A Lightweight-Transformer-Enhanced 3DCNN Network for Wetland Mapping With Time-Series Imagery

Due to the high spectral similarity among wetland vegetation types, their uneven spatial distribution, and irregular object boundaries, wetland remote sensing monitoring faces significant challenges. Time-series remote sensing imagery, with its rich spatial-spectral features and vegetation-specific phenological information, offers a new approache for accurate wetland vegetation classification. In recent years, 3D Convlutional Neural Network (3DCNN) have been progressively applied to temporal image classification tasks due to their capability to extract spatial-temporal local features. However, constrained by their limited local receptive fields, they struggle to capture global contextual information, which is one critical factor to classification performance. On the other hand, Transformers have gained widespread adoption due to their powerful ability to model long-range dependencies. Therefore, a hybrid architecture combining 3DCNN with Transformers can model global context on top of spatial-temporal local features, thereby improving classification performance. However, directly integrating a standard Transformer with a 3DCNN results in excessively high computational complexity and make it difficult to capture the key features of complex wetland scenes. To address this issue, this paper proposes a lightweight Transformer architecture combined with a 3DCNN, named U-TSDNet, which efficiently models spatial-temporal global dependencies on the basis of local spatial-temporal features. Additionally, a channel-time attention module tailored for temporal images is designed, enhancing feature representation capabilities through adaptive weighting of key channels and key phenological time points. To further improve the model's ability to capture features of irregularly shaped objects, a deformable convolution is introduced. This structure dynamically adjusts the positions of convolution sampling points to achieve precise modeling of irregular geometric structures. The proposed method was tested on time-series Sentinel-2 data from the Yellow River Delta wetland. Experimental results demonstrate that U-TSDNet, with an overall accuracy of 96.76%, achieves superior classification performance compared to existing mainstream methods.