Wetlands play an essential role in mitigating climate change through their remarkable capacity for carbon sequestration. However, their global degradation underscores the urgent need for precise mapping and monitoring.Deep learning has emerged as a promising solution for automated wetland delineation, enabling large-scale ecosystem monitoring. However, the sparse spatial distribution of wetlands poses a significant challenge for segmentation methods, as many satellite imagery regions contain little to no wetland presence. Traditional loss functions such as Dice Loss fail to provide meaningful gradients in these wetland-sparse scenarios. To address this limitation, we introduce a novel formulation of Flipped Dice Loss that transforms the original pixel-wise relationships to enable gradient propagation in wetland-sparse regions. Building upon this, we develop an Adaptive Dice Loss framework that dynamically adjusts the balance between standard Dice Loss and Flipped Dice Loss using a shifted sigmoid function. Experiments on our newly created Houston Wetland Dataset demonstrate that our method significantly improves wetland detection accuracy compared to state-of-the-art approaches. To facilitate future research in climate-oriented machine learning, we will release our multi-modal Houston Wetland Dataset.