Convolutional LSTM and Self-Attention Networks in PyTorch for Decade-Ahead Forecasting of Land-Cover Time Series
Abstract
Forest cover in Vietnam has changed profoundly since 1990 under agricultural expansion, urbanization and plantation programs. This study used the 31 annual JAXA High-Resolution Land-Use and Land-Cover maps of Vietnam (v21.09, 1990-2020) to quantify the dynamics of natural forest and plantations and to forecast land cover to 2030 with two deep learning models implemented in PyTorch: a convolutional long short-term memory network (ConvLSTM) and a Temporal Vision Transformer (ViT). The 250 m Level-2 maps were grouped into seven classes, and their area bias relative to the 30 m maps was quantified. The models were trained on data up to 2010 and validated by hindcasting the 2020 map against persistence, CA-Markov and Random Forest baselines. Natural forest covered 139,100 km 2 in 1990 and 136,800 km 2 in 2020, and plantations covered 66,600 and 49,400 km 2 , respectively, with the largest net forest losses in the Central Highlands and the northern uplands. In the hindcast, the ViT predicted change best (figure of merit 0.203, against 0.182 for the ConvLSTM and 0.165 for the Random Forest), whereas the Random Forest reached the highest overall accuracy (0.746); all trained models predicted about 79,000 km 2 of plantation for 2020, against 49,400 km 2 observed after the sharp plantation decrease in 2017-2018. Having learned the dynamics of 1990-2010, the models forecast for 2030 a further forest loss of 4-5%, with natural forest declining by 11-14% to 117,200-122,100 km 2 and plantations expanding by 13-21% to 56,000-59,700 km 2 , a scenario in which the plantation expansion of the 2000s resumes.
Related
- Rolling bearing fault diagnosis method based on dual attention multi-scale convolutional network
- Transfer Learning-Driven Convolutional Neural Networks for Classification of Plant Species from Leaf Vein Patterns
- Automated correction of training sets for biological image segmentation by convolutional networks
- Relational Graph Convolutional Networks for Glioblastoma Biomarker Discovery via ceRNA and Copy Number Variation Analysis
- Interpreting convolutional neural networks in population genetics.
BibTeX
@article{lemenkova2026convolutional,
title = {{Convolutional LSTM and Self-Attention Networks in PyTorch for Decade-Ahead Forecasting of Land-Cover Time Series}},
author = {Polina Lemenkova},
journal = {Geomatics},
year = {2026},
month = nov,
doi = {10.3390/geomatics6060133},
url = {https://hal.science/hal-05847671v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Convolutional LSTM and Self-Attention Networks in PyTorch for Decade-Ahead Forecasting of Land-Cover Time Series
AU - Polina Lemenkova
PY - 2026
DA - 2026/11/02
JO - Geomatics
DO - 10.3390/geomatics6060133
UR - https://hal.science/hal-05847671v1
AB - Forest cover in Vietnam has changed profoundly since 1990 under agricultural expansion, urbanization and plantation programs. This study used the 31 annual JAXA High-Resolution Land-Use and Land-Cover maps of Vietnam (v21.09, 1990-2020) to quantify the dynamics of natural forest and plantations and to forecast land cover to 2030 with two deep learning models implemented in PyTorch: a convolutional long short-term memory network (ConvLSTM) and a Temporal Vision Transformer (ViT). The 250 m Level-2 maps were grouped into seven classes, and their area bias relative to the 30 m maps was quantified. The models were trained on data up to 2010 and validated by hindcasting the 2020 map against persistence, CA-Markov and Random Forest baselines. Natural forest covered 139,100 km 2 in 1990 and 136,800 km 2 in 2020, and plantations covered 66,600 and 49,400 km 2 , respectively, with the largest net forest losses in the Central Highlands and the northern uplands. In the hindcast, the ViT predicted change best (figure of merit 0.203, against 0.182 for the ConvLSTM and 0.165 for the Random Forest), whereas the Random Forest reached the highest overall accuracy (0.746); all trained models predicted about 79,000 km 2 of plantation for 2020, against 49,400 km 2 observed after the sharp plantation decrease in 2017-2018. Having learned the dynamics of 1990-2010, the models forecast for 2030 a further forest loss of 4-5%, with natural forest declining by 11-14% to 117,200-122,100 km 2 and plantations expanding by 13-21% to 56,000-59,700 km 2 , a scenario in which the plantation expansion of the 2000s resumes.
SN - 2673-7418
ER -