Uncertainty-aware Deep Learning for Wetlands Typology Mapping from Multi-Source Satellite Remote Sensing Data

Wetlands are dynamic ecosystems whose health and functionality are continually shaped by seasonal fluctuations and long-term shifts in hydrology, climate, and land use. Effective wetlands mapping and monitoring requires methods capable of capturing temporal dynamics, spectral separability, and spatial patterns. Time series satellite observations are invaluable in this regard, as they reveal variations in vegetation phenology, water extent, and other key characteristics over time. To fully leverage temporal information, we employ the Continuous Change Detection and Classification (CCDC) algorithm, which robustly models temporal dynamics by detecting both abrupt and gradual changes, ensuring consistency across seasonal cycles and long-term trends. 

To overcome the limitations of individual sensors, we integrate multi-source satellite data. Sentinel-2 provides detailed spectral information related to vegetation conditions and water properties, while Sentinel-1 C-band SAR enables consistent, cloud-penetrating monitoring of surface water dynamics. PALSAR-2 L-band SAR complements them by capturing sub-canopy inundation and vegetation structure. This synergy of optical and multi-frequency SAR data enables a comprehensive characterization of both surface and sub-surface wetland properties across varying environmental conditions.  

Deep learning architectures such as U-Net outperform traditional pixel-based classifiers (e.g., Random Forests) by leveraging spatial context for object-level predictions. However, large‑scale wetland typology mapping remains challenging due to input‑dependent label noise arising from the integration of multi‑source maps at various spatial resolutions. We propose an uncertainty‑aware segmentation framework that fuses multi‑source satellite data and explicitly models heteroscedastic aleatoric uncertainty. Concretely, we combine a spatial overlap loss (Dice) with a heteroscedastic negative log-likelihood (NLL) to improve robustness to noisy labels and yield calibrated, per‑pixel uncertainty maps for quality control. 

We evaluate the performance of different feature representations derived from multi-source satellite data—including statistical metrics (minimum, maximum, and standard deviation), satellite embeddings, and CCDC-derived temporal features—using both Random Forests and deep learning models. Preliminary results indicate that CCDC features effectively capture temporal wetland dynamics, while spatial context plays a critical role in distinguishing specific wetland types such as marshes, forested wetlands, rivers, and lakes. The resulting uncertainty maps are spatially coherent and consistent with our expectations, showing higher uncertainty along wetland boundaries and lower uncertainty in homogeneous regions, ultimately contributing to more accurate and reliable wetland typology classification.