Clarify RasterDataset documentation for is_image and dtype

docs/api/geo_datasets.csv

@@ -2,15 +2,15 @@ Dataset,Type,Source,License,Size (px),Resolution (m)
 `Aboveground Woody Biomass`_,Masks,"Landsat, LiDAR","CC-BY-4.0","40,000x40,000",30
 `AgriFieldNet`_,"Imagery, Masks",Sentinel-2,"CC-BY-4.0","256x256",10
 `Airphen`_,Imagery,Airphen,-,"1,280x960",0.047--0.09
-`Aster Global DEM`_,Masks,Aster,"public domain","3,601x3,601",30
+`Aster Global DEM`_,DEM,Aster,"public domain","3,601x3,601",30
 `Canadian Building Footprints`_,Geometries,Bing Imagery,"ODbL-1.0",-,-
 `Chesapeake Land Cover`_,"Imagery, Masks",NAIP,"CC-BY-4.0",-,1
 `Global Mangrove Distribution`_,Masks,"Remote Sensing, In Situ Measurements","public domain",-,3
 `Cropland Data Layer`_,Masks,Landsat,"public domain",-,30
 `EDDMapS`_,Points,Citizen Scientists,-,-,-
 `EnviroAtlas`_,"Imagery, Masks","NAIP, NLCD, OpenStreetMap","CC-BY-4.0",-,1
 `Esri2020`_,Masks,Sentinel-2,"CC-BY-4.0",-,10
-`EU-DEM`_,Masks,"Aster, SRTM, Russian Topomaps","CSCDA-ESA",-,25
+`EU-DEM`_,DEM,"Aster, SRTM, Russian Topomaps","CSCDA-ESA",-,25
 `EuroCrops`_,Geometries,EU Countries,"CC-BY-SA-4.0",-,-
 `GBIF`_,Points,Citizen Scientists,"CC0-1.0 OR CC-BY-4.0 OR CC-BY-NC-4.0",-,-
 `GlobBiomass`_,Masks,Landsat,"CC-BY-4.0","45,000x45,000",100

docs/tutorials/custom_raster_dataset.ipynb

@@ -329,7 +329,11 @@
     "\n",
     "### `is_image`\n",
     "\n",
-    "If your data only contains image files, as is the case with Sentinel-2, use `is_image = True`. If your data only contains segmentation masks, use `is_image = False` instead.\n",
+    "If your data only contains model inputs (such as images), use `is_image = True`. If your data only contains ground truth model outputs (such as segmentation masks), use `is_image = False` instead.\n",
+    "\n",
+    "### `dtype`\n",
+    "\n",
+    "Defaults to float32 for `is_image == True` and long for `is_image == False`. This is what you want for 99% of datasets, but can be overridden for tasks like pixel-wise regression (where the target mask should be float32).\n",
     "\n",
     "### `separate_files`\n",
     "\n",

torchgeo/datasets/geo.py

@@ -55,9 +55,11 @@ class GeoDataset(Dataset[dict[str, Any]], abc.ABC):
     based on latitude/longitude. This allows users to do things like:
 
     * Combine image and target labels and sample from both simultaneously
-      (e.g. Landsat and CDL)
+      (e.g., Landsat and CDL)
     * Combine datasets for multiple image sources for multimodal learning or data fusion
-      (e.g. Landsat and Sentinel)
+      (e.g., Landsat and Sentinel)
+    * Combine image and other raster data (e.g., elevation, temperature, pressure)
+      and sample from both simultaneously (e.g., Landsat and Aster Global DEM)
 
     These combinations require that all queries are present in *both* datasets,
     and can be combined using an :class:`IntersectionDataset`:
@@ -69,9 +71,9 @@ class GeoDataset(Dataset[dict[str, Any]], abc.ABC):
     Users may also want to:
 
     * Combine datasets for multiple image sources and treat them as equivalent
-      (e.g. Landsat 7 and Landsat 8)
+      (e.g., Landsat 7 and Landsat 8)
     * Combine datasets for disparate geospatial locations
-      (e.g. Chesapeake NY and PA)
+      (e.g., Chesapeake NY and PA)
 
     These combinations require that all queries are present in *at least one* dataset,
     and can be combined using a :class:`UnionDataset`:
@@ -108,7 +110,7 @@ class GeoDataset(Dataset[dict[str, Any]], abc.ABC):
     def __init__(
         self, transforms: Optional[Callable[[dict[str, Any]], dict[str, Any]]] = None
     ) -> None:
-        """Initialize a new Dataset instance.
+        """Initialize a new GeoDataset instance.
 
