Roohum Jegan, Gajanan K. Birajdar, Sangita Chaudhari
{"title":"Deep Residual Multi-resolution Features and Optimized Kernel ELM for Forest Fire Image Detection Using Imbalanced Database","authors":"Roohum Jegan, Gajanan K. Birajdar, Sangita Chaudhari","doi":"10.1007/s10694-025-01729-7","DOIUrl":null,"url":null,"abstract":"<div><p>The growing incidence of wildfires, intensified by changing climate patterns, poses risks to human lives and the environment, leading to catastrophic impacts on agricultural and forest ecosystems. Consequently, timely wildfire detection becomes imperative to implement effective mitigation strategies. This article presents a new forest fire image detection technique to address a class imbalance problem using ResNet-18 multi-resolution features and kernel extreme learning machine (KELM). Shallow and deep layer ResNet-18 features are extracted and fused to create a comprehensive feature set that represents local and global characterization of the forest fire image data. The multi-resolution feature fusion effectively captures lower-level visual patterns and complex and abstract representations of the input image. The fused feature set is subsequently input into a kernel extreme learning machine, which effectively handles nonlinear data patterns for binary classification tasks like fire detection. However, the performance of the KELM heavily relies on its hyperparameters, which are optimized using the Newton–Raphson-Based Optimizer (NRBO) algorithm. The hyperparameters fine-tuning process ensures that the KELM operates with optimal settings, ultimately enhancing the accuracy and reliability of the fire detection process. The proposed algorithm is evaluated using two publicly available databases, Forest Fire and Flame, with a detection accuracy of 97.88% and 99.88%, respectively. Moreover, the contribution of each feature to the model’s predictions to interpret decisions is elaborated using SHAP (SHapley Additive exPlanations).</p></div>","PeriodicalId":558,"journal":{"name":"Fire Technology","volume":"61 5","pages":"3323 - 3349"},"PeriodicalIF":2.4000,"publicationDate":"2025-03-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Fire Technology","FirstCategoryId":"5","ListUrlMain":"https://link.springer.com/article/10.1007/s10694-025-01729-7","RegionNum":3,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENGINEERING, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
Abstract
The growing incidence of wildfires, intensified by changing climate patterns, poses risks to human lives and the environment, leading to catastrophic impacts on agricultural and forest ecosystems. Consequently, timely wildfire detection becomes imperative to implement effective mitigation strategies. This article presents a new forest fire image detection technique to address a class imbalance problem using ResNet-18 multi-resolution features and kernel extreme learning machine (KELM). Shallow and deep layer ResNet-18 features are extracted and fused to create a comprehensive feature set that represents local and global characterization of the forest fire image data. The multi-resolution feature fusion effectively captures lower-level visual patterns and complex and abstract representations of the input image. The fused feature set is subsequently input into a kernel extreme learning machine, which effectively handles nonlinear data patterns for binary classification tasks like fire detection. However, the performance of the KELM heavily relies on its hyperparameters, which are optimized using the Newton–Raphson-Based Optimizer (NRBO) algorithm. The hyperparameters fine-tuning process ensures that the KELM operates with optimal settings, ultimately enhancing the accuracy and reliability of the fire detection process. The proposed algorithm is evaluated using two publicly available databases, Forest Fire and Flame, with a detection accuracy of 97.88% and 99.88%, respectively. Moreover, the contribution of each feature to the model’s predictions to interpret decisions is elaborated using SHAP (SHapley Additive exPlanations).
期刊介绍:
Fire Technology publishes original contributions, both theoretical and empirical, that contribute to the solution of problems in fire safety science and engineering. It is the leading journal in the field, publishing applied research dealing with the full range of actual and potential fire hazards facing humans and the environment. It covers the entire domain of fire safety science and engineering problems relevant in industrial, operational, cultural, and environmental applications, including modeling, testing, detection, suppression, human behavior, wildfires, structures, and risk analysis.
The aim of Fire Technology is to push forward the frontiers of knowledge and technology by encouraging interdisciplinary communication of significant technical developments in fire protection and subjects of scientific interest to the fire protection community at large.
It is published in conjunction with the National Fire Protection Association (NFPA) and the Society of Fire Protection Engineers (SFPE). The mission of NFPA is to help save lives and reduce loss with information, knowledge, and passion. The mission of SFPE is advancing the science and practice of fire protection engineering internationally.