PANGAEA: A Global and Inclusive Benchmark for Geospatial Foundation Models

🔥 The pre-print is out!

📚 Introduction

While geospatial foundation models (GFMs) have proliferated rapidly, their evaluations remain inconsistent and narrow. Existing works often utilize suboptimal downstream datasets (e.g., EuroSAT) and tasks (e.g., land cover classification), which constrain comparability and real-world usability. Additionally, a lack of diversity in evaluation protocols, including image resolution and sensor types, further complicates the extensive assessments of GFM performance.

To bridge this gap, we propose a standardized evaluation protocol that incorporates a wide-ranging selection of datasets, tasks, resolutions, and sensor types, establishing a robust and widely applicable benchmark for GFMs.

In this repo, you can find the code to benchmark GFMs. For the moment we included several GFMs that present different approaches. We look forward to adding new models and datasets.

For the moment, we support the following models:

	Paper	GitHub	Keywords
SSL4EOS12	SSL4EO-S12: A Large-Scale Multi-Modal, Multi-Temporal Dataset for Self-Supervised Learning in Earth Observation	link	DINO, MAE, DATA2VEC, MOCO
Scale-MAE	Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning	link	Masked Autoencoders, Multiscale
SatlasNet	SatlasPretrain: A Large-Scale Dataset for Remote Sensing Image Understanding	link	Supervised, Multi-temporal
GFM	Towards Geospatial Foundation Models via Continual Pretraining	link	Swin, Continual Pre-training
SpectralGPT	SpectralGPT: Spectral Remote Sensing Foundation Model	link	MAE, Multi-spectral
DOFA	Neural Plasticity-Inspired Multimodal Foundation Model for Earth Observation	link	MAE, Dynamic bands
CROMA	CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders	link	Contrastive Learning, MAE
Prithvi	Foundation Models for Generalist Geospatial Artificial Intelligence	link	MAE, Multi-temporal
RemoteCLIP	RemoteCLIP: A Vision Language Foundation Model for Remote Sensing	link	Contrastive Learning

And the following datasets:

	Download	Domain	Task	Sensors	Location
HLS Burn Scars	link	Wildfire	Semantic Segmentation	HLS (Harmonized Landsat Sentinel-2)	USA
MADOS	link	Marine	Semantic Segmentation	S2	Global
PASTIS-R	link	Agriculture	Semantic Segmentation	S1, S2, SPOT-6	France
Sen1Floods11	link	Flood	Semantic Segmentation	S1, S2	Global
xView2	link	HADR	Change Detection	Maxar	Global
Five Billion Pixels	original version (custom version coming soon)	(Urban) Land Cover	Semantic Segmentation	Gaofen-2	China
DynamicEarthNet	link	(Urban) Land Cover	Semantic Segmentation	PlanetFusion	Global
CropTypeMapping-South Sudan	link	Agriculture	Semantic Segmentation	S1, S2, Planet	South Sudan
SpaceNet 7	link	Urban	Change detection/ Semantic Segmentation	Planet	Global
AI4SmallFarms	link	Agriculture	Semantic segmentation	S2	Cambodia/Vietnam
BioMassters	link	Forest	Regression	S1, S2	Finland

The repository supports the following tasks using geospatial (foundation) models:

• Single Temporal Semantic Segmentation
• Multi-Temporal Semantic Segmentation
• Change Detection
• Single Temporal Regression
• Multi-Temporal Regression

It is also possible to train some supervised baselines, based on UNet and ViT.

🗺️ Datasets details

Please refer to Dataset Guide to understand the processing requirements and commands specific to each dataset.

If you want to fast-prototype your model, maybe you want to run fast experiments on smaller datasets. We suggest starting with MADOS, HLSBurnScars, SpaceNet7 and Sen1Floods11 and AI4SmallFarms. They offer good diversity in satellites and domains. In the future, we will release stratified subsets for each dataset to facilitate fast prototyping across all datasets.

🛠️ Setup

Clone the repository:

git clone https://github.com/VMarsocci/pangaea-bench.git
cd pangaea-bench

Dependencies

We provide several ways to install the dependencies.

