Forest loss driver prediction pipeline on OlmoEarth platform

.github/workflows/build-test.yml

@@ -111,6 +111,8 @@ jobs:
             -e CI="true" \
             -v ${{env.GOOGLE_GHA_CREDS_PATH}}:/tmp/gcp-credentials.json:ro \
             -e GOOGLE_APPLICATION_CREDENTIALS=/tmp/gcp-credentials.json \
+            -e OEDATASETS_API_URL="${{ secrets.OEDATASETS_API_URL }}" \
+            -e DATASETS_API_TOKEN="${{ secrets.DATASETS_API_TOKEN }}" \
             ${{ needs.build.outputs.ghcr_docker_image }} \
             pytest tests/integration -vv --forked
 

.github/workflows/forest_loss_driver_prediction.yaml

@@ -0,0 +1,73 @@
+name: Forest-Loss-Driver-Prediction
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: '0 8 * * 6'  # Run at 8:00 AM UTC every Saturday
+
+env:
+  OE_PROJECT: "projects.forest_loss_driver.deploy"
+  OE_WORKFLOW: "integrated_pipeline"
+  # Name to use when creating Beaker image.
+  BEAKER_IMAGE_NAME: "forest_loss_driver"
+  # After creation, it is prefixed by username, so when we delete we need to use this
+  # full name.
+  BEAKER_IMAGE_FULL_NAME: "favyen/forest_loss_driver"
+  BEAKER_TOKEN: ${{ secrets.BEAKER_TOKEN }}
+  BEAKER_ADDR: ${{ secrets.BEAKER_ADDR }}
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build_and_start_beaker_job:
+    runs-on: ubuntu-latest-m
+    permissions:
+      contents: read
+      packages: write
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      # Avoid issues with running out of disk space.
+      - name: Cleanup disk space
+        run: |
+          sudo docker rmi $(docker image ls -aq) >/dev/null 2>&1 || true
+          sudo docker image prune --all --force >/dev/null 2>&1 || true
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /usr/local/share/boost
+
+      # Setup Beaker CLI.
+      - name: Setup Beaker
+        uses: allenai/setup-beaker@v2
+        with:
+          token: ${{ env.BEAKER_TOKEN }}
+          workspace: "ai2/earth-systems"
+
+      # Build Docker image.
+      - name: Build Docker image
+        id: build-push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: olmoearth_run_data/forest_loss_driver/deploy.Dockerfile
+          push: false
+          load: true
+          tags: ${{ env.BEAKER_IMAGE_NAME }}
+
+      # We upload the image from previous step to Beaker.
+      - name: Create Beaker Image
+        run: |
+          beaker image delete "$BEAKER_IMAGE_FULL_NAME" || true
+          beaker image create --name "$BEAKER_IMAGE_NAME" "$BEAKER_IMAGE_NAME"
+
+      # Now we can launch the Beaker job.
+      - name: Run integrated pipeline in Beaker job
+        run: |
+          docker run --rm \
+            -e BEAKER_TOKEN="${BEAKER_TOKEN}" \
+            -e BEAKER_ADDR="${BEAKER_ADDR}" \
+            "$BEAKER_IMAGE_NAME" \
+            python -m olmoearth_projects.projects.forest_loss_driver.deploy.start_beaker_job

docs/forest_loss_driver.md

@@ -83,6 +83,38 @@ If you open the `prediction_request_geometry.geojson` in qgis, you should see se
 small polygons. The model will be applied on a 128x128 pixel window centered at each of
 these polygons.
 
+### Historical and multi-region extraction
+
+`extract_alerts` exposes a few options that are useful when building a larger sample,
+e.g. a random sample of historical forest loss events across several countries:
+
+- `countries`: limit alerts to a list of two-letter country codes (requires the Natural
+  Earth `ne_10m_admin_0_countries.shp` shapefile in the working directory).
+- `prediction_utc_time`: set the reference time into the past to extract historical
+  alerts (the window is the `days` preceding this time).
+- `max_number_of_events`: randomly sample at most this many events.
+- `slice_days`: split the time window into consecutive slices of this many days and
+  extract events independently within each slice. Because the GLAD date raster stores a
+  single date per pixel, extracting over a long multi-year window at once can merge
+  spatially-adjacent forest loss from different time periods into a single connected
+  component. Slicing keeps these separate and gives more even temporal coverage. When
+  `slice_days` is set, `max_number_of_events` becomes a per-slice cap. Note that a loss
+  event straddling a slice boundary is split into two components.
+
+For example, to sample up to 2000 events per quarter from each GLAD tile across Peru,
+Bolivia, and Ecuador for Jan 2022 - Dec 2025:
+
+```
+python -m olmoearth_projects.main projects.forest_loss_driver extract_alerts \
+  --extract_alerts_args.gcs_tiff_filenames+=080W_20S_070W_10S.tif \
+  --extract_alerts_args.countries+=PE --extract_alerts_args.countries+=BO --extract_alerts_args.countries+=EC \
+  --extract_alerts_args.prediction_utc_time=2025-12-31T00:00:00+00:00 \
+  --extract_alerts_args.days=1460 \
+  --extract_alerts_args.slice_days=90 \
+  --extract_alerts_args.max_number_of_events=2000 \
+  --extract_alerts_args.out_fname='prediction_request_geometry.geojson'
+```
+
 To run inference:
 
 ```