Forest Fire detection
Project overview
During the NASA Space Challenge hackathon in 2019, our team "Code & Klystron" developed a Proof of Concept pipeline called "Forest Fire Detection" in just 2 days. This model uses real-time satellite data to detect early-stage forest fires and send alarm messages via Telegram about the fire and its location. Our system aims to prevent large forest fires by detecting and eliminating fires at an early stage.
Business Value
Forest fires cause enormous environmental, ecological, and financial damages. For example, the forest fires in California in 2018 resulted in $400 billion in losses, and the 2019-2020 Australian bushfires killed over 3 billion animals. However, these losses could be prevented if the forest fires were detected and eliminated before they grew to an enormous size. The Forest Fire Detection pipeline provides a service that notifies of possible early-stage forest fires by sending notifications with coordinates, allowing quick action to be taken to prevent the fire from growing.
Technical details
The model core is based on computer vision and digital signal processing. We used freely accessible micro-satellites that provide a gas map updated every hour. The analysis of CO2 emission changes over time allows us to determine possible fires as a new small source of gas emission. Our algorithm is based on finding small circles of gas emission that rapidly appear. We achieved this with computer vision applied together with frequency analysis.
Algorithm is based on finding small circles of gas emission that rapidly appeared. We achieved that with Computer Vision applied together with frequency analysis.
Challenges
Filtering the map. Detecting a source in the map is a difficult task as it already has many gas sources, such as cities and factories. Additionally, the wind causes gases to flow anywhere, sometimes creating a shape similar to the source shape. Therefore, these two phenomena must be filtered before applying analysis to the gas map. a. Solution. We implemented a band-frequency digital filter. We manually calculated the lowest frequency at which static sources make pulses of higher or lower CO2 emission, then passed the map through a low-frequency filter to remove all static sources. We also measured the optimal highest frequency that cuts off most of the wind flows and passed the map through a high-frequency filter to remove all gas flows. By combining both filters, we get a band-width digital filter. We used the Butterworth filter as it has smooth amplitude decay, minimizing data distortion.
Detecting new sources. Even with the filtered data, we still face the problem of finding new sources. We applied Fourier Transform and worked in the frequency domain. We expressed each pixel of the map as a time signal over the last M frames available, applied frequency analysis to it, and found a place where the CO2 emission value rose dramatically. We then created time tubes and applied Fourier Transform independently to them to find a possible new forest fire. Finally, we applied image processing to verify whether an image has a big enough circle, meaning there is a sufficient source of CO2 emission.
Results
Our project won 3rd place in the NASA Space Challenge hackathon in 2019. We built a working prototype of the model that loads the latest up-to-date gas map from satellites, applies digital signal processing and image analysis techniques, and notifies users via messenger if a possible forest fire is detected.
Future Works
The pipeline has significant potential for improvement, both in terms of accuracy and usability. One possible future development would be to integrate the pipeline with a predictive model that uses weather data to forecast fire risk. Additionally, expanding the pipeline to incorporate other types of satellite data, such as thermal imagery, could improve detection capabilities. Another avenue for improvement would be to develop a mobile application that would allow users to receive real-time alerts about detected fires and access information about current fire conditions in their area. These enhancements would make the Forest Fire Detection pipeline more effective and accessible to a wider audience.
