High-Accuracy Mapping of Coastal and Wetland Areas Using Multi-Sensor Data Fusion and Deep Feature Learning

Coastal and wetland regions play vital roles in ecosystem stability, economic development, and climate regulation, particularly along the rapidly changing coastline of eastern Asia. Timely and accurate mapping of these complex environments is essential for effective environmental monitoring and sustainable management. This study proposes a novel classification framework that integrates multi-sensor remote sensing data with deep feature extraction and machine learning algorithms to improve coastal and wetland mapping. High-resolution optical satellite images filtered by MLRSNet were used to classify six key coastal classes: beach, harbor port, wetland, river, lake, and island. Deep features were extracted using a pre-trained MobileNetV2 convolutional neural network and used to train Random Forest and Support Vector Classifier (SVC) models. Experimental results demonstrate that combining deep feature representations with multi-sensor data significantly enhances classification accuracy. The SVC model achieved a high overall classification accuracy of 97.3%, with nearly perfect performance in harbor port classification (F1-score = 0.999), along with high F1-scores for beach (0.986), island (0.981), river (0.960), lake (0.959), and wetland (0.952) classes. Despite the complex nature of wetland areas, the classifier maintained balanced performance across all classes, with macro-averaged accuracy, recall, and F1-score all reaching 0.973. These findings highlight the effectiveness of deep feature embeddings combined with SVC's non-linear decision boundaries for distinguishing natural and man-made coastal components. The framework shows strong potential for large-scale coastal ecosystem mapping and monitoring, even under subtle spectral differences, supporting future environmental management applications. Furthermore, the integration of Google Earth Engine (GEE) enables scalable visualization and operational deployment, demonstrating the framework's suitability for real-world coastal monitoring scenarios.