.png){kind=icon}
.png){kind=icon}
Return
Interpretable deep one-class model for forest fire detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127657.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Forest fires are notorious for their unpredictable nature, resulting in extensive ecological damage and loss of human life. To prevent a fire, deep binary classification models have widely applied in a vision-based forest fire surveillance system, due to the powerful performance in pattern representation and feature extraction. Existing methods typically prioritize training set construction, but neglecting the model's interpretability in fire detection. Furthermore, in order to meet the large-scale sample requirement for training a deep model, users have to collect data from a diverse range of fire and non-fire scenes. However, this may result in a loose decision for fire identification, ultimately leading to a failure to detect forest fires early or at a distance. In this work, we propose an interpretable deep one-class classification (OCC) model, named InDeepOCC (Interpretable Deep OCC), for forest fire detection. Concretely, we incorporate the spirit of data distribution, e.g., the unimodality of fire samples, into model-design; Then our model is trained on one-class fire samples and some points self-generated from the neighborhoods of border fire samples, aiming to obtain a tight decision boundary. Compared to state-ofthe-art (SOTA) methods, our advantages lie in three-fold: (1) The proposed InDeepOCC model provides superior interpretability; (2) It is easier to use because only a single class of samples is required; (3) It has more potential in small-area fire flame detection. Finally, we also provide an extensive comparison on public and our collected fire data. The result shows that our method outperforms SOTA, achieving a higher fire detection rate and a lower error warning rate simultaneously, and the ability in fine-grained fire detection, even at a granularity of 4 x 4pixel inputs.
Keywords:
Forest fire detection
Error warning
One-class classification
Granularity
Deep learning
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W