Pyro Precise Prediction (PPP)

Pyro Precise Prediction (PPP)| Spot That Fire V3.0

Awards & Nominations

Pyro Precise Prediction (PPP) 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.

Summary

Forest fires spread fast! To deal with it efficiently, firefighting strategies must also consider predictive measures. Knowing that certain conditions are essential for fire propagation, and that human activity plays a role on the matter, our project aims at predicting which regions are on the verge of burning by feeding a machine learning classification model with features such as temperature, humidity, rain index, land cover and presence of roads. We believe the algorithm could assist well-established monitoring systems provided by INPE (Instituto Nacional de Pesquisas Espaciais) and also shine some light on criminal fires by finding out the profile of its most likely origins.

How We Addressed This Challenge

Experimenting through a case study of Pantanal (Brazilian ecosystem with some similarities with a Savanna), three facts became clear:

Nowadays Brazil uses only mathematical models to predict wildfire;
The current system do not have features related to human caused fire - although it is known that is an important cause of this fires around the globe;
There is a huge challenge in fire prediction and control not only in Brazil but in many parts of the world and those models could use some improvements.

The team believes that Brazilian reality is not a single one and designed a system, through machine learning, to predict fire with precision, helping responsible organs not only to prevent it, but also to understand the causes of it, also enabling tracking its origin. For that, NASA’s datasets are used combined with national datasets enhancing society power to deal with wildfires and protecting wild species.

How We Developed This Project

Forest fires are a major global concern. At an increasing rate - or so it seems -, we are bombarded with news of flames wiping out unrecoverable natural resources. In Brazil, especially, but not exclusively, we also face the heavy burden of political apathy and disinformation campaign, which often leaves us under the impression of powerlessness. On the other hand, advanced institutes and agencies constantly put a lot of effort in supplying the world with open data; we, researchers or not, should take advantage of that, either to spread awareness or to develop tools. Our team is motivated by both.

When the media covers deforestation in Brazil, it usually refers to the Amazon forest. Smaller biomes, such as Pantanal, are often unspoken about. In September, 2020, however, fires of unprecedented levels brought the attention to the region and sparkled global commotion. Seizing the momentum, Pantanal posed to us as a great opportunity for a case study: besides the increased current relevance, the biome comprises a region of reasonable proportions to tackle as a primary research environment.

Regarding forest fires, the biggest challenge is, perhaps, to prevent it from spreading. After a fire detection, even a fast response might be insufficient, especially in places which are hard to reach. In order to mitigate this difficulty, firefighting strategies must consider predictive solutions. Although fires might occur for unpredictable causes, certain conditions are essential for its propagation, such as heat and low humidity. Besides climate factors, human activity is also relevant to consider: the presence of roads and farmland in a region might imply a higher chance of a fire event. Given these conditions, our hypothesis is that we can predict which regions are on the verge of burning. With this knowledge in hand, firefighting teams and stations could develop a more efficient approach.

The idea of the technical approach came from the desire to apply a widespread technique in industry 4.0 in an innovative way in this scope of fire predictions.

The technique, which is fault detection (normally used to detect machine failures) can be perfectly applicable for fire detection, since there is a clear parallel between the data obtained (on the one hand, industry sensors on the other, satellite data) and the moments we want to predict (on the one hand, forest fire on the other, machine failure).

The algorithm will receive the features separated by longitude, latitude and date, also knowing the respective labels of "Fire", "Pre-Fire", "Normal", characterizing the moments. All of these data are available on both FIRMS and SENTYNEL and other satellites from NASA, ESA, CSA and JAXA.

Once an information pipeline has been established between these satellite data and a local machine, such data will be treated in Python with libs such as Pandas, Numpy, SK-learn, etc., enabling the algorithm to train with the most impactful features for fire detection.

The proposed algorithm would be a simple supervised classification algorithm using either Random Forest or NLP. Such simplicity is useful for its quick and easy application in an efficient manner.

To validate the project, we would use part of the data to perform a cross-validation, where we evaluate the performance of the algorithm, if it is correctly identifying the moment type, whether it is "Normal" or "Pre-Fire".

A positive point for the cost of implementation is that there would be no need to spend resources on new hardware or devices, the overall cost would just be the cost of a team of engineers to deploy such a solution in real time adapting the data flow.

The team's greatest difficulty in delivering the solution was during the part of data processing and exploratory analysis.

On the part of exploratory analysis, we came across the immense amount of data provided from all satellites. Too much time was spent during this hackathon to understand what each feature meant and which data provided by the various satellites would best fit the proposed solution. However, we believe that taking this solution further, this difficulty would become an extra perk to help the algorithm, being able to combine such data and further enrich the algorithm (a fact that we cannot achieve during a 2-day hackathon).

On the part of data processing, also due to the immense amount of data, we were unable to treat the data in the best way we had in mind, which would be selecting batches of time, analyzing statistical distributions, etc. But if this work goes ahead, such tasks would be assigned to a data scientist capable of performing such functions.

In the end, the great technical achievements of this project were mainly due to the fact that we found data potentially indicative of criminal fires that would fit the feeding of the Machine Learning model making it resilient and able to identify and predict fires of the most diverse types, both criminal and natural. (through the analysis of the flow of trucks, the proximity to farms that carry out deliberate fires). A fact that has been proving extremely important for Brazilian reality.

How We Used Space Agency Data in This Project

Classification models are trained with data provided by Sentinel-2, MODIS and VIIRS in the closest available timestamp prior to a fire detection, as informed by FIRMS. Other important tools worth mentioning, used mostly for inspection, are FIRMS Fire Map, Sentinel Playground and INPE (Instituto Nacional de Pesquisas Espaciais) dashboards.