关键词: 时空融合, 稀疏学习, 单时相, 高分二号, Landsat-8

Abstract: Due to sensor material and satellite orbit parameters, spatial- and temporal- resolution, which it is hard to balance, are inherent characteristics of remotely sensed satellite-sensors and the issue of multitemporal data reconstruction from high resolution satellites has still been a key limitation for its widely application. Considering its effective composite of image information from spatial, spectral and temporal dimension, spatiotemporal fusion technology of multi-source remotely sensed images gets a fast-developing in the last decade and has been becoming a powerful means for solving the reconstruction of multitemporal data. Although learning-based fusion strategy generally has a higher prediction accuracy and potentially utilization than other spatiotemporal fusion methods, it probably generates high uncertainty owing to its high dependence on the process of dictionary training. Aiming at promoting prediction accuracy, computational efficiency and robustness of this fusion strategy, a spatiotemporal fusion frame designed for single high-resolution remotely sensed image is developed by a composite of atmospheric correction based on radiometric regression, data normalization conversion based on error constrains, an adaptive fusion strategy using multi-layer progression and an effective function library for sparse solution. Then remote sensing images from the domestic Gaofen-2 and Landsat-8 satellites are employed for validating the proposed fusion frame. The experimental results indicate that the proposed fusion frame not only raises the computational efficiency but also exhibits a better performance than the existing fusion models for single image pair in imagery fidelity, description of textural feature and spectral consistency.