Stochastic environmental research and risk assessment with machine learning classification and accuracy assessment of coastal wetlands
摘要
Coastal wetlands are ecologically significant ecosystems that provide essential services such as flood control, carbon sequestration, and biodiversity support. However, they are increasingly threatened by climate change, urbanization, and human activities. Effective monitoring and classification are crucial for conservation and sustainable management. Traditional methods often struggle with accuracy and scalability, but remote sensing and machine learning have emerged as powerful tools for large-scale, high-precision wetland mapping. The random forest algorithm is used to classify Level-2 water bodies in coastal wetlands, specifically distinguishing features such as rivers and aquaculture ponds within the broader wetland landscape and assess accuracy using key performance metrics. This study utilizes high-resolution satellite imagery from Sentinel-2 and Landsat 8 for feature extraction and wetland classification using machine learning techniques. The classification results demonstrate strong model performance, particularly for river features, which achieved an F1 score of 0.9362 based on a precision of 93.33% and a detection rate of 93.92%. Although aquaculture ponds showed high precision (96.15%) and a lower detection rate (71.43%), this resulted in a moderate F1 score of 0.8197. The overall classification reliability is supported by a Kappa coefficient of 0.93, reflecting strong agreement with reference data. Furthermore, the classified wetland types were evaluated against global datasets to assess consistency and spatial accuracy. The study identified 30,624 km2 of swamp, 15,774 km2 of marsh, 41,761 km2 of water bodies, and 3,358 km2 of mangrove forests, highlighting variations across different sources such as GWL-FCS30, GlobeLand30, and ESA World Cover. These findings demonstrate the effectiveness of machine learning in wetland classification and emphasize the importance of accuracy assessment. This study provides valuable insights for conservation efforts and suggests integrating multi-temporal and multi-sensor datasets to improve classification robustness and adaptability.
Research Highlights1. Random Forest classified coastal wetlands using Sentinel-2 & Landsat 8 data. 2. High river mapping accuracy (F1: 0.936; Kappa: 0.93) achieved. 3. Aquaculture ponds detected with high precision (96.15%) but moderate recall. 4. Wetland classes validated with global datasets for spatial consistency.