1. Using either Conda or Mamba:

   conda env create -f environment.yaml
   conda activate pangaea-bench
   

   Optional: install Mamba for faster resolution times

   wget https://github.com/conda-forge/miniforge/releases/download/24.3.0-0/Mambaforge-24.3.0-0-Linux-x86_64.sh
   sh ./Mambaforge-24.3.0-0-Linux-x86_64.sh
   
   mamba env create -f environment.yaml
   mamba activate pangaea-bench
   

2. Using pip, create a Python native virtual environment and install dependencies into it:

   export PANGAEA_PATH=/path/to/venv/pangaea-bench # change this
   python3 -m venv ${PANGAEA_PATH}
   source ${PANGAEA_PATH}/bin/activate
   
   pip install -r requirements.txt
   

Then install the code repository as a development package

pip install --no-build-isolation --no-deps -e .

🏋️ Training

To run experiments, please refer to configs/train.yaml. In it, in addition to some basic info about training (e.g. finetune for fine-tuning also the encoder, limited_label_train to train the model on a stratified subset of labels, num_workers, batch_size and so on), there are 5 different basic configs:

• dataset: Information of downstream datasets such as image size, band_statistics, classes etc.
• decoder: Downstream task decoder fine-tuning related parameters, like the type of architecture (e.g. UPerNet), which multi-temporal strategy to use, and other related hparams (e.g. nr of channels)
• encoder: GFM encoder related parameters. output_layers is used for which layers are used for Upernet decoder.
• preprocessing: Both preprocessing and augmentations steps required for the dataset, such as bands adaptation, normalization, resize/crop.
• task: Information about the trainer and evaluator. Most of the parameters are overwritten in run. Trainer and evaluator can be used for segmentation (SegTrainer) or regression (RegTrainer). Different parameter like precision training (precision) can be set in it.

Other 3 configs are used to set other training parameters:

• criterion: in which you can choose the loss for the training. Consider that if you want to add a custom loss, you should add to pangaea/utils/losses.py. Currently, we support cross_entropy, weigthed_cross_entropy, dice and mae loss functions.
• lr_scheduler: in which you can choose the scheduler. Consider that if you want to add a custom one, you should add to pangaea/utils/schedulers.py.
• optimizer: in which you can choose the optimizer. Consider that if you want to add a custom one, you should add to pangaea/utils/optimizers.py.

We provide several examples of command lines to initialize different training tasks on single GPU.

Please note:

• The repo adopts hydra, so you can easily log your experiments and overwrite parameters from the command line. More examples are provided later.
• To use more gpus or nodes, set --nnodes and --nproc_per_node correspondingly. Please refer to the torchrun doc.

💻 Decoder Finetuning

Single Temporal Semantic Segmentation

Take HLSBurnScars dataset, RemoteCLIP Encoder and Upernet Segmentation Decoder as example:

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=hlsburnscars \
   encoder=remoteclip \
   decoder=seg_upernet\
   preprocessing=seg_default \
   criterion=cross_entropy \
   task=segmentation

If you want to overwrite some parameters (e.g. turn off wandbe, change batch size and the path to the dataset, and use 50% stratified sampled subset for training):

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=hlsburnscars \
   encoder=remoteclip \
   decoder=seg_upernet\
   preprocessing=seg_default \
   criterion=cross_entropy \
   task=segmentation \
   dataset.root_path= /path/to/the/dataset/hlsburnscars \
   batch_size=16 \
   use_wandb=False \
   limited_label_train=0.5 \
   limited_label_strategy=stratified

Multi-Temporal Semantic Segmentation

• Multi-temporal decoder config (e.g. configs/decoder/seg_upernet_mt_ltae.yaml if you want to use ltae as a strategy to combine multi-temporal info) should be used.
• In addition, in the dataset config, indicate the number of time frames, e.g., multi_temporal: 6

An example of using SSL4EO-DINO on CropTypeMapping is as below

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=croptypemapping \
   encoder=ssl4eo_dino \
   decoder=seg_upernet_mt_ltae \
   preprocessing=seg_resize \
   criterion=cross_entropy \
   task=segmentation

To use SatlasNet encoder, the configs/encoder/satlasnet_mi.yaml is required

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=croptypemapping \
   encoder=satlasnet_mi \
   decoder=seg_upernet_mt_ltae \
   preprocessing=seg_resize \
   criterion=cross_entropy \
   task=segmentation

To overwrite parameters, please check the Single Temporal Semantic Segmentation example.

Change Detection

One of the change detection decoder should be used: configs/decoder/seg_siamupernet_conc.yaml employs feature concatenation strategy while configs/decoder/seg_siamupernet_diff.yaml uses feature differencing strategy. For example, Prithvi encoder on xView2:

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=xview2 \
   encoder=prithvi \
   decoder=seg_siamupernet_conc \
   preprocessing=seg_default \
   criterion=cross_entropy \
   task=change_detection

To overwrite parameters, please check the Single Temporal Semantic Segmentation example.

