Deploy forest loss driver Peru + Brazil + Colombia model

.github/workflows/forest_loss_driver_prediction.yaml

@@ -99,7 +99,7 @@ jobs:
               "bucket_name": "dfive-default",
               "mount_path": "/dfive-default"
             }]
-          RSLP_PREFIX: ${{ secrets.RSLP_PREFIX }}
+          RSLP_PREFIX: "/dfive-default/rslearn-eai/"
           TASK_ENV_VARS: |
             [{
               "class_path": "beaker.BeakerEnvVar",

data/forest_loss_driver/20251104/config.json

@@ -0,0 +1,225 @@
+{
+  "layers": {
+    "best_post_0": {
+      "band_sets": [
+        {
+          "bands": [
+            "R",
+            "G",
+            "B"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "best_post_1": {
+      "band_sets": [
+        {
+          "bands": [
+            "R",
+            "G",
+            "B"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "best_post_2": {
+      "band_sets": [
+        {
+          "bands": [
+            "R",
+            "G",
+            "B"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "best_pre_0": {
+      "band_sets": [
+        {
+          "bands": [
+            "R",
+            "G",
+            "B"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "best_pre_1": {
+      "band_sets": [
+        {
+          "bands": [
+            "R",
+            "G",
+            "B"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "best_pre_2": {
+      "band_sets": [
+        {
+          "bands": [
+            "R",
+            "G",
+            "B"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "label": {
+      "class_names": [
+        "agriculture",
+        "agriculture-generic",
+        "agriculture-mennonite",
+        "agriculture-rice",
+        "agriculture-small",
+        "airstrip",
+        "burned",
+        "coca",
+        "flood",
+        "human",
+        "hurricane",
+        "landslide",
+        "logging",
+        "mining",
+        "natural",
+        "none",
+        "river",
+        "road",
+        "unknown",
+        "unlabeled"
+      ],
+      "class_property_name": "new_label",
+      "type": "vector"
+    },
+    "mask": {
+      "band_sets": [
+        {
+          "bands": [
+            "mask"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "mask_vector": {
+      "type": "vector"
+    },
+    "output": {
+      "type": "vector"
+    },
+    "post_sentinel2": {
+      "alias": "sentinel2",
+      "band_sets": [
+        {
+          "bands": [
+            "B01",
+            "B02",
+            "B03",
+            "B04",
+            "B05",
+            "B06",
+            "B07",
+            "B08",
+            "B8A",
+            "B09",
+            "B11",
+            "B12"
+          ],
+          "dtype": "uint16"
+        }
+      ],
+      "data_source": {
+        "cache_dir": "cache/planetary_computer",
+        "duration": "180d",
+        "harmonize": true,
+        "ingest": false,
+        "name": "rslearn.data_sources.planetary_computer.Sentinel2",
+        "query_config": {
+          "max_matches": 4,
+          "space_mode": "CONTAINS"
+        },
+        "sort_by": "eo:cloud_cover",
+        "time_offset": "7d"
+      },
+      "type": "raster"
+    },
+    "pre_sentinel2": {
+      "alias": "sentinel2",
+      "band_sets": [
+        {
+          "bands": [
+            "B01",
+            "B02",
+            "B03",
+            "B04",
+            "B05",
+            "B06",
+            "B07",
+            "B08",
+            "B8A",
+            "B09",
+            "B11",
+            "B12"
+          ],
+          "dtype": "uint16"
+        }
+      ],
+      "data_source": {
+        "cache_dir": "cache/planetary_computer",
+        "duration": "180d",
+        "harmonize": true,
+        "ingest": false,
+        "name": "rslearn.data_sources.planetary_computer.Sentinel2",
+        "query_config": {
+          "max_matches": 4,
+          "space_mode": "CONTAINS"
+        },
+        "sort_by": "eo:cloud_cover",
+        "time_offset": "-300d"
+      },
+      "type": "raster"
+    }
+  }
+}

