System and method for estimating floods

Abstract

Disclosed are methods, systems, and techniques for flood estimation, the flood estimation method including receiving input data associated with a flood event for running a hydrological model, receiving external flood data for the flood event, validating the input data over one or more iterations based on the external flood data, and running the hydrological model with the validated input data based on a candidate input data set that provides a candidate model output having a satisfactory goodness of fit to the external flood data to estimate a water level of the flood event.

Description

[Technical Field]

[0001] The present disclosure relates to estimating floods and estimating water levels of flood events using data from a variety of sources. [Background technology]

[0002] Hydrological models can be used to simulate floods. Existing hydrological models receive input data, such as rainfall data, topographical data, and land use / land cover, and then model the input data to simulate floods. However, the output of hydrological models often contains significant inaccuracies due to errors in the input data, errors calculated from the model itself, and / or sensitivity to the input data or other parameters of the model. Thus, model outputs for a particular time or event may be unreliable for guiding decision-making. For example, governments and / or insurance companies may be interested in peak flood water levels during a flood event, but the peak flood water levels predicted by models are often inaccurate for the reasons mentioned above.

[0003] To improve the accuracy of output, complex hydrological models are used, requiring computational time of hours or even days, resulting in delays in output as well as significant computational and financial burdens. Furthermore, even these more complex hydrological models can be inaccurate due to difficulties in modeling surface water flow and the sensitivity of floods to variables such as localized rainfall and topographic changes. Complex hydrological models also cannot accurately estimate temporal changes in input data, resulting in inaccuracies in pre-calculated models due to an inability to predict and correct for changes in input data as flood events progress. Given the complexity of these hydrological models, it is impractical to continuously run them to account for updated input data. Summary of the Invention

[0004] A flood estimation method according to a first aspect includes receiving input data associated with a flood event for running a hydrological model; receiving external flood data for the flood event; generating, over one or more iterations based on the external flood data, a plurality of candidate input datasets for a current iteration, a plurality of candidate input datasets for a first iteration generated from the input data, and a plurality of candidate input datasets for a subsequent iteration generated based on one or more satisfactory candidate input datasets from a previous iteration; validating the input data by running the hydrological model using the plurality of candidate input datasets to generate a plurality of candidate model outputs; and comparing the plurality of candidate model outputs to the external flood data to identify one or more satisfactory candidate input datasets for the current iteration; and running the hydrological model using the validated input data based on the candidate input datasets that provide candidate model outputs having a satisfactory fit to the external flood data to estimate a water level of the flood event.

[0005] In some aspects, the input data includes measured data, predicted data, or a combination thereof.

[0006] In some embodiments, the multiple candidate input datasets in the current iteration are generated for the first iteration from a stochastic sampling of offsets of the input data, and for subsequent iterations from a stochastic sampling of offsets of one or more satisfactory candidate input datasets in a previous iteration.

[0007] In some embodiments, the offsets are stochastically sampled from a distribution of offsets in the input data, hi some embodiments, the distribution is Gaussian.

[0008] In some embodiments, the flood estimation method further includes, at each iteration, determining a measure of how well each of the plurality of candidate model outputs matches the external flood data, and validating the input data over one or more iterations is performed until (a) each of the plurality of candidate model outputs in the current iteration is within a predetermined threshold of the external flood data, (b) each of the plurality of candidate model outputs in the current iteration is unchanged from the previous iteration, or (c) a predetermined number of iterations have occurred without satisfying (a) or (b).

[0009] In some embodiments, running the hydrological model with a plurality of candidate input datasets includes running the hydrological model in parallel for each of the plurality of candidate input datasets in the current iteration.

[0010] In some embodiments, comparing each of the plurality of candidate model outputs to the external flood data includes scoring each of the plurality of candidate model outputs according to how well they match the external flood data.

[0011] In some embodiments, the external flood data includes a plurality of external datasets, and comparing each of the plurality of candidate model outputs to the external flood data includes comparing each of the plurality of candidate model outputs to each of the plurality of external datasets.

[0012] In some embodiments, comparing the plurality of candidate model outputs to the external flood data includes one or more of determining an average error between the candidate model outputs and the external flood data, determining whether the candidate model outputs meet the flood criteria set by the external flood data, and, if the candidate model outputs do not meet the flood criteria, determining a distance to the flood criteria.

[0013] In some embodiments, the external flood data includes one or more of earth observation data, gauge data, and social networking service data.

[0014] In some embodiments, the external flood data includes gauge data, and comparing the plurality of candidate model outputs to the external flood data includes evaluating a mean absolute error between the gauge data in the external flood data and sampling points from each of the plurality of candidate model outputs set at locations of the gauge data; and / or the external flood data includes social networking service data, and comparing the plurality of candidate model outputs to the external flood data includes setting a water depth criterion at the geographic location based on the SNS data, comparing a simulated water depth at the geographic location for each of the plurality of candidate model outputs to the criterion, and evaluating whether the simulated water depth at the geographic location meets the criterion, and if not, comparing the simulated water depth at the geographic location with the water depth from the SNS data. and / or the external flood data includes Earth observation data, and comparing the plurality of candidate model outputs with the external flood data includes predicting the probability of flooded areas within the domain of the Earth observation data, calculating an Earth observation data output vector using areas associated with areas with a high probability of being flooded and areas with a low probability of being flooded, generating a candidate model output vector for each of the plurality of candidate model outputs based on areas simulated to be flooded and areas simulated not to be flooded according to a predetermined simulated water depth, and comparing the Earth observation data output vector with the candidate model output vector for each of the plurality of candidate model outputs by calculating a cross-entropy between the Earth observation data output vector and the candidate model output vector.

[0015] In some aspects, the flood estimation method further includes generating boundary conditions from one or both of the input data and the external data to be used as boundaries constraining the plurality of candidate input data sets.

