Forest fires are a great cause of destruction in the wild, often leading to a loss of flora and fauna, and forcing the displacement of local population. It was during a casual chat with an old friend, who was closely involved with the Melghat Tiger Reserve in Maharashtra, that Dr Srivastava learned how these fires are detected. The task is fully reliant on manual effort by forest officials and wardens who go for long walks and, in the absence of any connectivity, run back to inform of a mishap, taking as long as two hours at times. Moreover, Melghat’s rugged terrain makes fighting forest fires a herculean task.
“So we just talked and I got this simple idea that has been followed in large parts of the world. Why don’t we have some sort of a Wireless Sensor Network deployment? These are motes that have sensors, and then they form an ad hoc network that can sense a fire,” says Dr Srivastava during a conversation with INDIAai. Dr Abhishek Srivastava is an Associate Professor of Computer Science and Engineering at the Indian Institute of Technology Indore.
Recently, wireless sensor network (WSN) has become a promising technology with a wide range of applications, including for environment surveillance. It is typically composed of multiple tiny devices equipped with limited sensing, computing and wireless communication capabilities. These tiny devices are called motes or wireless sensor nodes. A single mote has mainly three key ingredients- a microcontroller, sensors and low power radios.
The pandemic hit just a few months into the project, so Dr Srivastava’s team did a pilot deployment in IIT Indore. The biggest problem that they noticed was that of false positives, however. “Especially on a very hot summer day, the heat sensors will get triggered and it’s going to tell them that there is a fire. You have a bunch of people running even when there’s no fire” he says.
In addition, sending back false signals drains the battery that powers these motes. “The problem with these motes is that they’re extremely resource constrained. The battery is limited and is going to get exhausted very soon and the purpose is going to be defeated.”
And that’s where AI and ML came in. “We thought, why don’t we have these really simple algorithms? We have so many sensors they’ll send back signals. And only if there is a fire, the algorithm determines that there is indeed a fire. It’s not a false positive.”
So, the team decided to have AI algorithms deployed on the mote. These algorithms, however, had to be modified to work on the motes because most algorithms these days are developed with the assumption of a having resource liberal environment, facilitated by the infinite cloud storage. “That is totally nontrivial because you’re talking about something as small as 2 KB of RAM. So you cannot store any data.”
Explaining the process, he adds, “We’ve developed an anomaly detection algorithm and a classification algorithm. The anomaly detection tries to detect an anomaly and normally if there is one, it indicates the fire. And the classification algorithm classifies false positive and the real cases. So, in some ways, that part about eliminating false positives was taken care of.”
When designing for an area as expansive as 15,000 sq km, the other major problem is that of localisation. “So based on the algorithm, you know that there is a fire as soon as a signal goes back. But how do you locate where the fire is?” The problem is exacerbated with the inability to use GPS effectively in the thickly forested area. Also, the GPS module is very heavy to be attached to the modes.
Once the team started using localisation algorithms for multilateration (which is the technology that the GPS uses to locate objects) they soon ran into its limitations as well. “If there is a mode that is in the range of at least three other modes, we do triangulation to find this node. If you do that for one or two iterations, it’s going to be fine. But every time that you localise nodes, there will be some error. When you do it for hundreds of thousands of nodes, the error is going to keep increasing and ultimately spiral out of control. That’s when we started using a combination of multilateration and machine learning.”
They started training simple algorithms for localisation, offline, and tried to map both the RSSI (received signal strength indicator) and location using machine learning. “The first iteration of localisation was done using machine learning and a large number of modes were localised. Once we had a large amount of covered, we started using the triangulation technique.”
The project is still in the pilot stage, but the team plans to visit Melghat as soon as the lockdowns ease and deploy it on location. The actual experience will be more challenging than the pilot that is being carried out at the university campus, admits Dr Srivastava. “The thing with motes is that you have no way of going in and fixing it – if it’s down its down. Also, you can’t have two or three motes for one place because they are expensive and we’re talking about hundreds of thousands of motes.”
“Even if 70-80% of the fires are detected, I think it will be a success because scaling it up will probably be the biggest challenge,” says Dr Srivastava.
Source: indiaai.gov.in
