From AI Ultra Early Fire Detection to €10.5m of Funding

Author of an article and illustrations: Weronika Dorocka

Introduction

We partnered with the Dryad Team to create an early wildfire detection system with AI.

Since the project ended, thanks to the technology we developed together they have already been able to prevent a great amount of fires, implement it in new locations and recently closed €10.5m in Series A Funding to bring ‘Ultra-Early’ Wildfire Detection to the global market.

The round was led by eCAPITAL and additional investors, including Toba Capital, Semtech, and Marc Benioff’s TIME Ventures. 

Find out how we did it!

The Problem

Between 60,000 and 80,000 forest fires occur yearly, destroying between 3 and 10 million hectares of forested area.

Experts suggest the frequency of extreme fires will increase by further 14 percent by the end of this decade and by around 30 percent by 2050.

Let’s put this problem into perspective:

Since the start of the year to July 27, 2023 Italy has lost over 50,000 hectares of vegetation to wildfires. That’s equivalent to 95,000 soccer fields. 

Wildfires are a global problem that impacts everyone

When wildfires occur:

1. The release of high amounts of CO2 (adding to greenhouse effect) which is a huge impact on Global warming
2. Loss of forest cover (reducing the amount of CO2 absorbed by trees) and killion  great amounts of animals.
3. Loss of lives, money resources in huge amounts *Feel free to check more about the numbers behind on the page of Our World in Data here.
   

All of these factors make climate change even worse, hence creating conditions in which more fires occur with greater intensity.

The Background

With the increasing frequency and intensity of wildfires globally, there is a pressing need for innovative solutions to mitigate their devastating impact. Traditional methods of fire detection rely heavily on manual observation and periodic patrols, which can be time-consuming and prone to human error. 

However, advancements in technology, particularly artificial intelligence (AI), have revolutionized the field by offering more efficient and reliable detection methods. AI-powered systems, equipped with sophisticated algorithms, can analyze vast amounts of data from various sources, including satellite imagery, weather forecasts, and especially environmental sensors, to identify potential fire hotspots with unprecedented accuracy. 

To drastically reduce the effects of climate change, we need to be leveraging cutting-edge AI Artificial Intelligence technology in fire detection systems. It becomes increasingly essential for enhancing resilience, protecting critical infrastructure and wildlife too.

The Goal

The plan was to build an intelligent model that detects fire in different types of wood through analysis of the existing sensor data, optimize the model to work on the sensor, and build an end-to-end training pipeline, leveraging the dataset provided by the Dryad team.

We wanted to simply create a system better, and faster than other solutions already available out there, as a lot of them act too slow to detect the fire at its start.

The Partner

A German Startup Dryad was the partner of this project. The base of what they are doing is to use a large-scale IoT network of sensors to detect wildfires and to provide valuable forest insights and analytics.

Their main mission is to greatly reduce the reaction time for tackling forest fires. They also aim to enable our customers to monitor the vitality and growth of their forests and other ecosystems, helping to protect and restore our vital natural resources. 

Dryad uses sensors to detect particulate matter (smoke) in the air released by fire.

Solar power Gas sensors that detect fires within minutes 

They detect the pattern of the gas that is typical for a fire.

Our Approach to this Project

It was an Innovation Challenge type of project of ours where we gathered a group of 50 engineers to solve an innovative problem. It is one of our market advantages,as we can test different concepts at the same time, in a short time frame.We worked on the data that the sensors provide to make it intelligently detect fires and create the most effective and informative warning system, even knowing what type of material is being burned

Technical Backbone:

The backbone of the whole system of our partner and how the whole system world is around using IOT mesh Networks*

* Internet of Things Mesh Networks In an IoT mesh network, devices can communicate with each other directly or through neighboring devices, forming a decentralized network topology.

Benefits of utilizing IOT Network Technology:

• Reducing the risk of single points of failure.
• Enhanced coverage area and scalability, as devices can relay information through neighboring devices.
• Self-healing capabilities, where the network can dynamically reroute communication paths in case of failures.
• Improved reliability and flexibility in complex environments.

Step 1: Getting to know the Data available

1. Data Analysis: The process began with analyzing the collected data to understand its structure and identify patterns or trends. This step helped in determining which features were relevant to the problem being addressed. This data was then classified randomly to increase the accuracy of the model and minimize data leakage (data leakage is when data from outside the training dataset influences the model’s decision making).
2. Data Preprocessing: After analyzing the data, it was preprocessed to clean, transform, and prepare it for training. Tasks such as handling missing values, scaling numerical features, and encoding categorical variables were performed during this phase.
3. Feature engineering: To make the model more efficient, the Team Decided on the Features that were most important to detect the fires. Knowing this information allowed us to give higher priority to features that are more reliable and lower priority to ones that are less reliable.

Model Choosing and Training: Once the data was prepared, the machine learning model was trained using the labeled examples in the training dataset. During training, the model adjusted its parameters to minimize the difference between its predictions and the actual labels.

When it came to choosing and training a model for our AI system, we often try different ones with different teams simultaneously to see which ones give the best efficiency. There is no one-size-fits-all approach; it was all about testing and experimenting. 

The problem was initially considered a binary classification one because the outcome was known to fall into one of two classes: normal or in-smoke. This classification made it a binary problem, where the goal was to teach the computer to distinguish between these two classes based on labeled examples. However we never take things for granted, so we also tested other options to see if we can have an even more accurate system in use utilizing other models.

