“In addition to cameras and sensors, artificial intelligence (AI) is another technology that merits rapid adoption for fire prevention, detection, and suppression.” – Independent Institute1
Fast & Accurate Dual-Sensor
Our LWIR detection is fast (as small as 2 m2 flame at 5 km).
Accurately detect fire occurrence by dual-detection (visual & infrared)
Our patented technology accurately determines fire location.
Work reliably in tough environments (dark, light polluted, hazy, windy & hilly).
Automatic & Efficient AI Detection
Always-on Automatic AI detection. No delay nor tiredness.
Self-improving machine learning system analyses raw images on the edge
Large coverage (70k Ha per robot) and minimal manual operation.
More manpower-efficient, energy-efficient, and carbon-friendly than watchtowers and aerial fire patrol.
Intuitive & Versatile Management Platform
Easy-to-use GUI with 2D/3D maps. Supports cross platforms.
Situational awareness with real-time video & panoramic image.
Support visualization of 3rd party data via standard API.
Timestamped GIS-referenced fire records for post-event forensic and risk mitigation planning.
Figure 1: InsightFD Key Customer Benefits
Why AI?
What is wildfire smoke?
Smoke is made up of particulates, aerosols and gases. It is the by-product of the fuels it is burning. White smoke is off-gassing moisture and water vapor, meaning the fire is just starting to consume material. Thick, black smoke indicates heavy fuels that are not being fully consumed. In general, the darker the smoke, the more volatile the fire is.2
In the beginning, our InsightFD offered thermal detection and visual detection of smoke. Thermal detection detects the infrared emitted by burning trees and plants, while visual detection identifies the smoke of vaporized vegetation tissues.
Our visual detection is an expert system that detects smoke based on rules. These rules are defined by our experienced computer vision engineers. The rules describe the wildfire smoke: color, shape, size, motion, changes and change rates, etc.
Rule-based systems are deterministic in nature. Not having the right rule in place can result in false positives and false negatives. As we support more customers in more countries, we will update our rules to cover new scenarios. It takes time to learn a new environment and to become an expert.
Wildfires in different countries are different. The background cloud movement, cloud shadows, chimney smoke, tree vibrations, and sunlight reflection of bright surfaces, etc.
We need a more efficient way to optimize our detection for a new deployment than updating hundreds of rules. We believe we can make more effective use of our experience and knowledge.
What is our AI smoke detection?
We use Machine Learning (ML) to build our next-generation AI smoke detection. ML is probabilistic and uses statistical models rather than deterministic rules.3
We use YOLOv3 (You Only Look Once) to develop our real-time smoke detection system. We select YOLO because it is extremely fast and accurate compared to other ML-based detections. At the end of the day, from the perspective of early wildfire detection, a 99.9% confidence level of wildfire occurrence is only useful if you can detect it the second you observe it!
YOLO is great for forest fire detection because it looks at the whole image at test time so its predictions are informed by the global context in the image. This helps us to avoid false-positive cases such as chimney smoke.
How we build our AI smoke detection?
The power of Machine Learning is hidden in the self-learning algorithms, which when exposed to a huge amount of data, can study and learn for improved results.4 We can leverage our library of hundreds of thousands of HD images of early-stage wildfire and smoke to train our ML algorithm. With the support from our global customers around the world, we collected more than half a million of forest fire images of tropical, temperate and boreal forests in the past nine years. Pictures in both daytime and night-time, sunny and dry, cloudy and windy. All these are the foundations to build our CNN (convolution neural network) model which can work in Asia, Australia, Europe, North / Central / South Americas!
Our AI engineering team spent thousands of hours training our YOLOv3 model and configuring the hyperparameters – learning rate schedules, types of optimizer, loss weights, data augmentation method, numbers and types of hidden layers, weight initialization schemes, etc – to optimize the detection speed and accuracy.
Customer participation is the last puzzle to complete our AI smoke detection development! Since late 2018, we worked with some selected customers to deploy our beta ML algorithm on their InsightFD systems. We compared the detection performance against our thermal detection, rules-based visual detection, and new AI smoke detection. We base on the results to train our CNN model and improve it.
Our AI smoke detection is now outperforming our rule-based smoke detection! We are impressed by the AI smoke detection performance and its improvement rate! We update our rule-based visual smoke detection on a yearly basis. We can now update our AI smoke detection in months! It is an order of magnitude improvement!
In July of 2020, we made our AI smoke detection available to all our customers!
Our goal is to bring our AI smoke detection technology to more customers, to detect wildfire earlier, and to better protect our forests and communities!
Please download our InsightFD brochure if you want to learn more about our product.
Reference:
Black-faced Spoonbills: I see you/> 
4 Ways Camera-Based AI SaaS is Revolutionizing Wildfire Detection/> 