ER-GMMD: Cross-Scene Remote Sensing Classification Method of Tamarix chinensis in the Yellow River Estuary

Abstract

Tamarix chinensis effectively prevents coastal erosion, stabilizes the surface of coastal wetlands, and improves the soil quality of saline-alkali land, playing a crucial role in coastal wetland ecosystem restoration. Tamarix chinensis exhibits a wide distribution that is difficult to capture within a single remote sensing image, while its frequent interspersion with other vegetation results in significant intermixing. The characteristics of mixed tamarix chinensis vary substantially across remote sensing images from different scenarios, and spectral confusion further complicates the process. These factors hinder the extraction and alignment of mixed tamarix chinensis features during classification, resulting in low cross-scene classification accuracy. To address these challenges, this study proposes a deep learning-based cross-domain classification model, ER-GMMD, which leverages features extracted by deep residual networks for different mixed-growth patterns of tamarix chinensis, and integrates dual feature alignment to address the cross-scene classification challenges of mixed-species tamarix chinensis. Utilizing GF remote sensing images covering the tamarix chinensis research area in the Yellow River Delta, along with field survey data, the model achieves precise classification of different mixed tamarix chinensis types. Key results include: 1) The proposed model, trained with only 5% of the source domain samples, achieves an overall classification accuracy of 96.52% on the target domain samples, which is a 17.61% improvement compared with the traditional network U-Net without domain adaptation. 2) Compared with domain adaptation algorithms DAN and S-DMM, the proposed ER-GMMD model demonstrates higher accuracy on the constructed dataset, indicating its potential for high-precision classification of mixed vegetation in coastal wetlands.