[0016] In some aspects, the flood estimation method further includes selecting a hydrological model from among a plurality of hydrological models, where selecting the hydrological model includes running each of the plurality of hydrological models with an identical input dataset to generate a respective model output for each of the plurality of hydrological models, comparing each model output to external flood data, and selecting the hydrological model that generated a satisfactory model output for the external flood data.

[0017] In some aspects, the flood estimation method further includes generating and displaying a peak flood map of the flood event from the estimate of water level.

[0018] The system of the second aspect comprises a database storing input data and external flood data associated with a flood event, a processor, and a non-transitory computer-readable medium storing computer program code executable by the processor, which, when executed by the processor, causes the flood estimation method described in any of the above aspects to be performed.

[0019] In some aspects, the system further comprises a plurality of processors configured to run in parallel, the plurality of processors configured to run the hydrological model in parallel using each of the plurality of probabilistic input datasets to generate a plurality of candidate outputs.

[0020] A non-transitory computer-readable medium according to a third aspect stores computer program code executable by a processor, which, when executed by the processor, causes the processor to perform the method according to any one of the above aspects.

[0021] This summary does not necessarily describe the scope of all aspects, and other aspects, features, and advantages will become apparent to those skilled in the art upon review of the following description of specific embodiments. [Brief explanation of the drawings]

[0022] In order that the present disclosure may be more readily understood, preferred embodiments of the invention will now be described, by way of example only, with reference to the accompanying drawings, in which:

[0023] [Figure 1] 1 shows a system for implementing a flood estimation method. [Figure 2] 1 illustrates the general flow followed by a system for implementing a flood estimation method. [Figure 3] It also shows external flood data that is used to iteratively update the input data used in the hydrological model. [Figure 4] 1 shows a representation of water gauge data points in an example flood zone. [Figure 5] 1 shows a representation of social networking service (SNS) data points in an example flood area. [Figure 6] The flood estimation method is shown below. [Figure 7] Indicates how input data is validated. [Figures 8A-8C] Following the flood estimation method, we show the iterative refinement of candidate outputs towards external gauge data. [Figures 9A-9D] 10 shows examples of different peak water depth outputs simulated using different candidate input datasets. [Figure 10] An example of a peak flood map is shown below. [Figure 11] Figure 1 shows the representation of the uncertainty area in simulated floods. DETAILED DESCRIPTION OF THE INVENTION

[0024] In accordance with the present disclosure, a flood estimation method is described that can validate input data used to run a hydrological model. The flood estimation method generates a plurality of candidate input datasets and generates a plurality of candidate model outputs by running a hydrological model against the plurality of candidate input datasets. The plurality of candidate model outputs are compared with external flood data to identify one or more satisfactory candidate input datasets that produced satisfactory candidate model outputs based on the external flood data. The satisfactory candidate input datasets can be used to generate subsequent plurality of candidate input datasets, and in an iterative approach, the candidate input datasets produce better candidate model outputs that are validated by the external flood data. In this manner, a hydrological model can be run using validated input data that provide candidate model outputs that have a satisfactory fit to the external flood data.

[0025] Because external flood data is used to validate and update the quality of the input data and corresponding model outputs, the flood estimation methods described herein can overcome inaccuracies in the input data and generate more accurate model outputs representative of flood events. Also, because accurate model outputs can be generated by validating candidate model outputs against external flood data, simpler hydrological models can be used, which can generate model outputs much faster and at lower cost than complex hydrological models. A system for implementing the flood estimation methods may include multiple processors configured to operate in parallel, each generating a different candidate model output using a different candidate input dataset, to further reduce the computational time associated with running a hydrological model multiple times to generate multiple candidate model outputs.

[0026] Therefore, the flood estimation methods described herein may be performed more frequently than with complex hydrological models. Because the flood estimation methods can be performed more frequently than conventional techniques, they can be performed in response to dynamically changing input data (and / or external flood data), providing better and more up-to-date model outputs for dynamically changing flood events.

[0027] Embodiments will now be described, by way of example only, with reference to Figures 1 to 11.

[0028] 1 illustrates a system for implementing a flood estimation method. The system includes one or more computing devices 102 (e.g., servers) configured to implement the flood estimation method. The computing devices 102 may be distributed (cloud-based). The computing devices 102 may include multiple computing devices configured to operate in parallel to implement the flood estimation method, as described in more detail herein.

[0029] Each computing device 102 is configured to run a hydrological model, which may be stored locally on the computing device 102 or in a database, such as the model database 120 shown in FIG. 1 . The computing device 102 may retrieve the hydrological model from the model database 120 over the network 130. The hydrological model is configured to simulate a flood event and provide a model output of the flood event, such as the flood extent and water depth over time associated with the flood event. The hydrological model simulates the flood event using input data that includes both flood-specific data and model control parameters. The control parameters may be stored locally on the computing device 102 in association with one or more hydrological models, such as the model database 120, or may be stored in a separate database, such as the control parameter database 122 shown in FIG. 1 , and accessible over the network 130. The control parameters define spatial parameters associated with various regions on the Earth and may include digital terrain or elevation models, land use and land cover data, friction parameters, permeability parameters, and other information. The flood-specific data regarding a flood event may include time series data associated with the flood event, such as rainfall data, water gauge data, etc., and may be received over the network 130 from one or more external sources, such as a weather server 140, a gauge server 150 communicatively coupled to receive water gauge data from a water gauge 152, etc. The flood-specific data is preferably received at a real-time or near real-time rate (i.e., not subject to communication delays) so that the input data is as current as possible.However, as will become apparent from the present disclosure, the flood estimation method can also overcome inaccuracies in the input data and can be performed using, for example, input data provided by running a simulation in another domain (e.g., model output of a hydrological model run in an upstream domain can be used as input data for a downstream domain), as well as flood-specific data (e.g., a flood event is ongoing and the input data includes current flood-specific data and estimated / predicted flood-specific data based on the current flood-specific data).

