Add eurocrops data module. (microsoft#1869)

* Add eurocrops data module.

It is based on NAIPChesapeakeDataModule which splits bounding box of dataset
into 1/2 train, 1/4 val, and 1/4 test. This may not be the best way to train
an actual model.

* misc fixes

* various fixes per discussion

* update eurocrops test data

* fix

* fix version added placement

* fix failing test by forcing integrity test when checksum is requested

* Clarify SIZE setting in eurocrops test data

* fix currently remaining issues with eurocrops data module

* fix style

* more style fix

* Add documentation

---------

Co-authored-by: Adam J. Stewart <[email protected]>

docs/api/datamodules.rst

@@ -35,6 +35,7 @@ Sentinel
 ^^^^^^^^
 
 .. autoclass:: Sentinel2CDLDataModule
+.. autoclass:: Sentinel2EuroCropsDataModule
 .. autoclass:: Sentinel2NCCMDataModule
 .. autoclass:: Sentinel2SouthAmericaSoybeanDataModule
 

tests/conf/sentinel2_eurocrops.yaml

@@ -0,0 +1,17 @@
+model:
+  class_path: SemanticSegmentationTask
+  init_args:
+    loss: "ce"
+    model: "unet"
+    backbone: "resnet18"
+    in_channels: 13
+    num_classes: 3
+    num_filters: 1
+data:
+  class_path: Sentinel2EuroCropsDataModule
+  init_args:
+    batch_size: 2
+    patch_size: 16
+  dict_kwargs:
+    sentinel2_paths: "tests/data/sentinel2"
+    eurocrops_paths: "tests/data/eurocrops"

tests/data/eurocrops/AA_2022_EC21.cpg

@@ -0,0 +1 @@
+ISO-8859-1

tests/data/eurocrops/AA_2022_EC21.prj

@@ -0,0 +1 @@
+PROJCS["WGS_1984_UTM_Zone_16N",GEOGCS["GCS_WGS_1984",DATUM["D_WGS_1984",SPHEROID["WGS_1984",6378137.0,298.257223563]],PRIMEM["Greenwich",0.0],UNIT["Degree",0.0174532925199433]],PROJECTION["Transverse_Mercator"],PARAMETER["False_Easting",500000.0],PARAMETER["False_Northing",0.0],PARAMETER["Central_Meridian",-87.0],PARAMETER["Scale_Factor",0.9996],PARAMETER["Latitude_Of_Origin",0.0],UNIT["Meter",1.0]]

tests/data/eurocrops/data.py

@@ -12,15 +12,22 @@
 from rasterio.crs import CRS
 from shapely.geometry import Polygon, mapping
 
-SIZE = 100
+# Size of example crop field polygon in projection units.
+# This is set to align with Sentinel-2 test data, which is a 128x128 image at 10
+# projection units per pixel (1280x1280 projection units).
+SIZE = 1280
 
 
 def create_data_file(dataname):
     schema = {"geometry": "Polygon", "properties": {"EC_hcat_c": "str"}}
     with fiona.open(
-        dataname, "w", crs=CRS.from_epsg(31287), driver="ESRI Shapefile", schema=schema
+        dataname, "w", crs=CRS.from_epsg(32616), driver="ESRI Shapefile", schema=schema
     ) as shpfile:
         coordinates = [[0.0, 0.0], [0.0, SIZE], [SIZE, SIZE], [SIZE, 0.0], [0.0, 0.0]]
+        # The offset aligns with tests/data/sentinel2/data.py.
+        offset = [399960, 4500000 - SIZE]
+        coordinates = [[x + offset[0], y + offset[1]] for x, y in coordinates]
+
         polygon = Polygon(coordinates)
         properties = {"EC_hcat_c": "1000000010"}
         shpfile.write({"geometry": mapping(polygon), "properties": properties})
@@ -36,12 +43,12 @@ def create_csv(fname):
 
 if __name__ == "__main__":
     csvname = "HCAT2.csv"
-    dataname = "AA_2021_EC21.shp"
+    dataname = "AA_2022_EC21.shp"
     supportnames = [
-        "AA_2021_EC21.cpg",
-        "AA_2021_EC21.dbf",
-        "AA_2021_EC21.prj",
-        "AA_2021_EC21.shx",
+        "AA_2022_EC21.cpg",
+        "AA_2022_EC21.dbf",
+        "AA_2022_EC21.prj",
+        "AA_2022_EC21.shx",
     ]
     zipfilename = "AA.zip"
 

