Spatiotemporal fusion of multi-source Chinese Gaofen remote sensing data for mapping costal wetland using stacking ensemble classification method

The precise mapping of coastal wetlands holds great significance in the context of monitoring carbon sequestration and storage within coastal ecosystems, particularly in light of climate change and human-induced activities. Time series and multi-source remote sensing data offers distinct advantage in the spatial and temporal mapping of land use, particularly in wetland systems, encompassing various types of vegetation. The primary aim of this study was to delineate the spatial distribution of land use within the Liao River Delta (LRD) wetland. This was achieved by employing a stacking ensemble model that integrated time series GaoFen-1 (GF-1) optical imagery, GaoFen-3 (GF-3) synthetic aperture radar (SAR) imagery, and Moderate Resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index (NDVI) imagery. The first step involved the application of the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM) to fuse the GF-1 NDVI and MODIS NDVI datasets, resulting in the production of time series NDVI data. Subsequently, the parameters pertaining to vegetation phenology were obtained by employing the threshold method on time series NDVI data. We compiled feature datasets that encompassed GF-1 spectral bands, spectral indices, phenological parameters, and GF-3 SAR features. In order to mitigate data redundancy, the Recursive Feature Elimination and Cross-Validation (RFECV) model was employed to identify and select significant features. Finally, the stacking ensemble model was constructed by combining five base models [K-Nearest Neighbors (KNN), Random Forest (RF), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)] to perform wetland classification. The findings showed the following: (1) ESTARFM was able to successfully fuse GF-1 NDVI and MODIS NDVI data, resulting in a spatiotemporal fusion with a coefficient of determination (R²) of 0.85 and a root mean square error (RMSE) of 0.07. (2) In the process of wetland classification, the RFECV algorithm was employed to select relevant features. Specifically, 75 spectral band features, 89 spectral index features, 13 SAR features, and 7 phenological parameters were identified as significant for this task. (3) A stacking ensemble model was constructed using the aforementioned multi-source features. This model exhibited a robust and consistent performance in wetland classification, achieving the highest overall accuracy of 94.33%. Notably, this accuracy improvement ranged from approximately 0.09% to 10.02% when compared to the individual base models. Thus, the present study has the potential to be utilized in the context of fine-scale wetland monitoring, thereby offering valuable assistance to the field of wetland environmental research.