Systems and methods for generating flood hazard estimation using machine learning model and satellite data

Patent metadata

Field	Value
Publication number	US-2022108504-A1
Application number	US-202117493788-A
Country	US
Kind code	A1
Filing date	Oct 4, 2021
Priority date	Oct 5, 2020
Publication date	Apr 7, 2022
Grant date	—

Abstract

A system and method for flood hazard estimation inputs a satellite elevation map and applies a machine learning model to output a geographic map representing flood hazard areas. The machine learning model is trained using a generative adversarial network (GAN) to produce an output of a deterministic hazard mapping algorithm. A GAN objective applies a loss function, reweighted to increase the importance of high hazard areas. The method retrieves a DEM topography file representing elevation data of an identified terrain, and applies a sink-filling algorithm to detect and fill sinks in the DEM topography. The algorithm subtracts the DEM elevation data to generate a filled topography, and identifies flattest regions of the filled topography. The algorithm then generates a flood hazard map by merging the filled topography and the DEM elevation data, using a weighting function that balances the detected sinks and the flattest regions of the filled topography.

First claim

What is claimed is:

1 . A computer-implemented method comprising:
inputting, by a computer, a satellite elevation map; applying, by the computer, a machine learning model that is trained using a generative adversarial network to produce an output of a deterministic hazard mapping algorithm; and outputting, by the computer, a map representing flood hazard areas.

2 . The method of claim 1 , wherein the generative adversarial network comprises a deep learning convolutional neural network trained to abstract the satellite elevation map to a flooding-related feature map.

3 . The method of claim 1 , wherein the map representing flood hazard areas identifies geographic areas most vulnerable to pluvial flooding.

4 . The method of claim 1 , wherein the generative adversarial network includes an objective function, wherein in training the machine learning model a generator of the generative adversarial network tries to minimize the objective function against an adversarial of the generative adversarial network that tries to maximize the objective function.

5 . The method of claim 4 , wherein the objective function of the generative adversarial network includes a reconstruction loss with a weighting mechanism that increases importance of high hazard areas.

6 . The method of claim 5 , wherein the reconstruction loss comprises an L1 loss function, reweighted to increase the importance of the high hazard areas.

7 . The method of claim 6 , wherein the L1 loss function is reweighted via the weighting factor w=2.5y+2.5, clamped between 0.02 and 1.

8 . The method of claim 1 , wherein the generative adversarial network comprises a deep learning neural network, wherein the satellite elevation map is a digital elevation model (DEM) topography representing elevation data of an identified terrain, wherein the deep learning neural network applies the deterministic hazard mapping algorithm to the DEM topography to detect and fill sinks in the DEM topography.

9 . The method of claim 8 , wherein the deep learning neural network further applies the deterministic hazard mapping algorithm to subtract the elevation data of the DEM topography to generate a hidden layer representing a filled topography map and to identify flattest regions of the hidden layer representing the filled topography map.

10 . The method of claim 9 , wherein the deep learning neural network further applies the deterministic hazard mapping algorithm in an output layer of the deep learning neural network merging the hidden layer representing the filled topography map and the DEM topography representing the elevation data of the identified terrain.

11 . The method of claim 10 , wherein the merging the hidden layer representing the filled topography map and the DEM topography generates a weighted combination of the filled topography map and the DEM topography balancing the identified flattest regions of the filled topography map versus the detected sinks in the DEM topography.

12 . A computer-implemented method comprising:
retrieving, by the computer, a digital elevation model (DEM) topography file representing elevation data of an identified terrain; applying a sink-filling algorithm to the DEM topography file, wherein the sink-filling algorithm detects and fills sinks in the DEM topography file; subtracting the elevation data of the DEM topography file to generate a filled topography file; analyzing, by the computer, the filled topography file to identify flattest regions of the filled topography file; and generating, by the computer, a flood hazard map file by applying a weighting function in merging the filled topography file and the DEM elevation data, wherein the weighting function balances the detected sinks in the DEM elevation data and the flattest regions of the filled topography file.

13 . A system comprising:
a server comprising a processor and a non-transitory computer-readable medium containing instructions that when executed by the processor causes the processor to perform operations comprising: input a satellite elevation map; apply a machine learning model this is trained using a generative adversarial network to produce an output of a deterministic hazard mapping algorithm; and output a map representing flood hazard areas.

14 . The system of claim 13 , wherein the generative adversarial network includes an objective function, wherein in training the machine learning model a generator of the generative adversarial network tries to minimize the objective function against an adversarial of the generative adversarial network that tries to maximize the objective function.

15 . The system of claim 14 , wherein the objective function of the generative adversarial network includes a reconstruction loss with a weighting mechanism that increases importance of high hazard areas.

16 . The system of claim 15 , wherein the reconstruction loss comprises an L1 loss function, reweighted to increase the importance of the high hazard areas.

17 . The system of claim 13 , wherein the generative adversarial network comprises a deep learning neural network, wherein the satellite elevation map is a digital elevation model (DEM) topography representing elevation data of an identified terrain, wherein the deep learning neural network applies the deterministic hazard mapping algorithm to the DEM topography to detect and fill sinks in the DEM topography.

18 . The system of claim 17 , wherein the deep learning neural network further applies the deterministic hazard mapping algorithm to subtract the elevation data of the DEM topography to generate a hidden layer representing a filled topography map and to identify flattest regions of the hidden layer representing the filled topography map.

19 . The system of claim 18 , wherein the deep learning neural network further applies the deterministic hazard mapping algorithm in an output layer of the deep learning neural network merging the hidden layer representing the filled topography map and the DEM topography representing the elevation data of the identified terrain.

20 . The method of claim 19 , wherein the merging the hidden layer representing the filled topography map and the DEM topography generates a weighted combination of the filled topography map and the DEM topography balancing the identified flattest regions of the filled topography map versus the detected sinks in the DEM topography.

Assignees

• Bank Of Montreal

Inventors

• Kenyon-Dean Kian
• Zhao Bo
• Kasiri Keyvan
• Vahlis Yevgeniy
• Fraser Todd
• Morrison Lyndsay
• Torrance Michael
• Wu Stella

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title