[0030] Each computing device 102 is configured to receive input data associated with a flood event (i.e., flood-specific data and model control parameters) and execute a hydrological model. As shown in FIG. 1 , the computing device 102 includes a processing unit (e.g., a CPU 104), an input / output (I / O) interface 106, non-volatile storage 108, and non-transitory computer-readable memory 110. The computing device 102 is configured to receive input data via the I / O interface 106. The non-transitory computer-readable memory 110 includes computer-executable instructions that, when executed by the CPU 104, configure the computing device 102 to execute the hydrological model. As mentioned above, the computing device 102 may include multiple computing devices configured to execute the hydrological model in parallel, each of which may execute the hydrological model using different input data, as described further herein. The hydrological model itself may be executed using parallel processing, for example, using a cellular automaton model to represent the hydrological model in discretized cellular states, where each cell directly interacts with neighboring cells and is iteratively updated. It is also possible for different computing devices to perform different hydrological analyses using the same set of input data, which may be useful for selecting a hydrological model that provides satisfactory model output, as described in more detail herein. Parallel processing (e.g., a flood simulator) may include processing using graphics processors that can perform these operations efficiently.

[0031] At least one of the computing devices 102 is configured to execute an input data validation / update algorithm 111 stored as computer-executable instructions in the computing device's non-transitory computer-readable memory 110 and configured to be executed by the CPU 104. As described in more detail below, the input data validation / update algorithm 111 is configured to validate the input data by comparing candidate model outputs generated from the input data with external flood data. Based on this comparison, the input data validation / update algorithm 111 can identify one or more candidate input data sets that produce satisfactory candidate model outputs for the external data.

[0032] In one example, the computing devices 102 may be arranged in a master-slave configuration, where a master computing device generates and provides a set of input data to each slave computing device, which then runs the hydrological model using its respective set of input data and provides its respective output to the master computing device, which executes the input data validation / update algorithm 111. Alternatively, a single computing device may run the hydrological models sequentially using different sets of input data. However, using parallel processing, including multiple computing devices and / or multiple processors and / or graphics processors, can reduce computation time compared to running the models sequentially.

[0033] To validate the candidate model outputs, external flood data is received via the network 130 from one or more data sources. Various types of external flood data may be used to validate the candidate outputs. In FIG. 1 , the external flood data may include water level data received from a water level server 150 communicatively coupled to receive water level data from a water level gauge 152; social networking service (SNS) data, which may include images, videos, and other data related to flood events published by one or more user devices 162, 164, received from an application-accessible social networking service (SNS) server 160; and earth observation data, which may include, for example, images acquired from one or more radar systems 172 (e.g., using synthetic aperture radar) and / or one or more satellites 174, received from an image server 170. Further details regarding the use of external data in the model validation / update algorithm are provided below.

[0034] FIG. 2 shows the general flow followed by the system to perform the flood estimation method.

[0035] The system receives input data 202 associated with a flood, including flood-specific data and model control parameters. Figure 2 shows that the input data 202 includes flow or water level time series 206 provided by a water level service 204, rainfall time series 210 provided by a meteorological service 208, a digital terrain model (DTM) 214 provided by a digital elevation model (DEM) service 212, and global infiltration and friction parameters 216. It should be understood that the types of input data 202 shown in Figure 2 are not limited to these and that additional types of input data are possible. It should also be understood that not all of the input data 202 may be available for a given flood, such as flow or water level time series 206 obtained from one or more water gauges, and may not be available for a certain flood location.

[0036] The input data 202 is used by a hydrological model 220, which, when executed, simulates a flood event by modeling various hydrological processes using the input data. The hydrological model 220 generates model outputs 222 that may provide, for example, estimates of water depth and / or water level over time, and / or a flood peak map showing peak water depths at different locations for a flood event. The model outputs 222 are passed to an input data validation / update algorithm that performs data optimization 250. Furthermore, as described with reference to FIG. 1, the hydrological model 220 may be run multiple times in parallel, and therefore multiple model outputs 222 may be passed to data optimization 250.

[0037] As previously mentioned, outputs from hydrological models often contain inaccuracies due to errors in the input data, errors calculated from the model itself, and / or sensitivity to the input data or other model parameters. For example, flood-specific data may have measurement inaccuracies due to geolocation errors, missing values, missing natural variability, etc. Furthermore, inaccuracies may arise due to the complexity of hydrological modeling in general, and hydrological model outputs may not be perfect and may differ significantly from reality. In accordance with the present disclosure, data optimization 250 is performed to compare one or more model outputs 222 with external flood data 240 to verify whether the input data used to run the hydrological model 220 produces satisfactory model outputs. Based on the comparison, the input data can be updated to improve the accuracy of the model output relative to the external data 240. Thus, an iterative process of validating and updating the input data can be applied until the model output 222 has a satisfactory fit to the external flood data 240. The external flood data may comprise various data such as social networking service (SNS) points 242, water level graphs from water gauges 244, and Earth Observation (EO) data 246, for example, as described with reference to FIG.

[0038] Flood-specific input data and / or model control parameters may be updated. It should also be understood that only a portion of the input data may be updated during the data optimization process. Model input data for update 230 is passed to data optimization 250. Model input data for update 230 may include, for example, permeability and / or friction parameters 232, input gauge data and optional time offset 234, and / or input rainfall data and optional time offset 236. As described in more detail herein, because some input data is received as a time series (e.g., gauge data and rainfall data), not only can the values ​​of each input data be updated, but also the time of day in the time series (i.e., applying a time offset). For example, shifting rainfall input data forward or backward by one hour can create an output dataset from the hydrological model that better matches external flood data. This may be due to differences between the actual water velocity across the land surface and the velocity within the model. Meanwhile, other input data parameters, such as the DTM and spatial inputs like land use and land cover (LULC), may be held constant. However, the illustration of the model input data for updating 230 in FIG. 2 is by way of example only and is not intended to be limiting.

