Abstract
Super-resolution mapping (SRM) is an effective technology to solve the problem of mixed pixels because it can be used to generate fine-resolution land cover maps from coarse-resolution remote sensing images. Current methods based on deep neural networks have been successfully applied to SRM, as they can learn complex spatial patterns from training data. However, they lack the ability to learn structural information between adjacent land cover classes, which is vital in the reconstruction of spatial distribution. In this article, an SRM method based on graph convolutional networks (GCNs), named SRM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GCN</sub> , is proposed to improve SRM results by capturing structure information on the graph. In SRM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GCN</sub> , a supervised inductive learning strategy with mini-graphs as input is considered, which is an extension of the GCN framework. Furthermore, two operations are designed in terms of adjacency matrix construction and an information propagation rule to help reconstruct detailed information of geographical objects. Experiments on three datasets with different spatial resolutions demonstrate the qualitative and quantitative superiority of SRM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GCN</sub> over three other popular SRM methods.
Highlights
REMOTE sensing images are considered the most important input to produce land cover maps
Due to building the spatial contextual structure via graph, SRMGCN(DH) and SRMGCN(LOT) generate results that are more similar to the reference maps than other methods with the correct and complete linear information retained in the regions that are difficult to recover
Similar to the GlobeLand30 dataset, the results in Table IV indicate that the overall accuracy (OA), AA, and kappa coefficient (Kappa) of VBSPM and spatial attraction model (SAM) are at least 2% lower than those gained by the superresolution mapping (SRM) methods based on deep learning, i.e., SRMCNN, SRMGCN(DH), and SRMGCN(LOT)
Summary
REMOTE sensing images are considered the most important input to produce land cover maps. Mixed pixels, which refer to pixels that contain more than one land cover class, constitute a significant obstacle to accurate classification [1]– [3]. It is apparent that the performance of all methods is better than that of the GlobeLand and Slovenia datasets. This can be attributed mainly to the higher proportion of pure pixels (see the last column of Table I) and simpler spatial patterns of the Zurich Summer dataset. Consistent with visual perception, the statistical results of the evaluation indexes of all methods are similar; slight superiorities can still be observed when
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callMore From: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
