Spatiotemporal land-use segmentation of satellite imagery with graph neural networks

Trained model from the PhD dissertation Method for Spatiotemporal Semantic Segmentation of Land Use on Satellite Imagery Using Graph Neural Networks by Domen Kavran, University of Maribor, Faculty of Electrical Engineering and Computer Science (UM FERI), completed in March 2026. The dissertation is in Slovenian.

Dissertation: Digital Library of the University of Maribor
Code: SAMPA-Project/PhD-kavran

Method

Regions. Each image of the time series is segmented into regions.
Graph. A graph is built with spatial edges and temporally directed edges between the regions.
Subgraphs. For every target region (node), a directed subgraph of its spatial and temporal neighbours is sampled.
Classification. A CNN extracts features from the bounding box of every subgraph node, and a graph neural network classifies the target node.

The node predictions are then mapped back to pixels, which gives a time series of land-use maps.

This checkpoint


GNN	Heterogeneous Graph Transformer (`MyModelGraphTransformer`): 1 hidden HGT layer with 8 heads × 16 features, then an output HGT layer, with layer normalisation
Region feature extractor	ShuffleNetV2-x0.5, 1024-d features
Region input	32 × 32 px bounding boxes, 4 channels (RGB + NIR), standardised with `train_set_normalization_data.json`, then a 1×1 conv to 3 channels
Subgraph sampling	`ProposedSampler`, temporal neighbourhood: 0 spatial steps, 1 temporal step into the past, fanout 2
Classes (7)	Impervious surface, Agriculture, Forest & other vegetation, Wetlands, Soil, Water, Snow & ice
Parameters	0.9 M

All settings are in config.json.

Results on DynamicEarthNet

These numbers come from the dissertation abstract.

Method	mIoU	mF1
Proposed, spatial neighbourhood	0.4145 ± 0.0051
Proposed, temporal neighbourhood		0.5202 ± 0.0103
GASSL (best state-of-the-art foundation model)	0.3823 ± 0.0156	0.4908 ± 0.0198

The proposed method was also compared with SeCo, SatMAE and TOV, and its improvements are statistically significant against all compared methods.

The best-trained model, which uses the temporal neighbourhood, reached a wF1 of 0.6905. Per-class results:

Class	F1	IoU
Water	0.9159	0.8449
Forest & other vegetation	0.7890	0.6515
Soil	0.6561	0.4882

Usage

DGL has to match your PyTorch version:

import sys
!{sys.executable} -m pip install torch==2.4.0 torchvision==0.19.0
!{sys.executable} -m pip install dgl -f https://data.dgl.ai/wheels/torch-2.4/repo.html
!{sys.executable} -m pip install torchdata==0.8.0 pydantic huggingface_hub networkx matplotlib

import sys
from huggingface_hub import snapshot_download
repo = snapshot_download("SAMPA-Project/spatiotemporal-land-use-segmentation-gnn")
sys.path.insert(0, repo)
import landuse_model as lm

model, cfg = lm.load_model(local_dir=repo)
graph_nx, masks = lm.load_region("path/to/region/")      # graph.pickle + segmentation_masks.pickle
g = lm.to_dgl(graph_nx, cfg)
node_pred = lm.predict_nodes(model, g, cfg)
pred_maps = lm.nodes_to_maps(g, node_pred, masks)         # (T, H, W) class indices, 255 = no region

The input graphs are produced by source/data_preparation/1_prepare_graphs.ipynb in the GitHub repository. run_proposed_method.ipynb is the complete demo. proposed_method_lib/ is the unchanged library from the GitHub repository.

Files

best_model.pt: model weights (PyTorch state dict)
config.json: architecture, sampling, classes and colours
train_set_normalization_data.json: per-channel means and standard deviations of the training set
landuse_model.py: helper functions for loading and inference
proposed_method_lib/: graph construction, GNN architectures and ProposedSampler
run_proposed_method.ipynb: demo notebook

Acknowledgements

Mentor: Niko Lukač, University of Maribor, FERI.

Contact

domen.kavran1@um.si

Downloads last month: 18