[0039] Data optimization 250 can update update model input data 230 based on a comparison of model output 222 with external flood data 240 to provide updated model input data 252. If satisfactory, the updated model input data 252 can be used as input to the hydrological model 220 to simulate floods. Alternatively, if a more accurate prediction is desired, the updated model input data 252 can be used in a subsequent hydrological model 220 to generate another model output 222, and the process can be repeated.

[0040] Data Optimizer 250 may generate updated model input data by updating the model input data using an optimization routine. Various optimization routines, such as the Trust Region Dogleg or Levenberg-Marquardt algorithms, may be used to generate the updated model input data. However, these classical optimization methods are serial in nature and may require hundreds of hours of computation time depending on the number of iterations performed and the desired accuracy of the model output relative to the external data.

[0041] To facilitate parallel hydrological model execution, multiple candidate input datasets containing different variations of the input data are generated. In a preferred embodiment, this involves generating candidate input datasets containing stochastic sampling of offsets to be applied to the input data, running the hydrological model using these to generate candidate model outputs (i.e., implementing a Monte Carlo method). Furthermore, a preferred optimization routine may use a covariance matrix adaptive evolution strategy (CMA-ES) algorithm. As an example implementation, data optimization may function by fitting a distribution (e.g., a Gaussian distribution) to define possible offsets (e.g., rainfall offset, water level offset, time offset, etc.) for the model input data to be updated, and then sampling from the distribution to determine the offsets to be applied to the input data. Sampling from the distribution a predetermined number of times (e.g., 50 times) results in different offset values ​​(e.g., 50 offsets relative to the input data values), and applying these sampled offsets to the input data generates candidate input datasets. Thus, by sampling the offsets to be applied to the input data (i.e., offsets relative to the model input data to be updated), multiple candidate input datasets (e.g., 50) with unique combinations of input data parameters are generated. The goodness of fit of each of these candidate input datasets is evaluated by running the hydrological model with the candidate input dataset to generate a candidate output model for comparison with observed external data, e.g., as further described with reference to Figure 3. One or more satisfactory candidate input datasets that produce satisfactory candidate model outputs for the external data can be used to generate a further set of subsequent candidate input datasets (e.g., generated by sampling the distribution to determine an offset to apply to one or more satisfactory candidate input datasets). This iterative approach allows candidate input datasets to provide better candidate model outputs for the external data.

[0042] The generation of candidate input datasets is constrained by boundary conditions that provide physical constraints based on the received input data and / or external flood data. In the example CMA-ES algorithm, optimization is performed in "genotypic" space, while simulation is performed in "phenotypic" space. Thus, optimization operates in a space constrained between [-1, 1] in all dimensions, and the results are then scaled to real-valued space (i.e., an x3 value of 0.43 means that the forest area has an infiltration rate of 22 mm / hr). The boundary conditions are different for all parameters of the phenotype but identical for the genotype, i.e., the optimization space is symmetric but the real values ​​are not.

[0043] The goal of the optimization routine is to update the input data so that the candidate model outputs show a satisfactory match to the external flood data. Data optimization 250 may continue until a predetermined number of iterations have been performed, until each of the candidate model outputs in the current iteration is within a predetermined threshold of the external flood data, or until each of the candidate model outputs in the current iteration is unchanged from the previous iteration.

[0044] Figure 3 further illustrates the external flood data used to iteratively update the input data used in the hydrological model. The schematic diagram in Figure 3 shows only a subset of the flows as shown in Figure 2, the external flood data used to evaluate candidate model outputs.

[0045] As shown in Figure 3, the hydrological simulation uses multiple candidate input datasets to generate multiple candidate model outputs (302). The candidate model outputs are compared to external flood data, and one or more satisfactory candidate input datasets are selected (304). The external flood data, as shown in Figure 3, may include gauge curves 310, social networking service (SNS) points 312, and Earth Observation (EO) data 314. It should also be understood that the present disclosure is not limited to these examples of external flood data, as other types of external data may also be used.

[0046] The gauge curve 310 is obtained from one or more water gauges, if any, located in or upstream of the flooded area. Advantageously, the gauge curve provides time-stamped data spanning the entire flood event. However, water gauges are typically located only within waterways and are generally limited in number, if at all. Furthermore, water gauges may also experience geolocation errors or outliers. Figure 4 shows a representation of water gauge data points 402 in an example flood area 400.

[0047] The comparison of the candidate model output with the water level curve may be performed by evaluating the mean absolute error between the gauge data of the actual gauge in the external flood data and the sampling points set at the gauge locations in the simulated flood. For example, simulated water depths at the gauge locations in the simulated flood can be saved at predetermined time intervals starting from the same time as the gauge measurement times in the gauge curve. The saved simulated water depth over time provides a virtual gauge curve that is the same size as the actual gauge curve. The mean absolute error (meters) for the entire curve can be evaluated; the more similar the curves, the smaller the mean absolute error. The candidate model outputs can be ranked according to the mean absolute error for each gauge curve.

[0048] Social media data points 312 are found by searching and identifying useful images and videos associated with the flood. For example, images and videos associated with the flood may be found in tweets containing the flood hashtag, YouTube™, Facebook™, and other media posts, Google™ Images, local news, traffic cameras, etc. Geographic processing may be performed to identify where the image or video was taken. Various strategies for geolocating an image or video include finding something in the image to identify the location (e.g., the name of a business or a street sign), using pulses or landmarks in the image to identify the general location of the image and then using Google Maps™ to identify the location, using information provided outside the image (e.g., video description, text in a news article, text in a tweet, etc.), or randomly "searching" streets using Google Street View™ to attempt to identify the location. Water depth can be estimated by comparing images associated with the flood event, where water levels are shown, to a reference image such as Google Street View, and comparing the water level relative to specific buildings and reference objects such as vehicles, street curbs, and people.

