Multi-Sensor Data Fusion and Deep Feature-Based Classification for Mapping China's Coastal and Wetland Areas

Ecosystem stability, economic growth, and climate regulation are all greatly aided by coastal and wetland regions. Particularly in rapidly evolving regions, such as China's extensive coastline, this is the case. Environmental monitoring, shoreline change detection, and sustainable coastal zone management all depend on a precise definition of coastal limits. In this work, we present a deep learning-based method for binary coastline segmentation that combines both deterministic and probabilistic models, utilizing UAV-derived data. We build and evaluate a Residual U-Net and a Bayesian U-Net architecture, the former using Monte Carlo dropout for uncertainty quantification. Derived from UAV video frames and annotated using Label Studio, the collection comprises high-resolution pictures fit for fine-grained coastal edge identification. While the Bayesian U-Net gave equivalent segmentation performance (F1 = 0.78, IoU ≈ 0.65) along with calibrated confidence estimates, the Residual U-Net obtained a best validation F1 score of 0.812 and Intersection over Union (IoU) of 0.686. The Expected Calibration Error (ECE) was used for model calibration; the Bayesian U-Net achieved an ECE as low as 0.0007 with temperature scaling (T = 10). Consistent with the desired probabilistic behavior, uncertainty histograms showed that the model is more uncertain during erroneous predictions than during accurate ones. These results demonstrate the effectiveness of deep learning and uncertainty-aware modeling for UAV-based coastal mapping, highlighting its potential in decision-making systems that require trust, interpretability, and high-resolution coastal monitoring. By combining Bayesian modelling in coastal remote sensing with UAV-based photography, the proposed framework advances the expanding area of multi-sensor data fusion.