Remote Sensing Image Dehazing via Progressive Frequency Filtering and Adaptive Self-Attention
Lei Xu, Yijia Xu, Xuejian Gong, Yipeng Mao · IEEE Access · 2025
To meet the requirements of forest fire detection, a remote sensing (RS) image dehazing algorithm has been proposed, which always addresses the challenges posed by haze and smoke in real-world forest fire detection tasks. Moreover, RS image dehazing is a very difficult task due to the complex image degradation and haze distribution. Current learning-based methods have great performance for RS image dehazing while ignoring the frequency characteristics and treating various features equally. In this study, a stagewise PADehazeNet using progressive frequency filtering and the adaptive self-attention mechanism is proposed. Specifically, a progressive frequency filtering structure is presented to decompose high-frequency features from low-rank features. Progressive feature sharing tensors are used to concatenate and share the multiscale high-frequency features. In addition, an adaptive self-attention mechanism is used to select effective features from low-frequency components so that the global tone can be reconstructed in an efficient way. Extensive experimental results state that the proposed PADehazeNet achieves an outstanding performance compared to recent comparative models on both synthetic and real-world datasets. In summary, this algorithm should be used in the forest fire detection issue via RS.