[0049] SNS data points can potentially provide a large amount of data because they are often abundant and concentrated in interesting urban areas. However, timestamps may be inaccurate and / or missing (many sources of SNS data do not provide the time the image was taken), and depth estimates and geographic locations derived from SNS data points may be inaccurate. Thus, SNS data can provide water depth estimates but not timestamp information. It may be assumed that the images / videos are actual flooded images / videos taken during the flood event. Figure 5 shows a representation of social networking service (SNS) data points 502 in an example flood area 500.

[0050] Because SNS data may not provide timestamps, it typically does not provide a chronological picture of the flood event, particularly whether the flood subsequently rose above the observed water level at the SNS point. Comparing the multiple candidate model outputs with the SNS data may include establishing a water depth baseline at the geographic location based on the SNS data, comparing the simulated water depth at the geographic location in each of the multiple candidate model outputs to the baseline, and evaluating whether the simulated water depth at the geographic location meets the baseline, and if not, evaluating how low the simulated water depth at the geographic location is relative to the depth in the SNS data. That is, the baseline from the SNS data is that the simulation generated by the hydrological model should reach a water depth value equal to or greater than the water depth estimated from the SNS data at some point throughout the simulation time / flood event. This leads to an indication of asymmetry. That is, if the simulated flood simulates a water depth at a point in a geographic location that is greater than the water depth estimated from the SNS data, the candidate model output satisfies the SNS index; if the simulated flood simulates a water depth at a geographic location that is less than the water depth estimated from the SNS data, the candidate model output evaluates how much lower the simulated flood was than the SNS point should have been. Thus, the evaluation of a simulated flood may be whether the simulated flood meets the SNS data, and if not, how much lower the simulated water depth value is compared to the SNS data. The simulated flood evaluations can be calculated for each point separately and combined as a mean absolute error, which can be converted into a ranking similar to a gauge.

[0051] EO data 314 includes image data obtained from observation systems such as radar, satellites, etc. Image data for a region at a given time is converted using a neural network into a probability map that predicts the probability of flooding in discretized areas of the region at that time. Thus, the probability map can be thought of as geospatial data, e.g., a raster, overlaying the region of a flood event at a given time and associated with a probability indicating whether discretized areas within the region are inundated. There are areas with a high probability of being inundated, areas with a low probability of being inundated, and other areas with probabilities that indicate uncertainty. The areas with a high probability of being inundated and areas with a low probability of being inundated are used to evaluate the hydrological model, and the areas that indicate uncertainty may be excluded. An EO data output vector may be calculated by setting non-inundated areas to -1 and inundated areas to 1. The areas used in the EO data output vector are used to evaluate the output of the hydrological model and may be probabilistically sampled to meet the allocation of high- and low-probability areas. Additionally or alternatively, a target area may be determined using, for example, land use / land cover data, and the areas used to evaluate the model may be sampled from that target area. For example, open fields may be selected as target areas because Earth observation data are likely to be highly accurate, whereas areas such as urban areas or waterways may not be selected as target areas.

[0052] To evaluate the simulated flood, an area can be considered inundated when a predetermined water depth (e.g., 20 cm) is simulated in that area. Thus, the simulated flood area in the candidate model output may be converted to a candidate model output vector of -1s or 1s depending on the simulated flood. Then, the cross-entropy of the EO data output vector with respect to the candidate model output vector is calculated, converted to an index, and ranked.

[0053] As described above, each candidate model output in a given iteration is compared to each external data set and scored / ranked. That is, candidate model outputs may be scored according to their match with one or more gauge curves, one or more SNS data points, and one or more EO data sets (as well as any other type of external data). For example, candidate model outputs may be scored based on the mean absolute error between the simulated water depth at the gauge location and the external gauge data, and may be scored if they meet the SNS data points (i.e., if the simulated flood simulates a water depth at a point in a geographic location that is greater than the water depth estimated from the SNS data) and penalized if they do not meet the SNS data points (i.e., if the simulated water depth is less than the water depth estimated from the SNS data), with the amount of this penalty optionally further determined based on how much lower the simulated water depth is than the water depth estimated from the SNS points; and / or candidate model outputs may be scored according to how well the candidate model output vector matches the EO data output vector. A global ranking may be calculated taking into account scores indicating how well the candidate model output matches each external data set. Different weightings may be used in calculating the global rankings, for example, scores calculated for SNS data points may be weighted less than scores calculated for gauge data or EO data due to the higher uncertainty of SNS data points.

[0054] Candidate input datasets that produce one or more candidate model outputs with a satisfactory ranking (e.g., the best candidate model output, or the top n candidate model outputs) are identified. These candidate input datasets are then validated to produce candidate model outputs that provide satisfactory model outputs. For example, through the optimization process described with reference to FIG. 2, the candidate input datasets can be used to generate candidate input datasets for the next iteration, thereby optimizing the input data used to run the hydrological model.

[0055] 6 illustrates a flood estimation method 600. The flood estimation method 600 may be implemented, for example, by one or more computing devices 102 in FIG.

[0056] In method 600, input data associated with a flood event for running a hydrological model is received 602. As previously mentioned, the input data may include a combination of flood-specific data (e.g., rainfall data, input gauge data, etc.), which may be temporal data, and model control parameters (e.g., infiltration rates, friction parameters, digital elevation models, land use / land cover data, etc.), which may be spatial data. Additionally, external flood data for the flood event is received 604, which may include, for example, gauge data, SNS data points, Earth Observation data, etc.

