In this study,we extensively evaluated the viability of the state-of-the-art YOLOv8 architecture for object detec-tion tasks,specifically tailored for smoke and wildfire identification with a focus on agricultural and...In this study,we extensively evaluated the viability of the state-of-the-art YOLOv8 architecture for object detec-tion tasks,specifically tailored for smoke and wildfire identification with a focus on agricultural and environmen-tal safety.All available versions of YOLOv8 were initially fine-tuned on a domain-specific dataset that included a variety of scenarios,crucial for comprehensive agricultural monitoring.The‘large’version(YOLOv8l)was se-lected for further hyperparameter tuning based on its performance metrics.This model underwent a detailed hyperparameter optimization using the One Factor At a Time(OFAT)methodology,concentrating on key param-eters such as learning rate,batch size,weight decay,epochs,and optimizer.Insights from the OFAT study were used to define search spaces for a subsequent Random Search(RS).The final model derived from RS demon-strated significant improvements over the initial fine-tuned model,increasing overall precision by 1.39%,recall by 1.48%,F1-score by 1.44%,mAP@0.50 by 0.70%,and mAP@0.50:0.95 by 5.09%.We validated the enhanced model's efficacy on a diverse set of real-world images,reflecting various agricultural settings,to confirm its ro-bustness in detecting smoke and fire.These results underscore the model's reliability and effectiveness in scenar-ios critical to agricultural safety and environmental monitoring.This work,representing a significant advancement in the field of fire and smoke detection through machine learning,lays a strong foundation for fu-ture research and solutions aimed at safeguarding agricultural areas and natural environments.展开更多
文摘In this study,we extensively evaluated the viability of the state-of-the-art YOLOv8 architecture for object detec-tion tasks,specifically tailored for smoke and wildfire identification with a focus on agricultural and environmen-tal safety.All available versions of YOLOv8 were initially fine-tuned on a domain-specific dataset that included a variety of scenarios,crucial for comprehensive agricultural monitoring.The‘large’version(YOLOv8l)was se-lected for further hyperparameter tuning based on its performance metrics.This model underwent a detailed hyperparameter optimization using the One Factor At a Time(OFAT)methodology,concentrating on key param-eters such as learning rate,batch size,weight decay,epochs,and optimizer.Insights from the OFAT study were used to define search spaces for a subsequent Random Search(RS).The final model derived from RS demon-strated significant improvements over the initial fine-tuned model,increasing overall precision by 1.39%,recall by 1.48%,F1-score by 1.44%,mAP@0.50 by 0.70%,and mAP@0.50:0.95 by 5.09%.We validated the enhanced model's efficacy on a diverse set of real-world images,reflecting various agricultural settings,to confirm its ro-bustness in detecting smoke and fire.These results underscore the model's reliability and effectiveness in scenar-ios critical to agricultural safety and environmental monitoring.This work,representing a significant advancement in the field of fire and smoke detection through machine learning,lays a strong foundation for fu-ture research and solutions aimed at safeguarding agricultural areas and natural environments.