Step 3. Improving Decision-Making Speed of our System

Leveraging Model Quantisation to perform Big Tasks on Small Machines 

The Data is being collected on small devices called Edge Device. They don’t have the time or processing power to transmit  it to the cloud, therefore we needed to use clever machine learning techniques to reduce the amount of data they needed to understand what was going on. 

In this process called Model Quantisation, we changed the numbers in the model from very precise (like using lots of decimal points) to less precise (using fewer decimal points). It can be compared to rounding numbers to make them simpler, we did something similar with the numbers in the model. This makes the model take up less space and run faster, while still doing a good job at its task.

Step 4: Putting it all together

When we set out to bring our machine learning models to life on devices, we approached it step by step to make sure everything worked smoothly. 

First, we put our models and related software into neat, standardized containers using Dockerization. This made it easy to make them run consistently on different devices, like smartphones or IoT gadgets. 

Then, we made the whole process simpler using MLOps practices, which helped us automate tasks and get everything working together smoothly. 

With tools like MLflow, we could keep track of our experiments and see how well our models were doing in real-time. And with DVC, we made sure that all our data and projects were carefully organized and tracked. 

All these practices in general are under the name of Deployment, and in simple terms, it means getting our models up and running on these devices, so they can do their jobs.

Outcomes and The Impact

The system developed in this project is scalable and replicable. We were able to create an effective system to detect the fires and specific types of woods.

The next steps will be able to detect even faster and even more accurately.

They were able to detect in a very fast manner is the first unplanned wildfire detected by our Silvanet sensor system and it was detected ultra-fast. Since we caught it in its smoldering phase, it was an easy task for the fire services to put it out and stop it from spreading.

The more countries, municipalities, companies decide to help and incorporate this innovative system the more lives we can save, the more we gan act towards helping our planet against climate change.

Benefits and other Applications of this Technology

Applied Methodology 1:

IoT Technology for Ultra Early Detection

Can also be utilized for:

• Manufacturing: In factories, sensors can enhance equipment monitoring, potentially preventing costly breakdowns. AI analysis of sensor data offers the possibility of proactive maintenance, reducing downtime and boosting productivity.
• Healthcare: Wearable sensors and remote monitoring systems provide continuous health data collection. AI analysis holds the potential to identify health issues early, leading to better patient outcomes and potentially reducing healthcare costs.
• Smart Buildings: Installing IoT sensors in buildings enables efficient monitoring of environmental conditions. AI analysis offers the possibility of early hazard detection, minimizing property damage and ensuring occupant safety, thus safeguarding business assets.
• Transportation and Logistics: Equipping vehicles with IoT sensors allows for real-time monitoring and predictive maintenance. AI analysis holds the potential to optimize fleet operations, reducing maintenance costs and improving delivery reliability, thereby enhancing customer satisfaction.
• Energy Management: Implementing IoT sensors in energy infrastructure offers the potential for optimized resource usage. AI analysis could lead to reduced energy consumption and lower operational costs, contributing to improved profitability.
• Environmental Monitoring: Deploying IoT sensors for environmental monitoring supports regulatory compliance and sustainability efforts. AI analysis offers the possibility of early hazard detection, minimizing environmental risks and potential financial liabilities.
• Retail and Supply Chain: Integrating IoT devices enables efficient inventory management and product monitoring. AI analysis holds the potential to reduce stockouts, prevent product damage, and streamline supply chain operations, leading to increased sales and customer satisfaction.

Applied Methodology 2:

Minimizing the Size of the Data – Model Quantisation 

Can also be utilized for:

• Mobile Applications: Optimizing AI models through model quantization ensures faster app loading times and lower memory usage, enhancing user experience.
• Healthcare: Model quantization speeds up medical data processing for faster patient care and enhanced diagnostic accuracy.
• Finance: Utilizing model quantization improves the speed and accuracy of financial data analysis while reducing computational costs in fraud detection and risk assessment.
• Manufacturing: AI models optimized with model quantization enhance operational efficiency, reduce downtime, and improve product quality in predictive maintenance and quality control.
• Retail: Model quantization enhances marketing strategies and customer satisfaction through improved segmentation, forecasting, and recommendation systems.
• Autonomous Vehicles: Smaller AI models with model quantization let self-driving cars make decisions fast without needing a lot of computer power, making them safer and more reliable.
• Telecommunications: Model quantization makes phone and internet networks work better, reducing disruptions and improving customer service.
  

Time Frame

The whole Project took only 8 weeks knowing we needed to:

• Understand and Make the Data Usable
• Test and Choose the best performing system of work 
• Design and Implement several complicated data-based pipelines 
• Work with the devices that have a challenging parameters such as: small processing power, need to save the energy (the sensors that they are using that are installed on the trees are fully solar powered)

Further possibilities

Getting Variety of Data for More Accuracy

Adding more types of sensors can give Dryad’s system a clearer picture of fire risks. Sensors for things like humidity and wind speed help spot danger signs more accurately.

Getting Even More Precise

We can improve fire detection systems by getting better at recognizing the types of material that are burning and what’s causing the fire. This way, emergency teams can react more accurately and quickly, helping to prevent wildfires from spreading and causing damage.

Seeing Fires Even Faster

By using predictive models, Dryad’s system can predict how a fire might spread based on past data and what’s happening right now. 

Making Sure it Works Everywhere

Dryad’s system needs to be able to work in different places and situations. Making it more flexible and easy to set up means it can protect more areas from wildfires.

Want to work with us too?

Let’s see if we are a good fit