Official repository for:
A Comparative Benchmark of Geospatial Foundation Models for Land Use/Land Cover Segmentation
Can Michael Hucko, Elif Sertel
Status: The manuscript is currently unpublished. Benchmark results and additional repository content will be released soon.
Geospatial Foundation Models (GFMs) are increasingly used as general-purpose feature extractors for Earth observation, yet their performance can vary substantially across sensors, spatial resolutions, and spectral configurations.
This repository accompanies a systematic benchmark of foundation models for Land Use/Land Cover (LULC) semantic segmentation across six datasets spanning:
- very-high-resolution RGB imagery,
- historical panchromatic imagery,
- multispectral imagery with different band configurations,
- and medium-resolution Sentinel-2 imagery.
The benchmark evaluates models under a consistent experimental setting in order to study how model performance changes with sensor characteristics, spatial resolution, and spectral configuration.
A particular focus is placed on the Multi-Sensor Istanbul datasets, where SPOT-6, SPOT-5, and Sentinel-2 imagery share the same geographical extent, label scheme, and spatial train/validation/test partitions. This enables a controlled comparison of model behavior across sensors.
The benchmark contains six LULC segmentation datasets:
- DeepGlobe — 50 cm RGB imagery
- WorldView-3 — 30 cm RGB + NIR imagery
- HexaLCSeg — ~1 m historical panchromatic Hexagon KH-9 imagery
- SPOT6-LULC — 1.5 m B/G/R/NIR imagery
- SPOT5-LULC — 2.5 m G/R/NIR/SWIR imagery
- Sentinel-2 LULC — 10 m, 12-band multispectral imagery
SPOT6-LULC, SPOT5-LULC, and Sentinel-2 LULC form the Multi-Sensor Istanbul collection and share the same geographical area and LULC label map.
The DeepGlobe dataset was obtained from Kaggle; the Hexagon dataset and the WorldView-3 dataset were obtained from their respective GitHub repositories. The multi-sensor Istanbul dataset is available from the authors upon reasonable request.
The benchmark evaluates models from four main categories.
- SegFormer (MiT-B5)
- ResNeXt-50-UNet
- Mask2Former (Swin-L)
- DINOv2 (ViT-L/14)
- DINOv3-sat (ViT-L/16)
- Prithvi EO v2 300M
- TerraMind v1 Large
- SSL4EO-S12 ResNet-50
- SatMAE
- DOFA Large
- Clay v1.5
- RAMEN
- SMARTIES v1
- THOR Large v1
Two additional dense embedding products are evaluated using lightweight pixel-level classifiers without encoder fine-tuning:
- AlphaEarth
- TESSERA
To make comparisons between pretrained encoders as consistent as possible, the benchmark uses a shared downstream training setup.
All encoder–decoder models are evaluated using:
- a common U-Net decoder,
- identical decoder initialization,
- a two-stage fine-tuning strategy,
- AdamW optimization,
- Dice + Focal segmentation loss,
- dataset-specific normalization computed from the training split,
- spatially separated train/validation/test partitions,
- and identical evaluation procedures.
Stage 1 — Frozen Encoder
The pretrained encoder is frozen and only the decoder is optimized.
Stage 2 — Full Fine-Tuning
The complete encoder–decoder network is unfrozen and jointly optimized using a lower learning rate.
Early stopping is based on validation weighted F-score in both stages.
Results will be released soon.
Unless otherwise stated, the original benchmark data, results, model configurations, trained weights, and documentation provided in this repository are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.
Third-party datasets, pretrained model assets, and other external materials are excluded from this license and remain subject to their respective original terms and licenses.