With stubble burning being a pervasive practice in northern India, researchers have proposed a novel approach to address this pressing issue by developing a cost-effective neural network-based automated fire and smoke detection framework.
Researchers at the Thapar Institute of Engineering and Technology, Patiala, used a system based on the You Only Look Once (YOLO) platform that has significantly improved object detection capabilities as compared to other means. YOLO’s architecture was modified to enhance feature extraction and object localisation.
“The experimental findings indicate that the proposed approach outperforms existing state-of-the-art methods across multiple evaluation criteria,” the researchers said in their study that was published by Scientific Reports, a peer reviewed journal, on April 12.
In northern Indian states such as Punjab, Haryana, Uttar Pradesh and parts of Uttarakhand, around four million hectares of land are dedicated to rice and wheat cultivation.
After harvesting paddy in mid-October, farmers need to plant wheat by mid-November to ensure a successful harvest by mid-April.
Since paddy residues have limited economic value, farmers often burn the leftover stalks to prepare the fields for wheat. This practice significantly exacerbates winter pollution, the study observed. Further, stubble burning deteriorates air quality, depletes soil nutrients, increases the need for fertilisers and negatively impacts soil microbes and fauna, besides damaging electrical equipment and causing uncontrolled fires.
According to the researchers, various detection methods such as thermal remote sensing, temporal remote sensing and satellite-based remote sensing are being used to help identify burnt paddy fields and measure atmospheric pollution.
While such approaches provide large scale coverage, they often suffer from coarse spatial resolution, cloud interference, delayed detection and limited responsiveness to short-duration burning events.
In densely populated areas, the smog complicates the monitoring further. All these limitations restrict their suitability for real-time enforcement and rapid intervention.
“To address the challenges, recent studies have explored computer vision and deep learning frameworks using the UAV and the ground based imagery. One stage detectors such as the YOLO family have gained particular attention due to their real-time inference speed and suitability for the edge deployment. These models enable localised fire and smoke detection with higher precision,” the researchers said.
In practical deployments, aerial and edge systems often transmit compressed imagery due to bandwidth and storage constraints which significantly degrade detection performance. Hybrid feature extraction and multi-branch architectures have been shown to improve resilience of YOLO-based detectors under such distortions, they added.
Tests conducted by the three-member team showed that the system accomplished its primary objective of detecting stubble burning in varying weather conditions and demonstrated versatility in detecting fires under diverse lighting conditions and at any time of the day.
Enhancements in various system components resulted in notable reductions in overall response time, facilitating real-time tracking of burnings, while hardware upgrades utilising high-end hardware enabled the integration of additional sensors, addressing existing limitations and enhancing overall accuracy, the study said.