Single Temporal Regression

The regression decoder (e.g. configs/decoder/reg_upernet.yaml) and the regression task (e.g. configs/task/regression.yaml) configs should be used. E.g. Prithvi encoder on BioMassters

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=biomassters \
   encoder=prithvi \
   decoder=reg_upernet \
   preprocessing=reg_default \
   criterion=mse \
   task=regression

To use SatlasNet encoder, the configs/encoder/satlasnet_si.yaml is required. To overwrite parameters, please check the Single Temporal Semantic Segmentation example.

Multi-Temporal Regression

The multi-temporal regression decoder (e.g. configs/decoder/reg_upernet_mt_ltae.yaml or configs/decoder/reg_upernet_mt_linear.yaml) and the regression task (e.g. configs/task/regression.yaml) configs should be used.

Take Prithvi encoder on BioMassters as example:

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=biomassters \
   encoder=prithvi \
   decoder=reg_upernet_mt_ltae \
   preprocessing=reg_default \
   criterion=mse \
   task=regression

To use SatlasNet encoder, please refer to the multi-temporal semantic segmentation example. To overwrite parameters, please check the Single Temporal Semantic Segmentation example.

💻 End-to-end Finetuning

It is enough to add finetune=True to the command line.

For example, for single-temporal semantic segmentation:

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=hlsburnscars \
   encoder=remoteclip \
   decoder=upernet\
   preprocessing=default \
   criterion=cross_entropy \
   task=segmentation \
   finetune=True

💻 Fully Supervised Baseline

The repo supports also training fully supervised baselines (i.e. UNet and ViT). To run these, follow the same command line rules as for other models. Keep in mind that setting finetune=True is necessary since this fully supervised approach trains the model from scratch. An example for single temporal semantic segmentation with UNet is provided (Sen1Floods11 dataset):

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=sen1floods11 \
   encoder=unet_encoder \
   decoder=seg_unet \
   preprocessing=seg_default \
   criterion=cross_entropy \
   task=segmentation \
   finetune=True

There is no multi-temporal UNet supported.

An example for multi-temporal semantic segmentation with ViT is provided (CropTypeMapping-SS dataset):

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
   --config-name=train \
   dataset=croptypemapping \
   encoder=vit_scratch \
   decoder=seg_upernet_mt_ltae \
   preprocessing=seg_default \
   criterion=cross_entropy \
   task=segmentation \
   task.evaluator.inference_mode=whole \
   finetune=true

🔧 Customization

Using Your Own Dataset

Refer to: Adding a new downstream dataset

Using Your Own Model

Refer to: Adding a new geospatial foundation model

🏃 Evaluation

An evaluation step is always run after the training.

If you want to just run an evaluation, indicate the ckpt_dir where the checkpoints and configurations are stored.

torchrun pangaea/run.py --config-name=test ckpt_dir=path_to_ckpt_dir

✏️ Contributing

We appreciate all contributions. Please refer to Contributing Guidelines.

⚠️ TO DO

• host all weights/datasets/subsets on HF (the automatic download is working for all the datasets and models' weights but, respectively, Five Billion Pixels, BioMassters, and GFM. The GFM pretrained model can be downloaded from OneDrive.)
• add hyperparameters search (Optuna)
• support automatic running of all the experiments
• create an Arena to fast benchmark all the GFMs

🧮 Some results

Check the paper for all the insights!

NOTE: if you want to benchmark the results of your model, for a fair comparison do not change the hparams in the configs! Soon we will publish also a set of "benchmark-configs", to support automatic running.

📝 Citation

If you find this work useful, please cite:

@misc{marsocci2024pangaeaglobalinclusivebenchmark,
      title={PANGAEA: A Global and Inclusive Benchmark for Geospatial Foundation Models}, 
      author={Valerio Marsocci and Yuru Jia and Georges Le Bellier and David Kerekes and Liang Zeng and Sebastian Hafner and Sebastian Gerard and Eric Brune and Ritu Yadav and Ali Shibli and Heng Fang and Yifang Ban and Maarten Vergauwen and Nicolas Audebert and Andrea Nascetti},
      year={2024},
      eprint={2412.04204},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.04204}, 
}