Decoder-Based Anomaly Detection With Concentration Filtering for Distinguishing Wildfire Smoke From Clouds

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:For researchers in anomaly detection and remote sensing, the concentration filtering module offers a novel approach.

🏛政策担当者:Policymakers in disaster management may consider this for early warning systems.

Wildfires are natural disasters that cause extensive human casualties and economic losses worldwide. Rapid detection in the early stages is essential to minimize damage; however, conventional supervised learning-based object detection methods face limitations due to the difficulty of acquiring sufficient labeled data for rare wildfire events and their high implementation costs. This study proposes an unsupervised wildfire-detection framework that leverages the characteristics of fixed closed-circuit television (CCTV) environments. The core of the proposed method is a Learnable Concentration Filtering Module that quantifies the spatial distribution differences between wildfire smoke and atmospheric clouds. To extract robust feature representations, we employ ConvNeXt, a modernized convolutional architecture that integrates Vision Transformer design principles to achieve superior performance. Multi-scale features are extracted using a this backbone, and each feature map is divided into fixed $14\times 14$ patches. Spatial correlations between patches are learned through local-window self-attention. From the decoder-based reconstruction error map, the concentration ratio defined as the ratio of local maximum to mean values are computed at multiple scales, and concentrated anomaly patterns are selectively extracted through learnable thresholds. Experimental results on four regions of the HPWREN dataset achieved an average area under the receiver operating characteristic curve (AUROC) of 0.801, average precision (AP) of 0.689, and maximum F1-score (F1-max) of 0.769, demonstrating improved performance over other state-of-the-art anomaly detection models. The proposed method achieved a fast inference time of 10.53 ms, confirming its real-time processing capability. This study presents a practical approach for developing wildfire-detection systems that are robust to environmental variations and do not require labeled data.

• semanticscholar https://doi.org/10.1109/access.2026.3654564first seen 2026-06-29 08:08:41