tests/datasets/test_eurocrops.py

@@ -36,7 +36,7 @@ def dataset(
         monkeypatch.setattr(torchgeo.datasets.utils, "download_url", download_url)
         monkeypatch.setattr(torchgeo.datasets.eurocrops, "download_url", download_url)
         monkeypatch.setattr(
-            EuroCrops, "zenodo_files", [("AA.zip", "9a45308fc32b535318562c1394fd2911")]
+            EuroCrops, "zenodo_files", [("AA.zip", "b2ef5cac231294731c1dfea47cba544d")]
         )
         monkeypatch.setattr(EuroCrops, "hcat_md5", "22d61cf3b316c8babfd209ae81419d8f")
         base_url = os.path.join("tests", "data", "eurocrops") + os.sep

tests/trainers/test_segmentation.py

@@ -74,6 +74,7 @@ class TestSemanticSegmentationTask:
             "sen12ms_s2_all",
             "sen12ms_s2_reduced",
             "sentinel2_cdl",
+            "sentinel2_eurocrops",
             "sentinel2_nccm",
             "sentinel2_south_america_soybean",
             "spacenet1",

torchgeo/datamodules/__init__.py

@@ -30,6 +30,7 @@
 from .seco import SeasonalContrastS2DataModule
 from .sen12ms import SEN12MSDataModule
 from .sentinel2_cdl import Sentinel2CDLDataModule
+from .sentinel2_eurocrops import Sentinel2EuroCropsDataModule
 from .sentinel2_nccm import Sentinel2NCCMDataModule
 from .sentinel2_south_america_soybean import Sentinel2SouthAmericaSoybeanDataModule
 from .skippd import SKIPPDDataModule
@@ -53,6 +54,7 @@
     "L8BiomeDataModule",
     "NAIPChesapeakeDataModule",
     "Sentinel2CDLDataModule",
+    "Sentinel2EuroCropsDataModule",
     "Sentinel2NCCMDataModule",
     "Sentinel2SouthAmericaSoybeanDataModule",
     # NonGeoDataset

