OneScience-Group/SatlasPretrain

SatlasPretrain

Model Introduction

SatlasPretrain is a multi-task pretraining model for large-scale remote sensing understanding. It uses a dual-stream hierarchical backbone to process multi-temporal high-resolution RGB and Sentinel-2 multispectral imagery and jointly learns dense prediction and global classification tasks.

Paper: SatlasPretrain: A Large-Scale Dataset for Remote Sensing Image Understanding
https://arxiv.org/abs/2211.15660

Model Description

SatlasPretrain was proposed by a research team from the Allen Institute for AI. The model is trained with NAIP-style high-resolution imagery, Sentinel-2 multispectral imagery, and multiple types of remote sensing labels from SatlasPretrain. It is suitable for remote sensing multi-task learning, including semantic segmentation, regression, point, polyline, and polygon object prediction, property recognition, and scene classification.

Use Cases

Scenario	Description
Multi-temporal remote sensing fusion	Jointly process 4 time steps of high-resolution RGB and 8 time steps of Sentinel-2 data.
Remote sensing multi-task learning	Simultaneously predict segmentation, regression, point, polygon, polyline, property, and classification tasks.
Land-cover and crop mapping	Use multi-temporal imagery for semantic segmentation tasks such as land cover and crop type mapping.
Infrastructure feature extraction	Identify point, polyline, and polygon geographic objects such as buildings, roads, railways, airports, and energy facilities.
Local engineering validation	Use a small amount of synthetic data to check the training, inference, and evaluation workflows.
Multi-GPU training	Launch distributed training with torchrun.

Usage Guide

1. OneCode

Experience intelligent one-click AI4S programming through the OneCode online environment:

Click to Experience Intelligent One-Click AI4S Programming

2. Download and Installation

hf download OneScience-Group/SatlasPretrain --local-dir ./SatlasPretrain
cd SatlasPretrain

Environment Dependencies

Hardware Requirements

• A GPU or DCU is recommended.
• CPU can be used for small-configuration connectivity validation; full training and inference will be slow.
• DCU users must install DTK in advance. DTK 25.04.2 or above, or the OneScience recommended version matching the current cluster, is recommended.

DCU Environment

# Please activate DTK and CONDA first
conda create -n onescience311 python=3.11 -y
conda activate onescience311
# uv installation is supported
pip install onescience[earth-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

GPU Environment

# Please activate CONDA first
conda create -n onescience311 python=3.11 -y libstdcxx-ng=12 libgcc-ng=12 gcc_linux-64=12 gxx_linux-64=12
conda activate onescience311
# uv installation is supported
pip install onescience[earth-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

Training Data Introduction

By default, 1 training and 1 test synthetic sample are used to validate the engineering workflow. They are saved as data/train.npz and data/test.npz, respectively.

The synthetic data preserves the official multi-temporal configuration of 4 NAIP time steps, 8 Sentinel-2 time steps, a 512x512 spatial size, and seven task-label specifications.

Real data must be preprocessed and converted to the following NPZ training protocol. This protocol is consistent with the model input specification but is not the download format of the original datasets.

highres_images: float32 [N,4,3,512,512]
lowres_images: float32 [N,8,9,512,512]
valid_highres_times: bool [N,4]
valid_lowres_times: bool [N,8]
sample_ids: string [N]
segmentation: int64 [N,512,512]
regression: float32 [N,1,512,512]
point: float32 [N,1,512,512]
polygon: float32 [N,1,512,512]
polyline: float32 [N,1,512,512]
property: int64 [N]
classification: int64 [N]

fake_data.py automatically writes the protocol and source protocol metadata. These fields must be retained when using real data.

python scripts/fake_data.py

Training

python scripts/train.py

For multi-GPU training, use:

torchrun --nproc_per_node=8 scripts/train.py

Training jointly optimizes seven remote sensing task types and saves a checkpoint and aggregate training metrics. The default configuration is intended for quick workflow validation. Formal experiments should use the multi-temporal data, complete task labels, model configuration, and training duration corresponding to the paper.

result/checkpoints/satlaspretrain.pt
result/training/metrics.json

Training Weights

This repository will provide SatlasPretrain training weights in the weight/ folder. The weight files will be uploaded soon and are expected to be available in the near future.

Inference

python scripts/inference.py

Inference loads the training checkpoint, generates predictions for seven task types, retains sample identity and protocol metadata, and saves the results to:

result/output/predictions.npz

Evaluation and Visualization

python scripts/result.py

Evaluation covers segmentation, regression, point, polyline, and polygon objects, property, and classification tasks, and generates a multi-task prediction figure. Results on synthetic data are only for engineering workflow validation and do not represent full-paper performance.

result/evaluation/metrics.json
result/evaluation/multitask_predictions.png

Official OneScience Resources

Platform	OneScience Main Repository	Skills Repository
Gitee	https://gitee.com/onescience-ai/onescience	https://gitee.com/onescience-ai/oneskills
GitHub	https://github.com/onescience-ai/OneScience	https://github.com/onescience-ai/oneskills

Citation and License

This repository is a reproduction of the original SatlasPretrain paper.

Use of the code and data in this repository remains subject to the licenses and terms of use of their respective projects.