[0057] In some embodiments of the flood estimation method, a hydrological model may be selected 606. For example, different hydrological models may be available for modeling flood events. The different hydrological models may be run using the same set of input data and model outputs may be compared to external flood data. The hydrological model that produced satisfactory model outputs for the external flood data may be selected.

[0058] In some examples, one or more model control parameters in the input data may be selected (608). It should also be appreciated that some parameters in the input data, such as rainfall data, may be easily repeatable, while model control parameters, such as the digital elevation model, may not be suitable for repeated updating. Different digital elevation models may be available, and method 600 may include running hydrological models with identical sets of input data except for the digital elevation model and comparing the different model outputs with the external flood data. The digital elevation model used for the input data that produced satisfactory model outputs for the external flood data may be selected.

[0059] Furthermore, boundary conditions for the hydrological model may be generated (610). The boundary conditions may be generated in response to either or both of the input data and external data and may be used as boundaries to constrain the generation of candidate input datasets, as described below. For example, boundary conditions may represent locations where water flows into or out of the simulation domain due to external factors. Rainfall, lakes, rivers, stormwater manholes, and groundwater wells are all examples of boundary conditions. For example, rainfall may be measured as rainfall amount per unit time for each grid within the simulation domain, or may be measured for a single rain gauge within the simulation domain. Rainfall as a boundary condition may be set as a different rainfall amount per unit time for each grid, or as the same rainfall amount per unit time for the entire simulation domain. As another example, the flow rate (outflow) from a lake or upstream river may be measured by a sensor installed at the outlet gauge. Therefore, the contribution of the lake or river per unit time may be set as input data from the gauge as a boundary condition. Other measurements, such as stormwater runoff from a manhole, may also be useful in defining boundary conditions. Additionally, as noted above, the input data for the simulation may include simulated output data from a simulation performed for an upstream region, and such upstream data may be used to define boundary conditions for the simulation of the downstream region.

[0060] The flood estimation method 600 validates 612 the input data based on external flood data over one or more iterations. Specifically, in each iteration, multiple candidate input data sets are generated and used to run the hydrological model. The multiple candidate model outputs are compared to the external flood data to identify one or more satisfactory candidate input data sets that may be used in the next iteration. In this manner, the input data is validated by evaluating whether the input data produces satisfactory model outputs.

[0061] 7 shows a method 700 for validating input data. Method 700 may be performed at each iteration of validating input data at 612 in method 600.

[0062] A plurality of candidate input datasets for the current iteration are generated (702). The plurality of candidate input datasets for the first iteration are generated from the received input data, and the plurality of candidate input datasets for subsequent iterations are generated based on one or more satisfactory candidate input datasets from previous iterations. Generating the plurality of candidate input datasets may include stochastically sampling offsets for the input data (i.e., applying a Monte Carlo method) and applying offsets to the input data. If the input data includes a single value (e.g., friction), a single offset may be sampled from a distribution of offsets for that value. If the input data includes a time series of values ​​(e.g., water level data, rainfall data, etc.), both a value offset (e.g., water level, rainfall amount) and a time offset may be sampled from their respective distributions. In some embodiments, the distribution of offsets may be a Gaussian distribution set according to expected offset values ​​based on instrument error, position error, etc.

[0063] The hydrological model is run with multiple candidate input datasets to generate multiple candidate model outputs 704. As previously described, running the hydrological model with multiple candidate input datasets may include running the hydrological model in parallel for each of the multiple candidate input datasets in the current iteration. Thus, one computing device may instruct multiple other computing devices to run the hydrological model in parallel using different candidate input datasets.

[0064] The plurality of candidate model outputs are compared to the external flood data to identify one or more satisfactory candidate input data sets for the current iteration (706). As described above, each of the plurality of candidate model outputs may be scored according to how well it matches the external flood data. Comparing the plurality of candidate model outputs to the external flood data may include determining the average error between the candidate model output and the external flood data (e.g., when comparing the candidate model output to gauge data). Additionally or alternatively, comparing the plurality of candidate model outputs may include determining whether the candidate model output meets a flood criterion set by the external flood data, and, if the candidate model output does not meet the flood criterion, determining the distance to the flood criterion (e.g., when comparing the candidate model output to SNS data points or Earth Observation data). One or more satisfactory candidate model outputs may be identified by ranking the plurality of candidate outputs according to the above scoring and identifying a predetermined number of candidate outputs with the highest scores, or by identifying one or more satisfactory candidate outputs as candidate outputs with a score above a threshold. The external flood data may include a plurality of external datasets, and comparing each of the plurality of candidate model outputs to the external flood data may include comparing each of the plurality of candidate model outputs to each of the plurality of external datasets. The candidate model outputs may be ranked according to a calculated global score based on how well the candidate model outputs fit all of the external datasets.

[0065] From identifying the satisfactory candidate model outputs, one or more satisfactory candidate input datasets are identified 708. The satisfactory candidate input datasets may be used in subsequent iterations (i.e., used to generate multiple candidate input datasets in subsequent iterations at 702), or the model may be run using one of the satisfactory candidate input datasets or an ensemble of multiple satisfactory candidate input datasets.

[0066] Specifically, referring back to Figure 6, it is determined whether validation is complete 614. For example, validating the input data may include determining, in each iteration, a measure (e.g., a score) of how well each of the plurality of candidate model outputs matches the external flood data, and validation is performed until (a) each of the plurality of candidate model outputs in the current iteration is within a predetermined threshold of the external flood data, (b) each of the plurality of candidate model outputs in the current iteration is unchanged from the previous iteration, or (c) a predetermined number of iterations have occurred without satisfying (a) or (b).

[0067] If validation is not complete (NO at 614), the method returns to 612 to continue validating the input data (i.e., perform another iteration according to method 700). If validation is complete (YES at 614), a hydrological model is estimated using the validated input data to estimate water levels for the flood event (616). The validated input data is based on candidate input data sets that provide candidate model outputs with a satisfactory fit to external flood data. In some embodiments, method 600 may further include generating and displaying a peak flood map for the flood event from the water level estimates.

