Brute Force

https://1drv.ms/p/s!AmvxZ4n3-qs7jWEWkBANj6jDzHb3?e=B3suDR

https://github.com/ConstantinosAr/Wildfire_prediction_AI_model/

As a first step, we read through various papers to understand the mechanism and the main factors that could lead to an increased probability of a fire emerging. After reading through a few papers, factors like surface temperature, drought, atmospheric pressure, and elevation as the main contributors. To get access through the various data, Meteomatics API website was utilized to get access to the parameters mentioned before. This data will later be used for training a neural network model. It should be mentioned that as part of training, we did not use any human impact data like the population density or topological data which were also found to affect the probability of fire. With additional time, the training data could include these parameters since they are included in a specific format in the Meteomatics API. We used data from Kaggle about fire incidents in the period from 2013 to 2020 as cases where the fire probability was high.

For the training of the classification model, we also generated some random data that were close to the region of California around the dates in which fires did not occur. This data was used for the model to have examples of the parameters when a fire wasn't taking place, to give a more accurate result after the training. A more precise classification of the risk factor at a given time was determined by the model from available data. Inputs were associated with fire occurrence to train the network via a number of layers with a different amount of units. The depth of the network and the corresponding number of neurons in each layer were optimally selected to increase the accuracy of the testing data. After finding the relation between the different parameters with the fire probability, the different risk indices were classified using the colours of green, yellow and red.

For the regression model that estimates the severity of the fire, a parameter about the acres burned by the fire was also used to provide a target output for the model. This model is trained to determine the severity of the fire based on data from the 7 days before a fire occurred. Data from the previous 7 days was chosen for the model to know approximately the conditions during the wildfire that could affect its growth and spread.

Regarding the interface after selecting a start date (to generate inputs for the models from a server), a map would be displayed where the user can see locations marked for a potential fire. Areas are split into 3 grid squares. Locations marked with orange or red (higher probability to highest), will have a circle around them, with a radius according to the severity of the fire. The probability is calculated with our classification model, and the severity radius with our regression model.

Taking into account data that can be easily accessible to meteorological institutions, the wildfire risk index can provide valuable data that can reduce the cases of wildfires occurring in the future. In our project, Python was utilized for coding the data acquisition and machine learning models. The benefits our model provides is that with accurate data it can create a model that would predict accurately the risk probability factor in a desired place.

The following data were gathered from the Meteomatics API:

• drought_index:idx
• elevation:m
• leaf_wetness:idx
• precip_6h:mm
• relative_humidity_50m:p
• soil_moisture_index_-15cm:idx
•  solar_power:kW
• t_max_200m_6h:C
• wind_speed_max_200m_6h:kmh

Also, data from Kaggle were collected to find cases where fires emerged that will be needed to train the model. The gathering of the aforementioned data helped us in training our model to find the correlation between each factor and the probability of fire

The Space Apps experience has been a lot of fun, as trying to develop something in a short period of time is a great challenge. For the competition we wanted to choose a subject that can have an impact on our planet. Our inspiration to choose this challenge were the recent wildfires in Cyprus, which highlighted the need for tools to help combat these disasters. The outcome of the fires showed how unprepared we currently are to suppress them, and how quickly they can develop to claim lives and burn large areas.

Our approach to the project was to make sure firstly that the data we collected from the APIs were useful to our model by doing a lot of research on the factors that can cause a wildfire. After we had the right data, we made sure the coding of our model was optimal, to produce the results we wanted from the model, by minimizing the error. Lastly, we wanted to develop an easy-to-use interface for our project. Our only setbacks were in the training of our model, as during the duration of the hackathon we lost full access to Meteomatics’ API, which we used to request data needed for the training. We couldn’t get a response in time from support to reestablish our access, thus our training was limited, which affected the model’s results.

We would like to thank the Cyprus Space Exploration Organization for their support during the hackathon, and their willingness to work with our team to develop and potentially release our project in the future.

1. Polash Banerjee. (2021) Maximum entropy-based forest fire likelihood mapping: analyzing the trends, distribution, and drivers of forest fires in Sikkim Himalaya. Scandinavian Journal of Forest Research 36:4, pages 275-288
2. Ramesh P. Singh. (2020) Earth observation and sustainable development goals. Geomatics, Natural Hazards and Risk 11:1, pages i-vi.
3. Benson, R.P., Roads, J.O. and Weise, D.R., 2008. Climatic and weather factors affecting fire occurrence and behavior. Developments in Environmental Science, 8, pp.37-59.
4. González, J.R., Palahí, M., Trasobares, A. and Pukkala, T., 2006. A fire probability model for forest stands in Catalonia (north-east Spain). Annals of Forest Science, 63(2), pp.169-176.
5. Preisler, H.K., Brillinger, D.R., Burgan, R.E. and Benoit, J.W., 2004. Probability based models for estimation of wildfire risk. International Journal of wildland fire, 13(2), pp.133-142.
6. Kim, S.J., Lim, C.H., Kim, G.S., Lee, J., Geiger, T., Rahmati, O., Son, Y. and Lee, W.K., 2019. Multi-temporal analysis of forest fire probability using socio-economic and environmental variables. Remote Sensing, 11(1), p.86.
7. https://www.kaggle.com/ananthu017/california-wildfire-incidents-20132020

Tools / Coding Languages/ Software Used:

• Python
• PyCharm
• Visual Studio
• Keras Tensorflow API
• Google Earth Engine

#wildfires #critical_infrastructure #disaster_probability #quick_response #geosptail_technology #satellite_data #data_analysis

This project has been submitted for consideration during the Judging process.