[WIP] Sentinel-1 vessel detection pipeline

.github/workflows/sentinel1_vessel.yaml

@@ -0,0 +1,14 @@
+name: Sentinel1-Vessel-Detection
+
+on:
+  workflow_dispatch: # Manual trigger
+  push:
+    tags:
+      - "sentinel1_vessels_v*" # Trigger only when a version tag (e.g., sentinel1_vessels_v0.0.1) is pushed
+
+jobs:
+  build-and-push:
+    uses: ./.github/workflows/publish_project_docker_image.yaml
+    with:
+      rslp_project: "sentinel1_vessels"
+      image_name: "sentinel1-vessel-detection"

data/sentinel1_vessels/README.md

@@ -0,0 +1,38 @@
+These are the config files for Sentinel-1 vessel detection.
+
+The training data has separate groups for ascending orbit direction versus descending
+orbit direction. This is because the model only performs well with historical images,
+and the historical images are aligned with the current image when they share the same
+orbit direction (otherwise, terrain correction is needed, which is expensive).
+
+
+config.json
+-----------
+
+This is the configuration for the training data.
+
+Note that the orbit_direction attribute of the sentinel1_historical layer needs to be
+set differently for the ascending groups vs the descending groups. See
+`one_off_projects/convert_satlas_webmercator_to_rslearn/sentinel1_vessel/README.md` for
+details.
+
+This is also used for prediction when scene ID is provided and data is fetched from AWS
+bucket. In this case the prediction pipeline populates the items.json so the orbit
+direction in the config is unmodified (but has no effect).
+
+
+config_predict_local_files.json
+-------------------------------
+
+This is for prediction when the user provides the vv/vh files directly. The user must
+ensure the historical images are the same orbit direction as the target image.
+
+
+Dataset Versions
+----------------
+
+- 20250521: this is generated from the siv.sqlite3 in sentinel-vessel-detections.
+- 20250602: this is generated from `gs://satlas-explorer-data/siv-annotations/sentinel1.tar`.
+  It corresponds to the first subset of annotations which are included in 20250521. These
+  annotations may be higher quality but we seem to get better performance from 20250521, so
+  this dataset should not be used.

data/sentinel1_vessels/config.json

@@ -0,0 +1,65 @@
+{
+  "layers": {
+    "label": {
+      "type": "vector"
+    },
+    "mask": {
+      "band_sets": [
+        {
+          "bands": [
+            "mask"
+          ],
+          "dtype": "uint8",
+          "format": {
+            "format": "png",
+            "name": "single_image"
+          }
+        }
+      ],
+      "type": "raster"
+    },
+    "output": {
+      "type": "vector"
+    },
+    "sentinel1": {
+      "alias": "sentinel1",
+      "band_sets": [
+        {
+          "bands": [
+            "vv",
+            "vh"
+          ],
+          "dtype": "float32"
+        }
+      ],
+      "data_source": {
+        "name": "rslearn.data_sources.aws_sentinel1.Sentinel1"
+      },
+      "resampling_method": "cubic",
+      "type": "raster"
+    },
+    "sentinel1_historical": {
+      "alias": "sentinel1",
+      "band_sets": [
+        {
+          "bands": [
+            "vv",
+            "vh"
+          ],
+          "dtype": "float32"
+        }
+      ],
+      "data_source": {
+        "duration": "60d",
+        "name": "rslearn.data_sources.aws_sentinel1.Sentinel1",
+        "orbit_direction": "ASCENDING",
+        "query_config": {
+          "max_matches": 2
+        },
+        "time_offset": "-90d"
+      },
+      "resampling_method": "cubic",
+      "type": "raster"
+    }
+  }
+}

