用于遥感图像时空融合的渐进式特征对齐生成对抗网络OA

Spatiotemporal fusion aims to generate images with high spatiotemporal resolution by fus-ing one or more pairs of high spatial resolution and low spatial resolution images,along with a low spa-tial resolution image at the predicted time point.Traditional methods require at least a pair of high and low spatial resolution images from the same time point as references,but obtaining a sufficient number of image pairs from the same time point is challenging in practical applications.Although leading meth-ods only require one high spatial resolution image from a different time point as a reference,changes in object and ground information between images at different time points can lead to alignment issues be-tween the input images,resulting in information loss or mismatch.Therefore,this paper proposes a progressive feature alignment generative adversarial network(PFAGAN)that resolves alignment issues between images using only one reference image.For coarse-grained alignment,a dual-input spatial fea-ture transformation layer utilizes constraints derived from masks generated by the segment anything model(SAM)to achieve semantic-level alignment of features extracted from the two input images.For fine-grained alignment,a multi-scale spatiotemporal attention fusion module combines temporal and spatial attention to achieve pixel-level alignment.This alignment process,progressing from semantic to pixel levels,forms the basis of the PFAGAN.Comprehensive evaluations on two widely used bench-mark datasets,LGC and CIA,demonstrate that the proposed network achieves superior results.