Using SAM with DEMs?

[chameroy]

Would it be possible to use SAM with single-band rasters such as DEMs?

My understanding that SAM is intended to work with 3-bands such as RGB. However, could a DEM be converted to a 3-band raster to take advantage of SAM?

[giswqs]

yes. it should be.

[khalilurrrahmanridoykhan]

Yes. The current segment-geospatial implementation can read a single-band GeoTIFF directly. read_image_for_sam() passes the raster to prepare_image_for_sam(), which repeats one band into three channels and converts it to contiguous uint8 RGB input.

With the current release:

python -m pip install -U segment-geospatial

from samgeo import SamGeo

sam = SamGeo(model_type="vit_h", automatic=True)
sam.generate(
    "dem.tif",
    output="dem_segments.tif",
    unique=True,
    foreground=True,
)
sam.tiff_to_vector("dem_segments.tif", "dem_segments.gpkg")

Relevant source:
https://github.com/opengeos/segment-geospatial/blob/main/samgeo/common.py#L1149

However, simply copying elevation into R/G/B does not add information—the three channels remain identical. It only makes the tensor shape compatible with a model trained mainly on natural RGB imagery. Whether the masks are useful depends on whether boundaries are visually expressed in the DEM values.

For terrain features, a more informative three-channel visualization is often:

1. robustly stretched elevation,
2. hillshade,
3. slope (or another terrain derivative).

For example:

import numpy as np
import rasterio
from matplotlib.colors import LightSource


def stretch(a, valid):
    lo, hi = np.percentile(a[valid], (2, 98))
    out = np.clip((a - lo) / max(hi - lo, 1e-6), 0, 1)
    return (out * 255).astype("uint8")


with rasterio.open("dem.tif") as src:
    dem = src.read(1, masked=True).astype("float32")
    profile = src.profile.copy()

valid = ~np.ma.getmaskarray(dem)
filled = dem.filled(np.nanmedian(dem[valid]))

# Account for pixel spacing when calculating the gradient.
dy, dx = np.gradient(filled, abs(src.transform.e), abs(src.transform.a))
slope = np.hypot(dx, dy)
hillshade = LightSource(azdeg=315, altdeg=45).hillshade(filled)

rgb = np.stack([
    stretch(filled, valid),
    (hillshade * 255).astype("uint8"),
    stretch(slope, valid),
])
rgb[:, ~valid] = 0

profile.update(count=3, dtype="uint8", nodata=0)
with rasterio.open("dem_terrain_rgb.tif", "w", **profile) as dst:
    dst.write(rgb)

Then run SAM on dem_terrain_rgb.tif and compare the result with the repeated-band version.

Two important limitations:

• SAM produces object masks, not terrain-class labels. It will not inherently identify “ridge”, “valley”, or “landslide”. Classification requires labels, prompts, or a downstream classifier.
• Validate masks against manually reviewed reference polygons. Hillshade azimuth, DEM resolution, nodata edges, and stretch limits can materially change the segmentation.

So: a single-band DEM is supported directly, but a terrain-derivative composite is usually the better experiment when the target boundaries depend on landform rather than absolute elevation.