polarization SAR databases covering the CONUS area from January 2016 to June 2019 and acquires a total of 2,054 and 9,161 flooding and non-flooding images,based on which conducts the TD-RAPID algorithm.The result shows that the TD method provides an effective measure to cope with the noise of Sentinel-1 dual-polarized SAR images and improves the overall success rate of the RAPID algorithm for inundation monitoring.The visual comparison of several cases indicate that the inundation map is in good agreement with other similar data.The quantitative analysis of the 559 inundation mapping results with DSWE reveal that they are in high agreement,and the areas with deviations are concentrated in the areas where vegetation and water bodies intersect,such as forest,grassland and cropland.By leveraging the advantage features of SAR sensors,the proposed system achieves quasi-real-time,unsupervised and automated flood event inundation monitoring over CONUS area,and the derived flood maps have successfully provided supports for several departments and agencies,including the USGS Disaster Emergency Management Team,the DFO Flood Observation Center and the United Nations International Chater of Space&Major Disasters.(3)This study proposes a machine learning model-ICLAIM,which incorporates multiple sources of observation data,to achieve the prediction of properties insurance claims induced by flood events.Firstly,by overlaying the FEED with flood properties insurance claim records from National Flood Insurance Program(NFIP),the author defines a grid-event sample format and develop a flood property Insurance CLAIMs prediction model(ICLAIM)by fusing NFIP records,building locations,topography,basin morphometry,land cover,and multiple sources of hydrometeorological variables,including flood extent,precipitation,and operational river stage and oceanic water level measurements.Secondly,the model utilizes two steps,claim level classification and claim number regression,and accordingly designed subsampling strategies to reduce overfitting and underfitting caused by the flood claim samples,which are unevenly distributed and widely ranged.Finally,the author evaluates the model using 446,446 grid samples identified from 589 flood events occurring from 2016 to 2019 over the CONUS,overlapping 258,159 claims out of a total of 287,439 NFIP records of the same period.The rigorous validation yields acceptable and satisfactory performance at the grid/event level and county/event level,with R~2 being over 0.5 and 0.9,respectively.The prediction accuracy of ICLAIM is better than that of the traditional RF model,reflecting the advantages of the joint use of classification and regression model structure,and the applicability of the balanced resampling strategy to the unbalanced data prediction.The predictor importance results confirm the validity of the feature extraction in this study,which can provide a reference for feature selection in other related studies.The author concludes that the ICLAIM model can facilitate various applications,including assessing flood impact and improving flood resilience.