结合递进超分辨率与注意力机制的遥感时空融合方法

Remote sensing spatiotemporal fusion method combining progressive super-resolution and attention mechanism

针对遥感卫星获取图像的时间分辨率与空间分辨率之间的矛盾,提出一种结合递进超分辨率与注意力机制的遥感图像时空融合方法。该方法包含递进超分辨率和特征融合模块,前者通过多次2倍上采样实现MODIS和Landsat图像之间的16倍空间分辨率提升,后者则对五种不同尺度的超分辨率特征进行融合。在CIA数据集上与两种不同的时空融合算法进行实验,结果表明,所提方法具有较少的光谱损失和更高的细节重建能力。

Aiming at the contradiction between the temporal resolution and spatial resolution of images acquired by remote sensing satellites,a remote sensing image spatiotemporal fusion method combining progressive super-resolution and attention mechanism is proposed.The method includes progressive super-resolution and feature fusion modules.The former achieves a 16-fold spatial resolution improvement between MODIS and Landsat images by multiple 2-fold upsampling,while the latter fuses superresolution features of five different scales.Experiments with two different spatiotemporal fusion algorithms on the CIA dataset show that the proposed method has less spectral loss and higher detail reconstruction capabilities.