Context-Aware Satellite Remote Sensing Fire Point Detection Based on Energy Scores

Abstract

Mobile deployable deep models are crucial for forest fire point detection based on satellite remote sensing images. Existing convolutional neural networks (CNNs) are limited by their context-aware capabilities and the Transformer requires quadratic computational complexity for modeling long-distance dependency relationships, making it difficult to effectively deploy the model on mobile devices. To this end, this article constructs a context-aware Mamba network based on energy-based distillation for satellite remote sensing fire point detection. Firstly, we construct a feature extraction backbone network based on the Mamba module, which can achieve long-distance dependence modeling with linear computational complexity. In addition, we introduce a distillation learning mechanism based on energy score to improve the forest fire recognition performance. The results of the publicly available satellite remote sensing fire dataset have confirmed that our proposed method achieves the highest F1-Score in fire detection tasks.