ensemble/ensemble_global_2.json

{
  "type": "Feature",
  "stac_version": "1.1.0",
  "stac_extensions": [
    "https://stac-extensions.github.io/mlm/v1.5.0/schema.json",
    "https://stac-extensions.github.io/file/v2.1.0/schema.json"
  ],
  "id": "ENSEMBLE_2MODELS_MEAN_UNCERTAINTY_2025-10-27",
  "geometry": {
    "type": "Polygon",
    "coordinates": [
      [
        [
          -180.0,
          -90.0
        ],
        [
          -180.0,
          90.0
        ],
        [
          180.0,
          90.0
        ],
        [
          180.0,
          -90.0
        ],
        [
          -180.0,
          -90.0
        ]
      ]
    ]
  },
  "bbox": [
    -180,
    -90,
    180,
    90
  ],
  "properties": {
    "datetime": "2025-10-27T11:26:07Z",
    "created": "2025-10-27T11:26:07Z",
    "updated": "2025-12-02T12:53:01.472464Z",
    "description": "Ensemble of 2 models (1dpwunetpp, unetpp) with Mean aggregation and uncertainty quantification for cloud detection in VGT-1, VGT-2, and PROBA-V satellite imagery.",
    "title": "Ensemble Cloud Detection Model (2 Models + Uncertainty) - VGT1/VGT2/Proba-V",
    "mlm:name": "ensemble_2models_mean_uncertainty_fdr4vgt_cloudmask",
    "mlm:architecture": "Ensemble (Mean+Uncertainty): UNet+++PW, UNet++",
    "mlm:tasks": [
      "semantic-segmentation",
      "uncertainty-quantification"
    ],
    "mlm:framework": "pytorch",
    "mlm:framework_version": "2.5.1+cu121",
    "mlm:accelerator": "cuda",
    "mlm:accelerator_constrained": false,
    "mlm:accelerator_summary": "NVIDIA GPU with CUDA support (compute capability >= 7.0)",
    "mlm:accelerator_count": 1,
    "mlm:memory_size": 92291982,
    "mlm:batch_size_suggestion": 4,
    "mlm:total_parameters": 13651331,
    "mlm:pretrained": true,
    "mlm:pretrained_source": "Global VGT-1/VGT-2/PROBA-V cloud detection models (100k+ training samples)",
    "mlm:input": [
      {
        "name": "VGT_PROBA_TOC_reflectance",
        "bands": [
          "Blue (B0, ~450nm)",
          "Red (B2, ~645nm)",
          "Near-Infrared (B3, ~835nm)",
          "SWIR (MIR, ~1665nm)"
        ],
        "input": {
          "shape": [
            -1,
            4,
            512,
            512
          ],
          "dim_order": [
            "batch",
            "channel",
            "height",
            "width"
          ],
          "data_type": "float32"
        },
        "norm": {
          "type": "raw_toc_reflectance",
          "range": [
            0,
            10000
          ],
          "description": "Raw Top-of-Canopy reflectance values scaled by 10000"
        },
        "pre_processing_function": null
      }
    ],
    "mlm:output": [
      {
        "name": "cloud_probability",
        "tasks": [
          "semantic-segmentation"
        ],
        "result": {
          "shape": [
            -1,
            1,
            512,
            512
          ],
          "dim_order": [
            "batch",
            "channel",
            "height",
            "width"
          ],
          "data_type": "float32"
        },
        "classification:classes": [
          {
            "value": 0.0,
            "name": "clear",
            "description": "Clear sky (may contain cloud shadows)",
            "color_hint": "00000000"
          },
          {
            "value": 1.0,
            "name": "cloud",
            "description": "Cloud present",
            "color_hint": "FFFF00"
          }
        ],
        "post_processing_function": "Apply threshold to get binary mask. Recommended threshold: 0.5. Returns tuple: (probabilities, uncertainty)",
        "standard_threshold": 0.5,
        "recommended_threshold": 0.5,
        "value_range": [
          0.0,
          1.0
        ],
        "description": "Per-pixel mean probability across ensemble models. Built-in sigmoid activation. Values close to 1.0 indicate high confidence of cloud."
