Long-term wetland mapping at 10 m resolution using super-resolution and hierarchical classification – a case study in Jianghan Plain, China

Wetland is a critical and complex ecosystem but highly susceptible to human activities. It has experienced significant change along with the increase of human intervention since decades ago. Currently, remote sensing-based mapping is believed an effective method for monitoring wetland changes. However, long-term wetland mapping is difficult due to poor historical data quality and wetland complexity, resulting in the triangular contradiction of long-term, high-resolution, and comprehensive classification systems. To address these challenges, we first modified the single-image-super-resolution method Residual Channel Attention Network (RCAN) and combined it with the famous spatiotemporal fusion model STARFM to enhance the Landsat images to 10 m resolution, which enabled more wetland objects to be identified. Then we extracted water bodies using UNet and classified them by object-based shape indices. Next, we applied recursive feature elimination and random forest classifier to map vegetated wetlands. Our overall classification accuracy reached 85.65%. We observed that the wetland in Jianghan Plain expanded significantly since 2000, with ponds showing the most notable growth rate (>114.04%). Aquatic crops make up a large portion of the wetland area, remaining stable with slow growth since 2000, while lake and surrounding herbaceous wetland restoration has shown notable progress in recent years.