<h2 align="center"> Harnessing Massive Satellite Imagery with Efficient Masked Image Modeling</h2> Fengxiang Wang1    Hongzhen Wang2,‡    Di Wang3    Zonghao Guo2 Zhenyu Zhong4    Long Lan1,‡   Wenjing Yang1,‡    Jing Zhang3,‡&nbsp 1 National University of Defense Technology    2Tsinghua University     3Wuhan University   4Nankai University   <div align='center' style="font-size: larger; ">ICCV 2025</div> 📃 <a href="https://arxiv.org/abs/2406.11933" target="_blank">Paper</a> | 🤗 <a href="https://huggingface.co/datasets/initiacms/OpticalRS-4M" target="_blank"> OpticalRS-4M</a> | 🤗 <a href="https://huggingface.co/datasets/initiacms/OpticalRS-13M" target="_blank"> OpticalRS-13M</a> | 🤗 <a href="https://huggingface.co/initiacms/SelectiveMAE" target="_blank">Models</a>

🎯Intruduction

Dataset: OpticalRS-13M is a large-scale remote sensing dataset. This dataset, comprising 13 million optical images, is designed to fully leverage the representation learning capabilities of MIM methods in RS applications, distinguished by its diverse scene details. We also offer a light version, named OpticalRS-4M.
SelectiveMAE: A novel and efficient MIM method tailored for remote sensing images. This method incorporates a new PSTS module, which significantly accelerates convergence and enhances representation learning compared to the original MIM approach.

✅ To do List

[x] Initial release of checkpoint of SelectiveMAE.
[x] Pretraining codes and configs for SelectiveMAE have be released.
[x] OpticalRS-4M dataset has be released.
[x] OpticalRS-13M dataset will be released.
[x] Codes and configs for downstream tasks of Scene Classification.
[x] Codes and configs for downstream tasks of Object Detection and Semantic Segmentation.

🔥 News

2025.08 : The object detection and semantic segmentation codes have been released.
2025.07 : The classification codes have been released.
2025.06 : SelectiveMAE has been accepted by ICCV2025.
2025.06 : OpticalRS-13M has been released on 🤗HuggingFace.
2025.06 : Models have been released on 🤗HuggingFace.
2025.06 : OpticalRS-4M has been released on 🤗HuggingFace.
2025.06 : The pretraining codes of the SelectiveMAE have been released.
2024.06 : Paper has been released on arxiv.
2024.06 : The training logs and checkpoints of the SelectiveMAE have been released.

🚀OpticalRS-4M

Usage

OpticalRS-4M available on 🤗HuggingFace via OpticalRS-4M.

Use the following command to unzip:

# if 7z is available
7z x OpticalRS-4M.zip
# if zip and unzip is available
zip -s 0 OpticalRS-4M.zip --out whole.zip
unzip whole.zip

Experiments on OpticalRS-4M

OpticalRS-4M offers a significantly larger and more diverse image set compared to previous datasets. To evaluate its effectiveness, we pre-train a ViT-Base model using the vanilla MAE method. For comparison, we use the MillionAID dataset, maintaining an equal number of data points during training: 800 epochs for MillionAID's 1 million images and 200 epochs for our OpticalRS-4M dataset.

| Dataset | Pretrained model | Images Number | Epoch | Sence Classification | Sence Classification | Object Detection | Object Detection | Semantic Segmentation | Semantic Segmentation| |:----------:|:----------------:|:-------------:|:-----:|:---------------------:|:---------------------------:|:-------------------------:|:-----------------:|:--------:|:------------:| | | | | | AID | RESISC-45 | DIOR | DIOR-R | LoveDA | SpaceNetv1 | | | | | | OA (TR=20%/50%) | OA (TR=20%/50%) | mAP50 | mAP50 | mIoU | mF1 | | MillionAID | Weights | 1 million | 800 | 94.92/97.38 | 89.20/93.60 | 71.80 | 62.33 | 51.24 | 79.24 | | OpticalRS-4M | Weights | 2 million | 400 | 96.64/98.10 | 91.80/94.31 | 73.90 | 65.95 | 52.86 | 79.37 | | OpticalRS-4M | Weights | 3 million | 267 | 96.67/98.18 | 92.24/94.41 | 75.40 | 67.07 | 52.39 | 79.37 | | OpticalRS-4M | Weights | 4 million | 200 | 96.10/98.03 | 92.38/94.30 | 74.70 | 66.26 | 52.75 | 79.23 | | OpticalRS-4M | Weights | 4 million | 800 | 96.88/98.22 | 92.44/94.43 | 75.40 | 67.35 | 52.80 | 79.41 |

🚀OpticalRS-13M

OpticalRS-13M available on 🤗HuggingFace via OpticalRS-13M. Follow OpticalRS-4M to unzip.

