LiFE-BA-Net: A Lightweight Network With Depth Feature Enhancement and Boundary-Aware in Fire-Front Segmentation
Date:2026-10-26 clicks:
Affiliation of Author(s):东北林业大学
Journal:IEEE 期刊
Place of Publication:美国 (IEEE总部)
Key Words:Deep learning; fire front; model optimization; semantic segmentation; Wildfire
Abstract:Wildfires present a significant danger to forest ecosystems, underscoring the necessity for accurate and efficient monitoring systems that can track fire behavior beyond the initial ignition phase. Although recent advances in computer vision and deep learning have improved early-stage fire detection, current models hardly address the complexities of post-ignition fire-front evolution, primarily due to insufficient datasets. Existing approaches are also hindered by elevated computational cost, constrained accuracy, and inadequate real-time performance, limiting their utility in the swiftly changing wildfire contexts. To overcome these restrictions, we present a novel drone-based dataset on wildfire propagation, derived from multi-platform aerial imagery, for comprehensive investigation of fire-front dynamics. Leveraging this dataset, we introduce LiFE-BA-Net, a lightweight encoder-decoder architecture designed for real-time fire-front segmentation. LiFE-BA-Net utilizes a U-Net architecture with MobileViT2 as the backbone for feature extraction. It employs depthwise separable convolutions for downsampling and bilinear interpolation for upsampling, reducing model complexity. To enhance segmentation quality, the network incorporates a deep feature enhancement module in the encoder to preserve semantic information and a residual Sobel-Laplace (SL) module in the decoder for refining fire-front boundaries. Results show that LiFE-BA-Net significantly outperforms existing models with a precision of 86.13%, showing improvements of 2.16% to 8.33%. It achieves an intersection-over-union (IoU) of 81.98%, with enhancements of 2.72% to 9.1%, utilizing only 1.68 million parameters. Therefore, LiFE-BA-Net provides an effective and efficient solution for real-time fire-front monitoring, with strong potential for operational wildfire tracking and forest resource management in dynamic environments.
Co-author:Jungho Im,Guang Yang,Yinghui Zhao,Fengjun Zhao
First Author:Feng Zhu
Indexed by:Journal paper
Correspondence Author:Zhen Zhen*
Translation or Not:no
Date of Publication:2026-01-01
Included Journals:EI