YOLO based stubble burning detection system for Northern regions of India

Stubble burning (SB) is a pervasive practice in Northern India, contributing significantly to winter air pollution, particularly exacerbated by the widespread use of combined harvesters. Despite its illegality, SB continues, releasing harmful pollutants and significantly deteriorating air quality. Recent advancements in computer vision, particularly those utilizing neural networks like YOLO, have significantly improved object detection capabilities. This paper proposes a novel approach to address this pressing issue by developing a cost-effective neural network-based fire and smoke detection model for an automated SB detection system. The proposed model is built upon YOLOv5 to efficiently detect the stubble burning: Architectural modifications of basic YOLOv5 are done by updating the Neck, Backbone and Head enhancing the feature extraction and object localization. We assess the model’s performance on our custom dataset using F-score, Mean Average Precision (mAP), and accuracy. Additionally, we perform an ablation study to examine the effects on inference time and mAP, as well as to evaluate the relationship between modal weights and processing speed.