[0068] Depending on the input data validation method and the parallelism with which the hydrological model can be implemented using different candidate input data sets, method 600 can be implemented in a relatively short time (e.g., on the order of minutes or hours) and at a relatively low cost. Thus, the flood estimation method may be run each time the input data and / or external flood data are updated, using the most recent available data in the flood estimation method.

[0069] 8A-8C illustrate the iterative refinement of candidate outputs toward external gauge data according to the flood estimation method. As previously described, a gauge curve (solid black line 802) is generated from actual gauge data providing a time series of water depth or water level at a specific location. Simulated values ​​of water depth / water level from candidate model outputs generated from different candidate input datasets (light gray line 804) are compared to the gauge curve. FIG. 8A shows the time evolution of simulated values ​​of water depth / water level generated by 100 candidate model outputs generated from 100 candidate input datasets in the first iteration.

[0070] As the candidate input datasets are updated and validated through the process of iterative optimization of the flood estimation method, the candidate model outputs 804 converge towards the gauge curve 802 and each other, as shown in Figures 8B and 8C (note the scale of the y-axis). Figures 8B and 8C show the time evolution of simulated values ​​of water depth / level produced by 100 candidate model outputs generated from the 8th and 30th iterations, respectively.

[0071] It should be appreciated that in Figure 8C, there are still differences between the multiple candidate model outputs and the gauge curve 802. This is because, as described above, each of the multiple candidate model outputs is compared to each external data set of external data, and accordingly, satisfactory candidate model outputs can be identified according to their degree of agreement with all external data sets (e.g., based on a global score). In this example of Figures 8A-8C, the candidate model outputs were compared with the SNS data points as well as the data sets of the other three gauges.

[0072] 9A-9D show examples of different peak water depth outputs simulated using different candidate input data sets. Graphs 900a, 900b, 900c, and 900d visually illustrate the simulated flooding by showing peak water depths at different locations over the area of ​​interest.

[0073] Figure 10 shows an example of a peak flood map 1000. According to the flood estimation method described herein, the model output has been satisfactorily validated against external data, allowing accurate flood peaks to be determined.

[0074] Figure 11 shows a representation of uncertainty areas 1102 in a simulated flood. Based on available input data, some areas of the simulated flood may be uncertain, and accordingly, the candidate model output of the flood estimation method may not converge to the external flood data in these areas.

[0075] The embodiments have been described above with reference to flow, sequence, and block diagrams of methods, apparatuses, systems, and computer program products. In this regard, the illustrated flow, sequence, and block diagrams illustrate the architecture, functionality, and operation of example embodiments of various embodiments. For example, each block in the flow and block diagrams, as well as one or more actions in the sequence diagrams, may represent a module, segment, or portion of code that includes one or more executable instructions for performing the specified operation. In some alternative embodiments, one or more actions described in the blocks or operations may be executed in a different order than described in the diagrams. For example, two blocks or operations shown in succession may, in some embodiments, be executed substantially simultaneously, or the blocks or operations may be executed in the reverse order, depending on the functionality involved. While the above provides several specific examples, these examples are not necessarily the only examples. Each block in the flow and block diagrams, as well as each operation in the sequence diagrams, and combinations of these blocks and operations, may be implemented by a dedicated hardware-based system that performs the specified functions or operations, or a combination of dedicated hardware and computer instructions.

[0076] The terms used herein are for the purpose of describing particular embodiments only and are not intended to be limiting. Thus, as used herein, the singular forms "a," "an," and "the" are intended to encompass the plural unless the context clearly dictates otherwise (e.g., the use of "a" or "the" in "the challenge" in the claims does not exclude the use of plural challenges). It is further understood that, as used herein, the terms "comprises" and "comprising" refer to the presence of the specified features, integers, steps, operations, elements, or components, but do not exclude the presence of other features, integers, steps, operations, elements, components, or groups. In the following description, directional terms such as "top," "bottom," "upwards," "downwards," "vertically," and "laterally" are used merely to enable relative reference and are not intended to impose any restrictions on how the articles used should be positioned, assembled, or disposed relative to their environment when used. Additionally, in this description, the term "connect" and its variations, such as "connected," "connects," and "connecting," are used to encompass both direct and indirect connections unless otherwise specified. For example, when a first device is connected to a second device, the connection may be direct or indirect through other devices or connections. Similarly, when a first device and a second device are communicatively connected, the connection may be direct or indirect through other devices or connections. As used herein, the term "and / or" means any one or more of the listed items. For example, "A, B, and / or C" means "any one or more of A, B, and C."

[0077] It is contemplated that any part of any aspect or embodiment described herein can be implemented or combined with any part of any other aspect or embodiment described herein.

[0078] The scope of the claims should not be construed as being limited by the above-described exemplary embodiments, but should be given the broadest interpretation consistent with the entire description.

[0079] It should be understood that the features and aspects of the various embodiments described above may be combined in other embodiments within the spirit and scope of the present disclosure. Also, the figures are not to scale, and sizes and shapes may be exaggerated for illustrative purposes.

[0080] In this specification and claims, the terms "comprises" and "comprising" and variations thereof mean the inclusion of the specified features, steps, and integers. These terms are not intended to exclude the presence of other features, steps, or components.

[0081] The invention may be broadly referred to as consisting of any combination of any two or more of the components, elements, steps, examples and / or features individually or collectively referred to or illustrated herein, and in particular, one or more features of any embodiment described herein may be combined with one or more features of any other embodiment.

[0082] Features disclosed in one or more of the publications referenced herein may be claimed for protection in combination with the present disclosure.

[0083] Although specific exemplary embodiments of the present invention have been described, the appended claims are not intended to be limited to only these embodiments. The claims should be construed literally, pro formaly, and / or to encompass equivalents.

