Fine-Grained Natural Grassland Mapping for Sustainable Land Monitoring in Urban–Rural Landscapes Using Foundation Model Embeddings

Fine-grained information on natural grasslands is important for monitoring ecological land patterns and supporting sustainable land management in heterogeneous urban–rural landscapes. However, natural grasslands in subtropical southern China are highly fragmented and frequently embedded within forests, croplands, settlements, and other land-cover types, making their fine-grained identification challenging using conventional land-cover products and single-season remote sensing observations. This study focuses on Hunan Province and constructs a five-class dataset using field-survey plots independently collected during our field campaign, together with public land-cover products for spatially reliable sample screening. Seasonal–phenological features derived from spring, summer, and autumn Sentinel-2 observations, topographic information from DEM, and AlphaEarth Embedding were integrated into a proposed Grassland Mapping Network (GMapNet). GMapNet employs a Transformer to model dependencies among feature groups, Feature Group Attention (FGA) to adaptively aggregate seasonal and environmental information, and a Class Prototype Similarity Head (CPSH) for prototype-based fine-grained discrimination. GMapNet achieved an overall accuracy of 88.58 ± 0.98%, a macro-F1 of 88.46 ± 1.05%, and a Kappa coefficient of 85.72 ± 1.23%, outperforming all compared classification methods. Ablation experiments further confirmed the effectiveness of FGA and CPSH. Removing AlphaEarth Embedding reduced the overall accuracy to 76.58 ± 1.72%, demonstrating that foundation-model embeddings provide substantial complementary information beyond conventional seasonal–phenological–topographic features. The resulting maps for four representative cities displayed pronounced spatial heterogeneity, with predicted natural grasslands distributed mainly as fragmented patches and terrain-associated belts within complex land-cover mosaics. These maps provide finer thematic detail than broad grassland/open-vegetation categories and may inform future ecological monitoring and conservation planning.