data/sentinel1_vessels/config.yaml

@@ -0,0 +1,220 @@
+model:
+  class_path: rslearn.train.lightning_module.RslearnLightningModule
+  init_args:
+    model:
+      class_path: rslearn.models.multitask.MultiTaskModel
+      init_args:
+        encoder:
+          - class_path: rslearn.models.simple_time_series.SimpleTimeSeries
+            init_args:
+              encoder:
+                class_path: rslearn.models.swin.Swin
+                init_args:
+                  pretrained: false
+                  input_channels: 2
+                  output_layers: [1, 3, 5, 7]
+              image_channels: 2
+              groups: [[0], [1, 2]]
+          - class_path: rslearn.models.fpn.Fpn
+            init_args:
+              in_channels: [256, 512, 1024, 2048]
+              out_channels: 256
+        decoders:
+          detect:
+            - class_path: rslearn.models.faster_rcnn.FasterRCNN
+              init_args:
+                downsample_factors: [4, 8, 16, 32]
+                num_channels: 256
+                num_classes: 2
+                anchor_sizes: [[32], [64], [128], [256]]
+    lr: 0.0001
+    plateau: true
+    plateau_factor: 0.2
+    plateau_patience: 2
+    plateau_min_lr: 1e-6
+    plateau_cooldown: 10
+    restore_config:
+      restore_path: /weka/dfive-default/rslearn-eai/artifacts/satlaspretrain_models/sentinel1_swinb_mi.pth
+      remap_prefixes:
+        - ["backbone.backbone.backbone.", "encoder.0.encoder.model."]
+data:
+  class_path: rslearn.train.data_module.RslearnDataModule
+  init_args:
+    path: /weka/dfive-default/rslearn-eai/datasets/sentinel1_vessels/dataset_v1/20250521/
+    inputs:
+      image_0:
+        data_type: "raster"
+        layers: ["sentinel1"]
+        bands: ["vv", "vh"]
+        passthrough: true
+        dtype: FLOAT32
+      image_1:
+        data_type: "raster"
+        layers: ["sentinel1_historical"]
+        bands: ["vv", "vh"]
+        passthrough: true
+        dtype: FLOAT32
+      image_2:
+        data_type: "raster"
+        layers: ["sentinel1_historical.1"]
+        bands: ["vv", "vh"]
+        passthrough: true
+        dtype: FLOAT32
+      mask:
+        data_type: "raster"
+        layers: ["mask"]
+        bands: ["mask"]
+        passthrough: true
+        dtype: INT32
+        is_target: true
+      targets:
+        data_type: "vector"
+        layers: ["label"]
+        is_target: true
+    task:
+      class_path: rslearn.train.tasks.multi_task.MultiTask
+      init_args:
+        tasks:
+          detect:
+            class_path: rslearn.train.tasks.detection.DetectionTask
+            init_args:
+              property_name: "category"
+              classes: ["unknown", "vessel"]
+              box_size: 15
+              remap_values: [[0, 1], [0, 255]]
+              image_bands: [1, 1, 1]
+              exclude_by_center: true
+              score_threshold: 0.9
+              enable_map_metric: true
+              enable_f1_metric: true
+              f1_metric_thresholds: [[0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95], [0.1], [0.2], [0.3], [0.4], [0.5], [0.6], [0.7], [0.8], [0.9]]
+              f1_metric_kwargs:
+                cmp_mode: "distance"
+                cmp_threshold: 15
+                flatten_classes: true
+        input_mapping:
+          detect:
+            targets: "targets"
+    batch_size: 8
+    num_workers: 32
+    default_config:
+      transforms:
+        - class_path: rslearn.train.transforms.concatenate.Concatenate
+          init_args:
+            selections:
+              image_0: []
+              image_1: []
+              image_2: []
+            output_selector: image
+        - class_path: rslearn.train.transforms.normalize.Normalize
+          init_args:
+            mean: 0
+            std: 250
+            valid_range: [0, 4]
+        - class_path: rslp.transforms.mask.Mask
+    train_config:
+      patch_size: 512
+      transforms:
+        - class_path: rslearn.train.transforms.concatenate.Concatenate
+          init_args:
+            selections:
+              image_0: []
+              image_1: []
+              image_2: []
+            output_selector: image
+        - class_path: rslearn.train.transforms.normalize.Normalize
+          init_args:
+            mean: 0
+            std: 250
+            valid_range: [0, 4]
+        - class_path: rslp.transforms.mask.Mask
+        - class_path: rslearn.train.transforms.flip.Flip
+          init_args:
+            image_selectors: ["image"]
+            box_selectors: ["target/detect"]
+      groups:
+        - apr-2022-point-train_ascending
+        - apr-2022-point-train_descending
+        - jan-march-may-2022-point-train_ascending
+        - jan-march-may-2022-point-train_descending
+        - jun-2020-point-train_ascending
+        - jun-2020-point-train_descending
+        - jun-july-aug-2022-point-train_ascending
+        - jun-july-aug-2022-point-train_descending
+        - nov-2021-point-train_ascending
+        - nov-2021-point-train_descending
+    val_config:
+      load_all_patches: true
+      patch_size: 512
+      groups:
+        - apr-2022-point-val_ascending
+        - apr-2022-point-val_descending
+        - jan-march-may-2022-point-val_ascending
+        - jan-march-may-2022-point-val_descending
+        - jun-2020-point-val_ascending
+        - jun-july-aug-2022-point-val_ascending
+        - jun-july-aug-2022-point-val_descending
+        - nov-2021-point-val_ascending
+        - nov-2021-point-val_descending
+    test_config:
+      load_all_patches: true
+      patch_size: 512
+      groups:
+        - apr-2022-point-val_ascending
+        - apr-2022-point-val_descending
+        - jan-march-may-2022-point-val_ascending
+        - jan-march-may-2022-point-val_descending
+        - jun-2020-point-val_ascending
+        - jun-july-aug-2022-point-val_ascending
+        - jun-july-aug-2022-point-val_descending
+        - nov-2021-point-val_ascending
+        - nov-2021-point-val_descending
+    predict_config:
+      transforms:
+        - class_path: rslearn.train.transforms.concatenate.Concatenate
+          init_args:
+            selections:
+              image_0: []
+              image_1: []
+              image_2: []
+            output_selector: image
+        - class_path: rslearn.train.transforms.normalize.Normalize
+          init_args:
+            mean: 0
+            std: 250
+            valid_range: [0, 4]
+      groups: ["detector_predict"]
+      load_all_patches: true
+      skip_targets: true
+      patch_size: 512
+      overlap_ratio: 0.1
+trainer:
+  max_epochs: 500
+  callbacks:
+    - class_path: lightning.pytorch.callbacks.LearningRateMonitor
+      init_args:
+        logging_interval: "epoch"
+    - class_path: rslearn.train.prediction_writer.RslearnWriter
+      init_args:
+        path: placeholder
+        output_layer: output
+        selector: ["detect"]
+        merger:
+          class_path: rslp.utils.nms.NMSDistanceMerger
+          init_args:
+            grid_size: 64
+            distance_threshold: 10
+            property_name: "category"  # same as task.property_name
+            class_agnostic: false
+    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+      init_args:
+        save_top_k: 1
+        save_last: true
+        monitor: val_detect/mAP
+        mode: max
+    - class_path: rslearn.train.callbacks.freeze_unfreeze.FreezeUnfreeze
+      init_args:
+        module_selector: ["model", "encoder", 0, "encoder", "model"]
+        unfreeze_at_epoch: 4
+rslp_project: sentinel1_vessels
+rslp_experiment: data_20250521_model_20250530_satlaspretrain_unfreeze4_13