🚀SelectiveMAE

:gear: Installation for Pretraining

Please install the pretraining dependencies in SelectiveMAE/requirements.txt:

# Optionally create a conda environment
conda create -n selectivemae python=3.10 -y
conda activate selectivemae
# Install dependencies
pip install -r requirements.txt

:blue_car: Pretraining for SelectiveMAE

To pre-train ViT-Base, run the following on 8 GPUs:

torchrun --nproc_per_node=8 --nnodes 1 --master_port 16666 main_pretrain.py --batch_size 256 --selectivemae --dataset opticalrs-4m --dataset_path 'your_dataset_path' --model mae_vit_base_patch16 --output_dir output --norm_pix_loss --blr 1.5e-4 --weight_decay 0.05  --num_workers 12  --decoder_depth 12 --mask_ratio 0.85 --kept_mask_ratio 0.25 --epochs 800 --warmup_epochs 30

First, download the corresponding dataset, then set opticalrs-4m or opticalrs-13m, and update the dataset path accordingly. To train ViT-Small or ViT-Large, set --model mae_vit_small_patch16 or --model mae_vit_large_patch16. You can use --accum_iter to perform gradient accumulation if your hardware could not fit the batch size. FlashAttention 2 should be installed with pip install flash-attn --no-build-isolation.

:rocket: Results on downstream tasks

| Model | Publication | Backbone | Sence Classification | Sence Classification | Object Detection | Object Detection | Semantic Segmentation | Semantic Segmentation | |--------------|:-----------:|:----------:|:---------------------:|:-----------------:|:----------:|:----------:|:------------:|:----------:| | | | | AID | RESISC-45 | DIOR | DIOR-R | LoveDA | SpaceNetv1 | | | | | OA (TR=20%/50%) | OA (TR=20%/50%) | mAP50 | mAP50 | mIoU | mF1 | | SeCo | ICCV'21 | ResNet-50 | 93.47/95.99 | 89.64/92.91 | - | - | 43.63 | 77.09 | | GASSL | ICCV'21 | ResNet-50 | 93.55/95.92 | 90.86/93.06 | 67.40 | 65.65 | 48.76 | 78.51 | | TOV | JSTARS'23 | ResNet-50 | 95.16/97.09 | 90.97/93.79 | 70.16 | 66.33 | 49.70 | - | | CACo | CVPR'23 | ResNet-50 | 90.88/95.05 | 88.28/91.94 | 66.91 | 64.10 | 48.89 | 77.94 | | SatMAE | NIPS'22 | ViT-L | 95.02/96.94 | 91.72/94.10 | 70.89 | 65.66 | - | 78.07 | | ScaleMAE | ICCV'23 | ViT-L | 96.44/97.58 | 92.63/95.04 | 73.81 | 66.47 | - | - | | SSL4EO | GRSM'23 | ViT-S | 91.06/94.74 | 87.60/91.27 | 64.82 | 61.23 | - | - | | RingMo | TGRS'22 | Swin-B | 96.90/98.34 | 94.25/95.67 | 75.90 | - | - | - | | SatLas | ICCV'23 | Swin-B | 94.96/97.38 | 92.16/94.70 | 74.10 | 67.59 | - | - | | GFM | ICCV'23 | Swin-B | 95.47/97.09 | 92.73/94.64 | 72.84 | 67.67 | - | - | | RVSA | TGRS'23 | ViT-B+RVSA | 97.03/98.50 | 93.93/95.69 | 75.80 | 68.06 | 51.95 | - | | SelectiveMAE(OpticalRS-4M) | Baidu & [HuggingFace](https://huggingface.co/initiacms/S

SelectiveMAE

Install / Use

README