data/forest_loss_driver/20251104/config.yaml

@@ -0,0 +1,122 @@
+model:
+#  class_path: olmoearth_projects.train.classification_confusion_matrix.CMLightningModule
+  class_path: rslearn.train.lightning_module.RslearnLightningModule
+  init_args:
+    model:
+      class_path: rslearn.models.singletask.SingleTaskModel
+      init_args:
+        # We use SimpleTimeSeries not for time series but just to concatenate the
+        # feature maps for pre-forest-loss and post-forest-loss.
+        encoder:
+          - class_path: rslearn.models.simple_time_series.SimpleTimeSeries
+            init_args:
+              encoder:
+                class_path: rslearn.models.olmoearth_pretrain.model.OlmoEarth
+                init_args:
+                  model_id: OLMOEARTH_V1_BASE
+                  patch_size: 4
+              image_channels: 48
+              image_key: "sentinel2_l2a"
+              groups: [[0], [1]]
+        decoder:
+          - class_path: rslearn.models.pooling_decoder.PoolingDecoder
+            init_args:
+              in_channels: 1536
+              out_channels: 10
+              num_conv_layers: 1
+              num_fc_layers: 2
+          - class_path: rslearn.train.tasks.classification.ClassificationHead
+    scheduler:
+      class_path: rslearn.train.scheduler.PlateauScheduler
+      init_args:
+        factor: 0.2
+        patience: 2
+        min_lr: 0
+        cooldown: 10
+    optimizer:
+      class_path: rslearn.train.optimizer.AdamW
+      init_args:
+        lr: 0.0001
+data:
+  class_path: rslearn.train.data_module.RslearnDataModule
+  init_args:
+    path: /weka/dfive-default/rslearn-eai/datasets/forest_loss_driver/dataset_v1/combined/
+    inputs:
+      sentinel2_l2a:
+        data_type: "raster"
+        layers: ["pre_sentinel2", "pre_sentinel2.1", "pre_sentinel2.2", "pre_sentinel2.3", "post_sentinel2", "post_sentinel2.1", "post_sentinel2.2", "post_sentinel2.3"]
+        bands: ["B02", "B03", "B04", "B08", "B05", "B06", "B07", "B8A", "B11", "B12", "B01", "B09"]
+        passthrough: true
+        load_all_layers: true
+      targets:
+        data_type: "vector"
+        layers: ["label"]
+        is_target: true
+    task:
+      class_path: rslearn.train.tasks.classification.ClassificationTask
+      init_args:
+        property_name: "new_label"
+        classes: ["agriculture", "mining", "airstrip", "road", "logging", "burned", "landslide", "hurricane", "river", "none"]
+        allow_invalid: true
+        metric_kwargs:
+          average: "micro"
+        prob_property: "probs"
+        skip_unknown_categories: true
+    batch_size: 8
+    num_workers: 32
+    default_config:
+      transforms:
+        - class_path: rslearn.train.transforms.pad.Pad
+          init_args:
+            mode: "center"
+            size: 64
+            image_selectors:
+              - sentinel2_l2a
+        - class_path: rslearn.models.olmoearth_pretrain.norm.OlmoEarthNormalize
+          init_args:
+            band_names:
+              sentinel2_l2a: ["B02", "B03", "B04", "B08", "B05", "B06", "B07", "B8A", "B11", "B12", "B01", "B09"]
+        - class_path: rslearn.train.transforms.flip.Flip
+          init_args:
+            image_selectors:
+              - sentinel2_l2a
+        # To support visualization via image key.
+        - class_path: rslearn.train.transforms.concatenate.Concatenate
+          init_args:
+            selections:
+              sentinel2_l2a: [2, 1, 0]
+            output_selector: image
+    train_config:
+      tags:
+        split: train
+    val_config:
+      tags:
+        split: val
+    test_config:
+      tags:
+        split: val
+    predict_config:
+      skip_targets: true
+trainer:
+  max_epochs: 100
+  callbacks:
+    - class_path: lightning.pytorch.callbacks.LearningRateMonitor
+      init_args:
+        logging_interval: "epoch"
+    - class_path: rslearn.train.callbacks.freeze_unfreeze.FreezeUnfreeze
+      init_args:
+        module_selector: ["model", "encoder", 0, "encoder", "model"]
+        unfreeze_at_epoch: 20
+        unfreeze_lr_factor: 10
+    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+      init_args:
+        save_top_k: 1
+        save_last: true
+        monitor: val_accuracy
+        mode: max
+    - class_path: rslearn.train.prediction_writer.RslearnWriter
+      init_args:
+        path: placeholder
+        output_layer: output
+rslp_project: olmoearth_forest_loss_driver
+rslp_experiment: forest_loss_driver_00

data/forest_loss_driver/README.md

@@ -18,11 +18,19 @@ Dataset Versions
 Dataset Configurations
 ----------------------
 
-- config.json: current inference config that uses RGB PNGs.
 - config_ms.json: corresponds to dataset 20250429, it gets L2A images with all bands
   stored as GeoTIFF.
 - config_studio_annotation.json: this is original config used for Brazil+Colombia
   dataset. It creates RGB GeoTIFFs that are good for uploading to and visualizing in
   Studio. It also gets Planet Labs RGB images.
 - config_multimodal.json: this gets inputs that match what Helios can do, Sentinel-2 +
   Sentinel-1 + Landsat.
+
+
+Deployment Details
+------------------
+
+- 20251104: deploy OlmoEarth-v1-FT-ForestLossDriver-Base on Brazil, Peru, and Colombia.
+  The model uses Sentinel-2 L2A images from Microsoft Planetary Computer.
+- 20240912: original deployment trained on Peru only, applying Satlas on Sentinel-2 L1C
+  RGB PNGs.

data/forest_loss_driver/peru_brazil_colombia_model/README.md

@@ -0,0 +1,2 @@
+I tried training SatlasPretrain and OlmoEarth-SwinPretrain on forest loss driver but they
+did not perform very well so did not use them.