基于全局先验和误差补偿的稀疏表示遥感图像时空融合

基于稀疏表示的遥感图像时空融合方法存在2个问题:当低分辨率图像退化严重时,即使在字典非常理想的前提下,也无法由低分辨率图像的稀疏表示重建理想的高分辨率图像;参考日期的高分辨率图像可以帮助细节重建,但如果预测日期与参考日期场景发生变化,对局部先验的利用会引入错误。针对上述问题,文章提出基于全局先验和误差补偿的稀疏表示融合算法。首先,提出全局跨尺度先验引导的稀疏重建,利用参考日期高分辨率图像里的相似块对稀疏系数结构进行约束,构建新的目标函数,并提出自适应投影策略,将高、低分辨率图像投影到一个兼顾投影精度和消除二义性的中间尺度上,进行相似块匹配和相似块权重的确定;其次,提出基于稀疏表示的误差估计,在训练阶段,利用参考日期高-低分辨率图像对和稀疏重建结果计算高-低分辨率图像误差对,学习误差字典对,在预测阶段利用耦合算法估计稀疏系数重建图像误差。实验结果表明:相较于次优算法,该算法在 BOREAS 数据集中,光谱角制图(spectral angle mapper, SAM)提高了8%,结构相似性(structural similarity, SSIM)提高了2%;在 CIA 数据集中,SAM 提高了7%,SSIM 提高了2%;在 LGC 数据集中,SAM 提高了8%,SSIM 提高了4%。文章所提算法在不同的数据集下均有最优的表现,不仅提升了预测效果,还具有一定的鲁棒性。

There are two problems with spatiotemporal fusion methods for remote sensing images based on sparse representation: one is that when the low-resolution image is severely degraded, it is not possible to reconstruct a desirable high-resolution image from the sparse representation of the low-resolution image, even when the dictionary is highly desirable; the other is that the high-resolution image of the reference date can facilitate detail reconstruction, but if the scene discrepancies occur between the predicted date and the reference date, the utilization of the local prior introduces errors. To address the problem as above, this paper proposes a sparse representation fusion algorithm based on global prior and error compensation. Firstly, the global cross-scale prior guided sparse reconstruction is proposed, the sparse coefficient structure is constrained by using similar blocks in the high-resolution image of the reference date, a new objective function is constructed, and an adaptive projection strategy is proposed to project the high- and low-resolution images onto an intermediate scale that takes into