      },
      {
        "name": "prediction_uncertainty",
        "tasks": [
          "uncertainty-quantification"
        ],
        "result": {
          "shape": [
            -1,
            1,
            512,
            512
          ],
          "dim_order": [
            "batch",
            "channel",
            "height",
            "width"
          ],
          "data_type": "float32"
        },
        "value_range": [
          0.0,
          1.0
        ],
        "description": "Normalized standard deviation across 2 ensemble members. Values close to 1.0 indicate high disagreement between models (high uncertainty). Automatically returned as second element of output tuple."
      }
    ],
    "mlm:hyperparameters": {
      "ensemble_size": 2,
      "ensemble_members": [
        "1dpwunetpp",
        "unetpp"
      ],
      "aggregation_method": "mean",
      "uncertainty_method": "normalized_std",
      "avg_val_loss": 0.0644,
      "member_details": [
        {
          "model": "1dpwunetpp",
          "epoch": 22,
          "val_loss": 0.0625
        },
        {
          "model": "unetpp",
          "epoch": 5,
          "val_loss": 0.0663
        }
      ]
    },
    "file:size": 61527988,
    "custom:export_format": "torch.export.pt2",
    "custom:has_sigmoid": true,
    "custom:sigmoid_location": "built-in per-model wrapper",
    "custom:export_datetime": "2025-12-02T12:53:01.472464Z",
    "custom:training_stage": "ensemble-mean-uncertainty",
    "custom:project": "FDR4VGT",
    "custom:project_url": "https://fdr4vgt.eu/",
    "custom:sensors": [
      "VGT-1",
      "VGT-2",
      "PROBA-V"
    ],
    "custom:sensor_notes": "Model applicable to SPOT-VGT1, SPOT-VGT2, and PROBA-V imagery",
    "custom:spatial_resolution": "1km",
    "custom:tile_size": 512,
    "custom:recommended_overlap": 64,
    "custom:applicable_start": "1998-03-01T00:00:00Z",
    "custom:applicable_end": null,
    "custom:returns_tuple": true,
    "custom:tuple_format": "(probabilities, uncertainty)",
    "dependencies": [
      "torch>=2.0.0",
      "segmentation-models-pytorch>=0.3.0",
      "pytorch-lightning>=2.0.0",
      "numpy>=1.20.0"
    ]
  },
  "links": [
    {
      "rel": "about",
      "href": "https://fdr4vgt.eu/",
      "type": "text/html",
      "title": "FDR4VGT Project - Harmonized VGT Data Record"
    },
    {
      "rel": "license",
      "href": "https://creativecommons.org/licenses/by/4.0/",
      "type": "text/html",
      "title": "CC-BY-4.0 License"
    }
  ],
  "assets": {
    "model": {
      "href": "https://huggingface.co/isp-uv-es/FDR4VGT-CLOUD/resolve/main/ensemble/ensemble_global_2.pt2",
      "type": "application/octet-stream; application=pytorch",
      "title": "PyTorch ensemble model weights",
      "description": "Ensemble of 2 models in torch.export .pt2 format. Returns tuple: (probabilities, uncertainty).",
      "mlm:artifact_type": "torch.export.pt2",
      "roles": [
        "mlm:model",
        "mlm:weights",
        "data"
      ]
    },
    "example_data": {
      "href": "https://huggingface.co/isp-uv-es/FDR4VGT-CLOUD/resolve/main/ensemble/example_data.safetensor",
      "type": "application/octet-stream; application=safetensors",
      "title": "Example VGT/PROBA-V image",
      "description": "Example VGT/PROBA-V Top-of-Canopy reflectance image for model inference.",
      "roles": [
        "mlm:example_data",
        "data"
      ]
    },
    "load": {
      "href": "https://huggingface.co/isp-uv-es/FDR4VGT-CLOUD/resolve/main/ensemble/load.py",
      "type": "application/x-python-code",
      "title": "PyTorch Ensemble Loader",
      "description": "Python helper code to load the exported .pt2 ensemble model. Includes predict_large() function for large images.",
      "roles": [
        "code"
      ]
    }
  },
  "collection": "ENSEMBLE_2MODELS_FDR4VGT_CloudMask_MeanUncertainty"
}