Authors
Siyu Lu, Wenfeng Zheng, Chunlai Wu, Lirong Yin
Published in
Environmental research. Pages 126019. Nov 03, 2026. Epub Nov 03, 2026.
Abstract
Wetlands are crucial ecosystems, they store carbon, participate in material dynamic cycles, and provide urban ecosystem services. However, wetlands are shrinking at an alarming rate. Predicting wetland changes is vital for wetland protection. Existing spatiotemporal prediction models have made significant progress in temporal sequence modeling, yet they commonly rely on local convolutional operations for spatial representation and tend to decouple spatial feature learning from temporal memory updating, limiting their ability to capture long-range spatial dependencies and multi-scale geospatial dynamics. To this end, this paper proposes a wetland multiscale spatiotemporal aggregation network (WMS-SAN) to enhance the spatial information capture of multi-temporal wetland data. At the same time, a geospatial information-aware enhanced recurrent memory unit (GRM unit) is proposed in combination with CMT and LSTM to better integrate the local and global features of multi-temporal wetland images. WMS-SAN models time step associations in a single-input and single-output recursive mode, and learns strong dependencies with the help of LSTM hidden state transfer. Experiments show that the model can accurately simulate the evolution range of wetlands, highlighting the advantages of spatial feature learning in comparison with other models. The Moving MNIST experiment provides a standardized supplementary evaluation of the model's generic spatiotemporal prediction capability and enables comparison with established forecasting models.
PMID:
42977234
Bibliographic data and abstract were imported from PubMed on 04 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 8
- Comments 0