Remote Sensing Image Dehazing Based on an Attention Convolutional Neural Network

Haze may affect the quality of optical remote sensing images, thus limiting the scope of their application. Remote sensing image dehazing has become important in remote sensing image preprocessing, promoting the use of remote sensing data and the precision of target recognition. Existing remote sensing dehazing methods based on simplified atmospheric degradation models are not suitable for the removal of heterogeneous haze that exist in remote sensing images. For this purpose, this study proposes an end-to-end convolutional neural network based on attention mechanism, in which the residual block structure combines both channel and spatial attention mechanisms, and establishes a synthetic high-resolution haze image dataset for full training. Thus, it obtains the desired dehazing model. Finally, this study investigates the dehazing model using a GF-1 image and compares it with existing dehazing methods. The results show that the proposed method improved the image similarity, color authenticity, and haze residue level.