Sunflower Coconut| Spot That Fire V3.0

Awards & Nominations

Sunflower Coconut has received the following awards and nominations. Way to go!

Global Nominee

Spot That Fire V3.0

Recent wildfires worldwide have demonstrated the importance of rapid wildfire detection, mitigation, and community impact assessment analysis. Your challenge is to develop and/or augment an existing application to detect, predict, and assess the economic impacts from actual or potential wildfires by leveraging high-frequency data from a new generation of geostationary satellites, data from polar-orbiting environmental satellites, and other open-source datasets.

Fire Away

Summary

This project aims to detect wildfires with real time inference. Many technologies for wildfire detection currently is not in real time, and we would like to improve on existing technologies to provide a better wildfire detection system, so we could minimize environmental loss and human losses in case of a wildfire occurrence. The project uses existing convolutional neural networks models to provide a balance of speed and performance. The model is trained on data combined from multiple sources, as to increase the data amount.

How We Addressed This Challenge

We developed tools to immediately recognize fires in videos, created a new fire dataset and are working on fire risk and impact prediction.


Our project is to create a one-stop tool with machine learning to mitigate the effects of wildfires on humans and the biosphere in real time all around the world. Traditionally, data for fire prediction is accessed through the local observatory, and thus any predictions were local. To improve on the situation, we combined global datasets provided by NASA and other organizations on different areas such as surface temperature, windspeed, relative humidity, altitude and landscape to be our input and train them on the historical fire data (also from multiple sources) to predict the risk and potential impact of fires. We referenced our input parameters from research papers on mathematical modelling of fires and machine learning approaches to find the most impactful variables.  


We created a new image binary classification dataset of fires by combining data from MIT, Kaggle and DeepQuestAI and used OpenCV to capture images from online videos on wildfires. Originally there were only 6648 images, but after conducting data augmentation using albumentations, the dataset now has 27321 images. We hope this dataset can help talented programmers around create more fire-related apps so that we can all contribute to the cause.


We then made a wildfire-detection AI that can detect fires in videos or images in real time using our new dataset. This can be used in surveillance cameras inside forests or from phones of hikers or campers alike. Wildfires can be detected and reported to authorities quickly and the fire-fighting process can begin as soon as possible.


Currently, there are no public datasets with a comprehensive view of fire-causing factors and occurrences of fire. Therefore we are in the process of creating one by combining multiple sources. We are using the Google Maps API to get longitude and latitude information from word descriptions in current datasets, then map weather and geoform information of surrounding areas in the past.

How We Developed This Project

This project is inspired by the many wildfires recently happening all over the world, including but not limited to the recent Australian and American wildfire that killed many plants and animals, and made many civilian homes lost. We wish to make a wildfire detection system so firefighters can stop the spread of wildfire before it spreads into a large scale disaster.


Our approach to this project is to use the existing computer vision framework, EfficientNet, as it provides a good balance of performance and speed. Without real time inference, the detection system will not be able to prevent large scale spread of wildfire in time.


For the tools, we used 8 titan X in parallel to train the model, with python being the language of choice as the libraries and existing code helps to make a prototype quicker. If this is made into an actual product, other languages might be used as the performance of python is not that great.

During the making of the project, we encountered many problems and roadblocks. For example, there are many bugs related to the pytorch framework, and also the modification of EfficientNet. There are also many challenges on making different dataset work together to make a larger data sample. Augmentation of the data is also problematic.

How We Used Space Agency Data in This Project
  • https://doi.pangaea.de/10.1594/PANGAEA.895835 Global Wildfire Database for GWIS. An individual fire event focused database. Post processing of MCD64A1 providing geometries of final fire perimeters including initial and final date and the corresponding daily active areas for each fire.
  • https://www.globalfiredata.org/fireatlas.html Global dataset on individual fire behavior reveals location of largest, longest and fastest wildfires
  • https://www.kaggle.com/carlosparadis/fires-from-space-australia-and-new-zeland Fires from Space: Australia, NASA Satellite Data MODISC6 and VIIRS 375m from 2019-08-01 to 2020-01-11
  • https://firms.modaps.eosdis.nasa.gov/active_fire/#firms-txt Active Fire Data
  • https://firms.modaps.eosdis.nasa.gov/download/list.php Archive Fire data (We submited request to download, we got 3 different zips folders there) - For Fire Detection
  • https://search.earthdata.nasa.gov/search/granules?p=C1426442798-LANCEMODIS&pg[0][gsk]=-start_date&m=-1.6881083994900195!-136.125!0!1!0!0%2C2&as[science_keywords][0]=Atmosphere%3AAtmospheric%20Water%20Vapor%3AWater%20Vapor%20Indicators%3AHumidity%3ASurface%20Humidity&tl=1585994203!4!!&fst0=Atmosphere&fsm0=Atmospheric%20Water%20Vapor&fs10=Water%20Vapor%20Indicators&fs20=Humidity&fs30=Surface%20 Humidity Precipitable water Vapor 5-Min L2 Swath 1km and 5km - NRT (NASA Earth Data) - For Fire Prediction
  • https://search.earthdata.nasa.gov/search/granules?p=C1239897978-GES_DISC&pg[0][gsk]=-start_date&q=surface%20temperature&m=5.0625!-6.46875!2!1!0!0%2C2&fdc=Goddard%20Earth%20Sciences%20Data%20and%20Information%20Services%20Center%20(GES%20DISC)!Soil%20Moisture%20Active%20Passive%20(SMAP)&tl=1585994203!4!! MODIS/Terra Monthly mean Day-Time Land Surface Temperature at 1x1 degree V005 (MOD11CM1D) at GES DISC (NASA Earth Data) - For Fire Prediction
  • https://search.earthdata.nasa.gov/search/granules?p=C1598621092-GES_DISC&pg[0][gsk]=-start_date&q=surface%20humidity&m=5.0625!-6.46875!2!1!0!0%2C2&fdc=Goddard%20Earth%20Sciences%20Data%20and%20Information%20Services%20Center%20(GES%20DISC)!Soil%20Moisture%20Active%20Passive%20(SMAP)&tl=1585994203!4!! GPM IMERG Final Precipitation L3 1 month 0.1 degree x 0.1 degree V06 (GPM_3IMERGM) at GES DISC - For Fire Prediction
Project Demo

https://sunflowercoconut.000webhostapp.com/

https://docs.google.com/presentation/d/1zQuKYl8S_1z9ACKiUOmcb_2eWMYVrTqxxqeDx0d7U0g/edit?usp=sharing

Project Code
Data & Resources

Refer to "How did you use space agency data in your project?"

Judging
This project was submitted for consideration during the Space Apps Judging process.