EDIF: boosting unsupervised cross-domain forest fire smoke detection with enhanced domain-invariant features

In forest fire smoke detection tasks, variations in data distribution caused by deployment environments, acquisition devices, and smoke characteristics, along with the scarcity of fire incidents, make it difficult for models to generalize across all potential scenarios. Unsupervised domain adaptation, as an effective paradigm that enhances model generalization without additional annotation costs, has therefore shown great potential. This paper proposes an unsupervised domain adaptation (UDA) method with Enhanced Domain-Invariant Features (EDIF) to improve cross-domain object detection by jointly enhancing image- and instance-level features. At the image level, domain-invariant features are refined by explicitly modeling their interdependencies with spatial context to strengthen feature expressiveness. At the instance level, key attribute correlations are integrated to reinforce domain invariance. A dynamically adjusted gradient reversal layer mitigates over-/under-learning during adaptation. Evaluated on a public cross-domain dataset, EDIF outperforms existing methods, achieving 8.8% and 13.6% accuracy gains in synthetic-to-real and real-to-synthetic smoke adaptation, respectively. These results validate its effectiveness in addressing domain shifts caused by environmental variations and limited labeled data in forest fire smoke detection scenarios.