A Noval Super-Resolution Model for 10-m Mangrove Mapping With Landsat-5

Abstract

Existing temporal mangrove products are at a 30-m resolution from Landsat, facing challenges such as unclear delineation of mangrove community edges, difficulty in identifying creeks and open spaces within communities, and ineffective recognition of small patches. Therefore, there is an urgent need to produce higher resolution temporal mangrove products (e.g., 10-m) with Landsat, particularly considering the absence of available Sentinel imagery before 2015. To this end, we propose a novel super-resolution model that incorporating Residual Channel Attention Networks (RCAN) and Texture Transformer Network (TTSR) to generate 10-m Landsat-5, namely RCAN-TTSR. RCAN and TTSR play crucial roles from different perspectives in the super-resolution process, respectively. TTSR accurately transfers texture information from Sentinel-2 to Landsat by computing the texture correlation between them. On the other hand, RCAN assigns different weights to multiple low-frequency features and a small number of high-frequency features derived from the raw bands of Landsat imagery, thus achieving better super-resolution outcomes. The results demonstrate that images produced by this model significantly outperform existing super-resolution models in terms of PSNR and SSIM metrics. Furthermore, the random forest classifier was employed for mangrove mapping. Compared to 30-m products, our 10-m map shows higher mapping accuracy and finer spatial details.