Predicting wildfires on the basis of biophysical indicators and spatiotemporal properties using a long short term memory network

Patent metadata

Field	Value
Publication number	US-11275989-B2
Application number	US-201715601739-A
Country	US
Kind code	B2
Filing date	May 22, 2017
Priority date	May 22, 2017
Publication date	Mar 15, 2022
Grant date	Mar 15, 2022

Abstract

The present disclosure involves systems, software, and computer implemented methods for predicting wildfires on the basis of biophysical indicators and spatiotemporal properties. A method includes receiving a request for a wildfire prediction for at least one geographical area. At least one biophysical indicator is identified. Each biophysical indicator provides biophysical data for the at least one geographical area. The at least one biophysical indicator is provided to a long short term memory (LSTM) network. The LSTM network includes a convolutional neural network (CNN) for each of multiple LSTM units. Each LSTM unit and each CNN are associated with a historical time period in a time series. The LSTM is used to generate at least one prediction for wildfire risk for the at least one geographical area for an upcoming time period. The at least one prediction is provided responsive to the request.

First claim

What is claimed is:

1. A computer-implemented method for providing a wildfire prediction, the computer-implemented method comprising:
receiving a request for the wildfire prediction for at least one geographical area;
retrieving image data for the at least one geographical area, wherein the image data comprises overhead images of the at least one geographical area for at least one time point;
generating, from the image data, at least one biophysical indicator, each biophysical indicator providing image-based biophysical data for the at least one geographical area;
providing the at least one biophysical indicator to a long short term memory (LSTM) network, the LSTM network comprising a LSTM layer and a fully connected layer, the LSTM layer comprising a series of LSTM units, each LSTM unit in the series of LSTM units being connected to a respective convolutional neural network (CNN), each LSTM unit and each respective CNN being associated with a particular historical time period in a time series, wherein each LSTM unit uses input from the respective CNN and each LSTM unit, except a last LSTM unit in the series of LSTM units, generates output data for use by a successive LSTM unit in the series of LSTM units, and wherein the output data generated by the last LSTM unit in the series of LSTM units of the LSTM layer is transmitted to the fully connected layer of the LSTM to generate at least one prediction for a wildfire risk for the at least one geographical area for an upcoming time period;
evaluating the at least one prediction to generate at least one corresponding prediction evaluation by comparing the at least one prediction to ground truth data for the at least one geographical area using categorical cross-entropy as an objective function; and
providing the at least one prediction and the at least one corresponding prediction evaluation responsive to the request.

2. The computer-implemented method of claim 1 , wherein outputs of each CNN except a last CNN associated with a last historical time period are respectively provided to a CNN associated with a more recent time period in the time series.

3. The computer-implemented method of claim 2 , wherein the outputs of each CNN are outputs generated by a hidden layer of a respective CNN.

4. The computer-implemented method of claim 2 , wherein outputs of the last CNN are provided to the fully connected layer of the LSTM.

5. The computer-implemented method of claim 4 , wherein the fully connected layer is used to generate the at least one prediction.

6. The computer-implemented method of claim 4 , wherein the at least one prediction comprises a first output neuron and a second output neuron, the first output neuron and the second output neuron indicating a probability of a wildfire and a probability of no wildfire for the geographic area for an upcoming time period, respectively.

7. The computer-implemented method of claim 1 , wherein providing the at least one prediction comprises presenting prediction information on a map that displays the at least one geographic area.

8. The computer-implemented method of claim 1 , wherein evaluating the at least one prediction to generate at least one corresponding prediction evaluation by comparing the at least one prediction to ground truth data for the at least one geographical area using categorical cross-entropy as an objective function comprises:
evaluating the at least one prediction to generate the at least one corresponding prediction evaluation by comparing the at least one prediction for the wildfire risk to the ground truth data for the at least one geographical area using categorical cross-entropy as the objective function, wherein the ground truth data comprises historical information about wildfires for the at least one geographical area.

9. A system, comprising:
at least one processor; and
a memory communicatively coupled to the at least one processor, the memory storing instructions which, when executed by the at least one processor, cause the at least one processor to perform operations comprising:
receiving a request for a wildfire prediction for at least one geographical area;
retrieving image data for the at least one geographical area, wherein the image data comprises overhead images of the at least one geographical area for at least one time point;
generating, from the image data, at least one biophysical indicator, each biophysical indicator providing image-based biophysical data for the at least one geographical area;
providing the at least one biophysical indicator to a long short term memory (LSTM) network, the LSTM network comprising a LSTM layer and a fully connected layer, the LSTM layer comprising a series of LSTM units, each LSTM unit in the series of LSTM units being connected to a respective convolutional neural network (CNN), each LSTM unit and each respective CNN being associated with a particular historical time period in a time series, wherein each LSTM unit uses input from the respective CNN and each LSTM unit, except a last LSTM unit in the series of LSTM units, generates output data for use by a successive LSTM unit in the series of LSTM units, and wherein the output data generated by the last LSTM unit in the series of LSTM units of the LSTM layer is transmitted to the fully connected layer of the LSTM to generate at least one prediction for a wildfire risk for the at least one geographical area for an upcoming time period;

evaluating the at least one prediction to generate at least one corresponding prediction evaluation by comparing the at least one prediction to ground truth data for the at least one geographical area using categorical cross-entropy as an objective function; and
providing the at least one prediction and the at least one corresponding prediction evaluation responsive to the request.

10. The system of claim 9 , wherein outputs of each CNN except a last CNN associated with a last historical time period are respectively provided to a CNN associated with a more recent time period in the time series.

11. The system of claim 10 , wherein the outputs of each CNN are outputs generated by a hidden layer of a respective CNN.

12. The system of claim 10 , wherein outputs of the last CNN are provided to the fully connected layer of the LSTM.

13. The system of claim 12 , wherein the fully connected layer is used to generate the at least one prediction.

14. The system of claim 13 , wherein the at least one prediction comprises a first output neuron and a second output neuron, the first output neuron and the second output neuron indicating a probability of a wildfire and a probability of no wildfire for the geographic area for an upcoming time period, respectively.

15. A non-transitory computer-readable media storing instructions which, when executed by at least one processor, cause the at least one processor to perform operations comprising:
receiving a request for a wildfire prediction for at least one geographical area;
retrieving image data for the at least one geographical area, wherein the image data comprises overhead images of the at least one geographical area for at least one time point;
generating, from the image data, at least one biophysical indicator, each biophysical indicator providing image-based biophysical data for the at least one geographical area;
providing the at least one biophysical indicator to a long short term memory (LSTM) network, the LSTM network comprising a LSTM layer and a fully connected layer, the LSTM layer comprising a series of LSTM units, each LSTM unit in the series of LSTM units being connected to a respective convolutional neural network (

Assignees

• Sap Se

Inventors

• Tschernezki Vadim
• Blum Oliver
• Gildhoff Hinnerk
• Wyss Michèle
• Deiseroth Bjoern
• Svojanovsky Wenzel

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/082 Primary

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/0442 Primary

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title