Long-term series wetland classification of Guangdong-Hong Kong-Macao Greater Bay Area based on APSMnet

Abstract

Wetlands play a crucial role in achieving carbon peak and carbon neutrality goals. Exploring spatiotemporal distribution is one of the fundamental task in wetland research. However, existing large-scale and long-term series wetland mapping methods have challenges related to classification accuracy and obtaining inter-annual wetland samples. Therefore, a rapid sample collection and precise classification method are needed to support wetland resource assessment, conservation, and ecological restoration. In this paper, we propose a novel deep learning method suitable for large-scale and long-term series wetland classification. First, we acquire and preprocess the long-term series remote sensing data on the Google Earth Engine (GEE) platform. Second, the maximum wetland extent is extracted using decision trees and the Otsu algorithm (OTSU). Third, the attention and pyramid structure-based multidimensional feature extraction network (APSMnet) and the inter-annual sample migration algorithm (ISM) are proposed for realizing inter-annual wetland classification and wetland change analysis. Experiments conducted on Landsat 7 data of the Guangdong-Hong Kong-Macao Greater Bay Area (GBA) from 2000 to 2022 demonstrate that the overall accuracy (OA), average accuracy (AA), and kappa coefficient (Kappa) achieved by the proposed wetland classification method are 0.9664, 0.9396, and 0.9570, respectively. Long-term series trend analysis reveals that mangrove/swamp areas experience fluctuations influenced by urbanization and ecological policies, ultimately undergoing restoration and becoming stable. The codes and datasets are made available publicly at https://github.com/louanjun/APSMnet.

Keywords

• Wetland classification
• Google earth engine
• Deep learning
• Landsat