Mapping and Classification of the Liaohe Estuary Wetland Based on the Combination of Object-Oriented and Temporal Features

For the protection, restoration, and sustainable management of wetland ecosystems, precision in extracting high-quality wetland land cover information is crucial. This study focused on the National Nature Reserve of Liaohe Estuary in Panjin City, Liaoning Province, China. To enhance the classification accuracy of wetland land covers exhibiting similar spectral characteristics and alleviate the occurrence of the 'salt-and-pepper' effect, where certain land parcels are erroneously classified into multiple categories by pixel-based methods, an approach integrating object-oriented techniques and temporal features was employed for precise wetland land cover classification. The analysis utilized multi-temporal Sentinel-2 multispectral images. Initially, the images underwent segmentation using the SNIC method to generate uniform polygons, effectively mitigating misclassification issues. Subsequently, texture, geometry, band reflectance, and spectral deviation features were extracted for each segmented object. A total of 57 features, including vegetation and moisture components, were integrated to construct temporal characteristics. By applying the Random Forest (RF) algorithm in combination with Extreme Randomized Trees (ERT), 18 significant features influencing wetland extraction were identified. These selected features were then utilized to train a Random Forest (RF) model for classifying wetland land cover in the study area. The findings revealed that the integrated object-oriented and temporal feature classification approach achieved an impressive overall accuracy of 95.52% and a Kappa coefficient of 0.95 for the Liaohe Estuary wetland region. The accuracy for various land cover types reached 0.87 for both user and producer accuracy. Compared to alternative machine learning algorithms such as SegUnet++, SVM, and RF, the proposed method demonstrated a performance increase of 16.35%, 14.06%, and 6.14%, respectively. The incorporation of temporal features notably reduced land cover misclassifications, resulting in a 6.14% increase in overall accuracy and a 0.07 improvement in the Kappa coefficient compared to a method lacking temporal features. Particularly for categories like canals, aquaculture, rivers, and reservoirs, producer accuracy improved by over 7.5% and user accuracy by more than 2.9%. The effectiveness of the object-oriented approach was evident in addressing the "salt-and-pepper" effect, showcasing a rise of 2.81% in overall accuracy and 0.03 in Kap