Аңдатпа
Accurate wetland vegetation mapping in alpine environments remains challenging because of high spectral similarity among vegetation classes, complex spatial patterns, seasonal hydrological variation, and the limited availability of field-labelled samples. This study suggests a self-supervised learning (SSL) based multimodal remote sensing approach for vegetation mapping in wetland in the Bosten Lake Alpine Wetland, Xinjiang, China. Single-sensor, paired-fusion, and full multisource fusion inputs were built by combining Pléiades, PlanetScope-3, Sentinel-2, Zhuhai-1 and RadarSat-2 SAR imagery to take advantage of the complementary spatial, spectral and structural information. Four SSL methods (Barlow Twins, SimCLR, SwAV, and SimSiam) were used to pre-train a ResNet-50 model on unlabelled multi-source image patches, and the representations learned by the ResNet-50 were adopted by two segmentation models: DeepLabV3+ and HRNet. The supervised models without SSL pretraining were used for comparison, and the further evaluation was performed in a limited-label case with 3, 5, 7, and 10 labelled polygons per class. By using paired-fusion dataset (PS + ZH), the highest mean Intersection over Union (mIoU) value of 0.65 was obtained as the baseline model DeepLabV3+ performed well using the full multisource fusion input. For all the evaluated configurations, SSL-pretrained models consistently performed better than the supervised baseline models, while multimodal fusion achieved more consistent segmentation performance by fusing complementary optical and SAR multistate information. The results also showed that merely increasing the dimensionality of the spectrum didn't help to realize better segmentation results, as the segmentation results of Zhuhai-1 were still greatly influenced by class confusion and spatial inconsistency. HRNet retained more detailed spatial information, but was more sensitive to local variation and complex class boundaries, while DeepLabV3+ produced more spatially coherent segmentation maps. In summary, the proposed framework clearly illustrates that self-supervised representation learning with multi-modal remote sensing is a promising tool for the mapping of vegetation in wetlands with limited labels and shows great potential for ecological monitoring and planning in alpine wetland environments.