SkySearch: Satellite Video Search at Scale

Official Repository for the paper SkySearch: Satellite Video Search at Scale (KDD 2025 ADS Track).

---

COMS Satellite Image Dataset

We provide 50GB of real-world satellite images captured by the COMS satellite. The dataset is publicly available via Dropbox:

🔗 Download COMS Dataset (Dropbox)

Folder Structure

• Yearly folders of raw satellite images (2014/, 2015/, ..., 2019/)

  • Each folder contains zipped image files.
  • To extract all .zip files at once, run:

    unzip '*.zip'

• Supporting metadata files

  • data_list_ir02: List of selected IR02 image paths.
  • date_list_ir02: Corresponding date information for each IR02 frame.

---

Video Compression

We provide the codebase for training and inference of self-supervised compression models for satellite images and videos. The deployed code can be found in the compression directory.

1. Image Preprocessing

Preprocess raw images and generate downscaled versions for faster ranking and embedding extraction:

python image_preprocess.py \
    --rawdatapath [RAW_IMAGE_DIR] \
    --outputpath [PREPROCESSED_IMAGE_DIR] \
    --smaller_outputdir [DOWNSCALED_IMAGE_DIR_FOR_FAST_RANKING] \
    --rate 0.50

2. Image Encoder Training and Embedding

• Generate triplet training data:

python make_traindata_image.py \
    --imagepath [PREPROCESSED_IMAGE_DIR] \
    --repeat_num 2 \
    --posdiff_list 4 \
    --negdiff_list 1 2 \
    --split_type train \
    --savename default

• Train the image encoder:

python train_image.py \
    --gpu 0 \
    --inputdata default \
    --channel [CHANNEL] \
    --batch_size 128 \
    --epochs 20 \
    --learning_rate 1e-6 \
    --dim 256 \
    --gamma 0.5 \
    --savedir [OUTPUT_DIR]

• Generate image embeddings:

python make_embeddings_image.py \
    --gpu 0 \
    --modelpath [TRAINED_IMAGE_ENCODER_PATH] \
    --outputdir [OUTPUT_DIR] \
    --outputname [OUTPUT_NAME] \
    --channel [CHANNEL]

3. Video Encoder Training and Embedding

• Prepare video sequences:

python preprocess_video.py --length 12 --gap 60

• Generate video triplets:

python make_traindata_video.py \
    --repeat_num 2 \
    --posdiff_list 4 \
    --negdiff_list 1 2 \
    --split_type train

• Train the video encoder:

python train_video.py \
    --gpu 0 \
    --inputdata default \
    --inputemb [IMAGE_EMBEDDING_PATH] \
    --channel [CHANNEL] \
    --batch_size 128 \
    --epochs 20 \
    --learning_rate 1e-6 \
    --dim 256 \
    --gamma 0.5 \
    --videolength 12 \
    --savedir [OUTPUT_DIR]

• Generate video embeddings

python make_embeddings_video.py \
    --gpu 0 \
    --modelpath [TRAINED_VIDEO_ENCODER_PATH] \
    --inputemb [IMAGE_EMBEDDING_PATH] \
    --outputdir [OUTPUT_DIR] \
    --outputname [OUTPUT_NAME] \
    --channel [CHANNEL]

---

Video Prediction (Query Augmentation)

We leverage video prediction to augment the queries and improve long-term search quality. The prediction code can be found here.

To train the model:

python train.py \
  --mse_step [MSE WEIGHT] \
  --gen_step [GENERATOR WEIGHT] \
  --disc_step [DISCRIMINATOR WEIGHT]

• You may additionally tune other hyperparameters such as learning rate, weight decay, etc.

To run inference:

python inference.py \
  --gpu [GPU_ID] \
  --data_path [DATA_PATH] \
  --ckpt_path [CHECKPOINT_PATH] \
  --output_path [OUTPUT_PATH]

• A pre-trained model checkpoint can be found in the above Dropbox directory:

  • ckpt_mse30_gen10_disc60_feat1_orth1_step80000.pth

• The generated images (video frames) will be saved to the specified OUTPUT_PATH in the format:

{index}_{order}.png
- index: the test sample index
- order: the frame number (0 to 11)

• Acknowledgement: This implementation is based on the official SimVP repository.