Claims

1. receiving input data associated with a flood event for running a hydrological model; receiving external flood data for the flood event; validating the input data based on the external flood data over one or more iterations, wherein in each of the one or more iterations: generating a plurality of candidate input datasets for a current iteration, the plurality of candidate input datasets for a first iteration generated from the input data, and the plurality of candidate input datasets for a subsequent iteration generated based on one or more satisfactory candidate input datasets from a previous iteration; running the hydrological model with the plurality of candidate input datasets to generate a plurality of candidate model outputs; comparing the plurality of candidate model outputs to the external flood data to identify one or more satisfactory candidate input data sets for the current iteration; validating the input data by running the hydrological model using the validated input data based on a candidate input dataset that provides a candidate model output with a satisfactory fit to the external flood data to estimate water levels of the flood event; Flood estimation methods, including:

2. 2. The flood estimation method of claim 1, wherein the plurality of candidate input datasets in the current iteration are generated for the first iteration from a probabilistic sampling of offsets of the input data, and for subsequent iterations from a probabilistic sampling of offsets of the one or more satisfactory candidate input datasets in a previous iteration.

3. The flood estimation method of claim 2 , wherein the offsets are probabilistically sampled from a distribution of offsets in the input data.

4. The flood estimation method according to claim 3 , wherein the distribution of the offsets is a Gaussian distribution.

5. and determining, at each iteration, a measure of how well each of the plurality of candidate model outputs matches the external flood data; Validating the input data over the one or more iterations includes: (a) until each of the plurality of candidate model outputs in the current iteration is within a predetermined threshold of the external flood data; (b) until each of the plurality of candidate model outputs in the current iteration remains unchanged from the previous iteration; or (c) until (a) or (b) is not satisfied and a predetermined number of repetitions are performed; The flood estimation method according to claim 1 , wherein the method is carried out by

6. 6. The flood estimation method of claim 1, wherein running the hydrological model with the plurality of candidate input datasets comprises running the hydrological model in parallel for each of the plurality of candidate input datasets in the current iteration.

7. 6. The flood estimation method of claim 1, wherein comparing each of the plurality of candidate model outputs with the external flood data comprises scoring each of the plurality of candidate model outputs according to how well it matches the external flood data.

8. the external flood data includes a plurality of external data sets; 8. The flood estimation method of claim 1, wherein comparing each of the plurality of candidate model outputs with the external flood data comprises comparing each of the plurality of candidate model outputs with each of the plurality of external data sets.

9. Comparing the plurality of candidate model outputs to the external flood data includes: determining a mean error between the candidate model output and the external flood data; determining whether the candidate model output satisfies a flood criterion established by the external flood data, and if the candidate model output does not satisfy the flood criterion, determining a distance to the flood criterion; 9. The flood estimation method according to claim 1, further comprising one or more of:

10. The flood estimation method according to claim 1 , wherein the external flood data comprises one or more of earth observation data, gauge data, and social networking service data.

11. the external flood data includes gauge data; 11. The flood estimation method of claim 10, wherein comparing the plurality of candidate model outputs with the external flood data includes evaluating a mean absolute error between the gauge data in the external flood data and a sampling point from each of the plurality of candidate model outputs set at a position of the gauge data.

12. The external flood data includes social networking service (SNS) data, 12. The flood estimation method of claim 10 or 11, wherein comparing the plurality of candidate model outputs with the external flood data comprises: setting a water depth criterion at a geographic location based on the SNS data; comparing a simulated water depth at the geographic location for each of the plurality of candidate model outputs to the criterion; evaluating whether the simulated water depth at the geographic location meets the criterion; and if not, evaluating how shallow the simulated water depth at the geographic location is relative to the water depth from the SNS data.

13. the external flood data includes Earth observation data; 13. The flood estimation method of claim 10, wherein comparing the plurality of candidate model outputs with the external flood data comprises predicting a probability of flooded areas within the domain of the Earth observation data, calculating an Earth observation data output vector using areas associated with areas with a high probability of being flooded and areas with a low probability of being flooded, generating a candidate model output vector for each of the plurality of candidate model outputs based on areas simulated to be flooded and areas simulated not to be flooded according to a predetermined simulated water depth, and comparing the Earth observation data output vector with the candidate model output vector for each of the plurality of candidate model outputs by calculating a cross-entropy between the Earth observation data output vector and the candidate model output vector.

14. 14. A flood estimation method according to any one of claims 1 to 13, further comprising generating boundary conditions from one or both of the input data and the external data, the boundary conditions being used as boundaries to constrain the plurality of candidate input data sets.

15. selecting the hydrological model from among a plurality of hydrological models; selecting the hydrological model comprises: running each of the plurality of hydrological models using the same input data set to generate a respective model output for each of the plurality of hydrological models; comparing each of the model outputs with the external flood data; selecting the hydrological model that produced satisfactory model output for the external flood data; 15. The flood estimation method of claim 1, comprising:

16. 16. A flood estimation method according to any preceding claim, further comprising generating and displaying a peak flood map of the flood event from the water level estimate.

17. 17. The flood estimation method of claim 1, wherein the input data comprises measured data, forecast data, or a combination thereof.

18. a database containing input data and external flood data associated with flood events; a processor; a non-transitory computer readable medium storing computer program code executable by the processor, the computer program code executing the flood estimation method of any one of claims 1 to 17; A system comprising:

19. further comprising a plurality of processors configured to operate in parallel; 20. The system of claim 18, wherein the multiple processors are configured to run a hydrological model in parallel using each of a plurality of probabilistic input datasets to generate a plurality of candidate outputs.

20. A non-transitory computer readable medium having stored thereon computer program code executable by a processor, the computer program code causing the processor to perform the method of any one of claims 1 to 17 when executed by the processor.

Images