Search

With the increasing demand for high temporal and spatial resolution multispectral image sequences, many studies have been carried out on fusion on Landsat-8 and Sentinel-2 images to obtain image sequences with a revisit cycle of 2 and 3 days and a spatial resolution of 10 m. However, current fusion methods suffer from complex computation and loss of spectral and spatial information. To address these issues, a Landsat-8 and Sentinel-2 image fusion based on multiscale smoothing-sharpening filter (MSSF) method is proposed. MSSF combines well the initial spatial prediction obtained from the Landsat-8 image at the target date and the detailed image extracted from the Sentinel-2 image at the reference date. Thin plate spline interpolation with morphological opening-closing algorithm is implemented on the Landsat-8 image at the target date, and the Laplacian of Gaussian enhancement algorithm is applied to the Sentinel-2 image at the reference date in the preprocessing stage. Smoothing-sharpening filter (SSIF) is employed to separate the high and low frequency components of the two preprocessed images. The multiscale SSIF is then utilized to migrate the details from the preprocessed Sentinel-2 image to the preprocessed Landsat-8 image. The performance of MSSF and five compared methods was evaluated qualitatively and quantitatively. Experiments on three remote sensing data sets gathered from different experimental sites confirm that the proposed MSSF method could efficiently generate Sentinel-2-like images with high spatial and spectral resolution.

Subpixel mapping (SPM) could handle the mixed pixels in coarse original spectral image (COSI) to obtain the fine land-cover class mapping result. In recent years, with the auxiliary spatiotemporal information provided by the same region fine prior spectral image (FPSI), spatiotemporal subpixel mapping (SSPM) has shown greater potential than the traditional SPM methods. However, the inaccurate spatiotemporal information of the FPSI is rarely effective identified due to variation differences in the current SSPM methods, affecting the mapping accuracy. To address the abovementioned issues, SSPM based on priori remote sensing image with variation differences (CVDBI) is proposed. First, the coarse abundance images of COSI and the fine thematic images of FPSI are obtained by unmixing COSI and classifying FPSI. Second, the degradation observation model (DOM) is established to use downsampling matrix to correlate the coarse abundance images of COSI with the ideal thematic images of COSI, and the variation difference observation model (VDOM) is established to use variation difference factor to correlate the fine thematic images of FPSI with the ideal thematic images of COSI. Third, a separable convex optimization model is established for DOM and VDOM. This model optimizes the variation difference factor and the ideal thematic images of COSI. Finally, we use the alternating direction method of multipliers to solve the separable convex optimization problem to produce the final mapping result. Experimental results on the three spectral images show that the proposed CVDBI yields the more accurate mapping result than the traditional SPM methods.