{
  "model_name": "Scale-MAE",
  "model_type": "scale_mae",
  "architectures": [
    "ScaleMAE"
  ],
  "framework": "PyTorch",
  "domain": "earth-observation",
  "task": "multiscale-geospatial-representation-learning",
  "implementation": {
    "entry_point": "model/scalemae.py",
    "scope": "GSD-aware masked autoencoding with paired low- and high-frequency target reconstruction",
    "train_script": "scripts/train.py",
    "inference_script": "scripts/inference.py",
    "evaluation_script": "scripts/result.py",
    "synthetic_data_script": "scripts/fake_data.py"
  },
  "architecture": {
    "family": "scale-aware masked autoencoder",
    "input_size": 16,
    "target_size": 32,
    "patch_size": 4,
    "input_patches": 16,
    "in_channels": 3,
    "embed_dim": 64,
    "encoder_depth": 2,
    "encoder_heads": 4,
    "decoder_dim": 48,
    "decoder_depth": 1,
    "decoder_heads": 4,
    "mask_ratio": 0.75,
    "reference_gsd": 1.0,
    "blur_kernel": 5,
    "band_config": {
      "low": {"kernel": 5, "target": "low_frequency"},
      "high": {"residual": true, "target": "high_frequency"}
    }
  },
  "data": {
    "datasets": [
      "FMoW-RGB",
      "RESISC-45",
      "EuroSAT",
      "UCMerced",
      "AID",
      "MLRSNet"
    ],
    "protocol": "synthetic_multiscale_geospatial_rgb",
    "format": "NPZ",
    "input_key": "images",
    "input_shape": ["N", 3, 16, 16],
    "target_key": "targets",
    "target_shape": ["N", 3, 32, 32],
    "gsd_key": "gsd",
    "gsd_shape": ["N"],
    "gsd_values": [0.5, 1.0, 2.0, 4.0],
    "label_key": "labels",
    "label_shape": ["N"],
    "num_classes": 4,
    "default_train_file": "data/train.npz",
    "default_test_file": "data/test.npz"
  },
  "configuration_sources": [
    "conf/config.yaml",
    "model/scalemae.py",
    "scripts/fake_data.py",
    "scripts/train.py",
    "scripts/inference.py",
    "scripts/result.py"
  ]
}