         Args:
             transforms: a function/transform that takes an input sample
@@ -344,7 +346,14 @@ class RasterDataset(GeoDataset):
     #: ``start`` and ``stop`` groups.
     date_format = "%Y%m%d"
 
-    #: True if dataset contains imagery, False if dataset contains mask
+    #: True if the dataset only contains model inputs (such as images). False if the
+    #: dataset only contains ground truth model outputs (such as segmentation masks).
+    #:
+    #: The sample returned by the dataset/data loader will use the "image" key if
+    #: *is_image* is True, otherwise it will use the "mask" key.
+    #:
+    #: For datasets with both model inputs and outputs, a custom
+    #: :func:`~RasterDataset.__getitem__` method must be implemented.
     is_image = True
 
     #: True if data is stored in a separate file for each band, else False.
@@ -363,6 +372,10 @@ class RasterDataset(GeoDataset):
     def dtype(self) -> torch.dtype:
         """The dtype of the dataset (overrides the dtype of the data file via a cast).
 
+        Defaults to float32 if :attr:`~RasterDataset.is_image` is True, else long.
+        Can be overridden for tasks like pixel-wise regression where the mask should be
+        float32 instead of long.
+
         Returns:
             the dtype of the dataset
 
@@ -382,7 +395,7 @@ def __init__(
         transforms: Optional[Callable[[dict[str, Any]], dict[str, Any]]] = None,
         cache: bool = True,
     ) -> None:
-        """Initialize a new Dataset instance.
+        """Initialize a new RasterDataset instance.
 
         Args:
             paths: one or more root directories to search or files to load
@@ -605,7 +618,7 @@ def __init__(
         transforms: Optional[Callable[[dict[str, Any]], dict[str, Any]]] = None,
         label_name: Optional[str] = None,
     ) -> None:
-        """Initialize a new Dataset instance.
+        """Initialize a new VectorDataset instance.
 
         Args:
             paths: one or more root directories to search or files to load
@@ -873,9 +886,11 @@ class IntersectionDataset(GeoDataset):
     This allows users to do things like:
 
     * Combine image and target labels and sample from both simultaneously
-      (e.g. Landsat and CDL)
+      (e.g., Landsat and CDL)
     * Combine datasets for multiple image sources for multimodal learning or data fusion
-      (e.g. Landsat and Sentinel)
+      (e.g., Landsat and Sentinel)
+    * Combine image and other raster data (e.g., elevation, temperature, pressure)
+      and sample from both simultaneously (e.g., Landsat and Aster Global DEM)
 
     These combinations require that all queries are present in *both* datasets,
     and can be combined using an :class:`IntersectionDataset`:
@@ -896,7 +911,12 @@ def __init__(
         ] = concat_samples,
         transforms: Optional[Callable[[dict[str, Any]], dict[str, Any]]] = None,
     ) -> None:
-        """Initialize a new Dataset instance.
+        """Initialize a new IntersectionDataset instance.
+
+        When computing the intersection between two datasets that both contain model
+        inputs (such as images) or model outputs (such as masks), the default behavior
+        is to stack the data along the channel dimension. The *collate_fn* parameter
+        can be used to change this behavior.
 
         Args:
             dataset1: the first dataset
@@ -1026,9 +1046,9 @@ class UnionDataset(GeoDataset):
     This allows users to do things like:
 
     * Combine datasets for multiple image sources and treat them as equivalent
-      (e.g. Landsat 7 and Landsat 8)
+      (e.g., Landsat 7 and Landsat 8)
     * Combine datasets for disparate geospatial locations
-      (e.g. Chesapeake NY and PA)
+      (e.g., Chesapeake NY and PA)
 
     These combinations require that all queries are present in *at least one* dataset,
     and can be combined using a :class:`UnionDataset`:
@@ -1049,7 +1069,12 @@ def __init__(
         ] = merge_samples,
         transforms: Optional[Callable[[dict[str, Any]], dict[str, Any]]] = None,
     ) -> None:
-        """Initialize a new Dataset instance.
+        """Initialize a new UnionDataset instance.
+
+        When computing the union between two datasets that both contain model inputs
+        (such as images) or model outputs (such as masks), the default behavior is to
+        merge the data to create a single image/mask. The *collate_fn* parameter can be
+        used to change this behavior.
 
         Args:
             dataset1: the first dataset