torchgeo/datamodules/sentinel2_eurocrops.py

@@ -0,0 +1,121 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+"""EuroCrops datamodule."""
+
+from typing import Any
+
+import kornia.augmentation as K
+import torch
+from kornia.constants import DataKey, Resample
+from matplotlib.figure import Figure
+
+from ..datasets import EuroCrops, Sentinel2, random_grid_cell_assignment
+from ..samplers import GridGeoSampler, RandomBatchGeoSampler
+from ..samplers.utils import _to_tuple
+from ..transforms import AugmentationSequential
+from .geo import GeoDataModule
+
+
+class Sentinel2EuroCropsDataModule(GeoDataModule):
+    """LightningDataModule implementation for the EuroCrops and Sentinel2 datasets.
+
+    Uses the train/val/test splits from the dataset.
+
+    .. versionadded:: 0.6
+    """
+
+    def __init__(
+        self,
+        batch_size: int = 64,
+        patch_size: int | tuple[int, int] = 256,
+        length: int | None = None,
+        num_workers: int = 0,
+        **kwargs: Any,
+    ) -> None:
+        """Initialize a new Sentinel2EuroCropsDataModule instance.
+
+        Args:
+            batch_size: Size of each mini-batch.
+            patch_size: Size of each patch, either ``size`` or ``(height, width)``.
+            length: Length of each training epoch.
+            num_workers: Number of workers for parallel data loading.
+            **kwargs: Additional keyword arguments passed to
+                :class:`~torchgeo.datasets.EuroCrops` (prefix keys with ``eurocrops_``)
+                and :class:`~torchgeo.datasets.Sentinel2`
+                (prefix keys with ``sentinel2_``).
+        """
+        eurocrops_signature = "eurocrops_"
+        sentinel2_signature = "sentinel2_"
+        self.eurocrops_kwargs = {}
+        self.sentinel2_kwargs = {}
+        for key, val in kwargs.items():
+            if key.startswith(eurocrops_signature):
+                self.eurocrops_kwargs[key[len(eurocrops_signature) :]] = val
+            elif key.startswith(sentinel2_signature):
+                self.sentinel2_kwargs[key[len(sentinel2_signature) :]] = val
+
+        super().__init__(
+            EuroCrops,
+            batch_size,
+            patch_size,
+            length,
+            num_workers,
+            **self.eurocrops_kwargs,
+        )
+
+        self.train_aug = AugmentationSequential(
+            K.Normalize(mean=self.mean, std=self.std),
+            K.RandomResizedCrop(_to_tuple(self.patch_size), scale=(0.6, 1.0)),
+            K.RandomVerticalFlip(p=0.5),
+            K.RandomHorizontalFlip(p=0.5),
+            data_keys=["image", "mask"],
+            extra_args={
+                DataKey.MASK: {"resample": Resample.NEAREST, "align_corners": None}
+            },
+        )
+
+        self.aug = AugmentationSequential(
+            K.Normalize(mean=self.mean, std=self.std), data_keys=["image", "mask"]
+        )
+
+    def setup(self, stage: str) -> None:
+        """Set up datasets and samplers.
+
+        Args:
+            stage: Either 'fit', 'validate', 'test', or 'predict'.
+        """
+        self.sentinel2 = Sentinel2(**self.sentinel2_kwargs)
+        self.eurocrops = EuroCrops(**self.eurocrops_kwargs)
+        self.dataset = self.sentinel2 & self.eurocrops
+
+        generator = torch.Generator().manual_seed(0)
+        (self.train_dataset, self.val_dataset, self.test_dataset) = (
+            random_grid_cell_assignment(
+                self.dataset, [0.8, 0.1, 0.1], grid_size=8, generator=generator
+            )
+        )
+        if stage in ["fit"]:
+            self.train_batch_sampler = RandomBatchGeoSampler(
+                self.train_dataset, self.patch_size, self.batch_size, self.length
+            )
+        if stage in ["fit", "validate"]:
+            self.val_sampler = GridGeoSampler(
+                self.val_dataset, self.patch_size, self.patch_size
+            )
+        if stage in ["test"]:
+            self.test_sampler = GridGeoSampler(
+                self.test_dataset, self.patch_size, self.patch_size
+            )
+
+    def plot(self, *args: Any, **kwargs: Any) -> Figure:
+        """Run EuroCrops plot method.
+
+        Args:
+            *args: Arguments passed to plot method.
+            **kwargs: Keyword arguments passed to plot method.
+
+        Returns:
+            A matplotlib Figure with the image, ground truth, and predictions.
+        """
+        return self.eurocrops.plot(*args, **kwargs)

torchgeo/datasets/eurocrops.py

@@ -138,6 +138,10 @@ def _check_integrity(self) -> bool:
         Returns:
             True if dataset files are found and/or MD5s match, else False
         """
+        # Check if the extracted files already exist
+        if self.files and not self.checksum:
+            return True
+
         assert isinstance(self.paths, str)
 
         filepath = os.path.join(self.paths, self.hcat_fname)

torchgeo/datasets/geo.py

@@ -608,6 +608,19 @@ class VectorDataset(GeoDataset):
     #: Not used if :attr:`filename_regex` does not contain a ``date`` group.
     date_format = "%Y%m%d"
 
+    @property
+    def dtype(self) -> torch.dtype:
+        """The dtype of the dataset (overrides the dtype of the data file via a cast).
+
+        Defaults to long.
+
+        Returns:
+            the dtype of the dataset
+
+        .. versionadded:: 0.6
+        """
+        return torch.long
+
     def __init__(
         self,
         paths: str | Iterable[str] = "data",
@@ -734,6 +747,7 @@ def __getitem__(self, query: BoundingBox) -> dict[str, Any]:
         # Use array_to_tensor since rasterize may return uint16/uint32 arrays.
         masks = array_to_tensor(masks)
 
+        masks = masks.to(self.dtype)
         sample = {"mask": masks, "crs": self.crs, "bbox": query}
 
         if self.transforms is not None: