Multiscale Deformable DenseNet for Wetland Mapping Using Hyperspectral Images

Wetlands are vital for maintaining ecosystems and supporting biodiversity, but they face increasing threats from climate change and human activities. Accurate mapping of wetlands is essential to detect detrimental changes and guide effective conservation efforts. However, many wetland mapping (WM) methods using convolutional neural networks (CNNs) rely on kernels with fixed sizes and shapes, limiting their ability to capture the multiscale features of wetlands. To enhance their ability, we propose multiscale deformable DenseNet (MDD) by integrating deformable convolutions (DConv) into the DenseNet architecture and using a dual feature extractor. The DConv adapt kernel shapes and sampling to capture spatial patterns across scales, while the dual feature extractor uses varied kernel sizes for diverse receptive fields. These innovations significantly improve the classification accuracy for the complex WM task, where classes are often highly similar. Experimental results demonstrate that MDD achieves the highest overall accuracy (OA) in three hyperspectral datasets, with an OA of 97.24%, 98.23%, and 94.59%, compared with the best competing OAs of 96.33%, 97.16%, and 92.70%, respectively. These results highlight MDD’s superiority in WM.