Skip to content

depth_fm.data.image_processing.terrain

terrain

Terrain-level scalar metrics derived from elevation patches.

compute_topographic_residual 

compute_topographic_residual(elevation: Tensor, valid_mask: Tensor) -> float

Fit a 2D plane to elevation and return the RMS residual.

Removes macroscopic slopes so the returned scalar isolates true topographic roughness (used as a manifest filter to reject overly-flat patches).

Source code in src/depth_fm/data/image_processing/terrain.py 
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
def compute_topographic_residual(elevation: torch.Tensor, valid_mask: torch.Tensor) -> float:
    """Fit a 2D plane to elevation and return the RMS residual.

    Removes macroscopic slopes so the returned scalar isolates true topographic
    roughness (used as a manifest filter to reject overly-flat patches).
    """
    if elevation.ndim > 2:
        elevation = elevation.squeeze()
        valid_mask = valid_mask.squeeze()

    H, W = elevation.shape

    y = torch.linspace(-1, 1, H, dtype=elevation.dtype, device=elevation.device)
    x = torch.linspace(-1, 1, W, dtype=elevation.dtype, device=elevation.device)
    Y, X = torch.meshgrid(y, x, indexing="ij")

    valid_bool = valid_mask.bool()
    if not valid_bool.any():
        return 0.0

    X_v = X[valid_bool].unsqueeze(1)
    Y_v = Y[valid_bool].unsqueeze(1)
    Z_v = elevation[valid_bool].unsqueeze(1)

    A = torch.cat([X_v, Y_v, torch.ones_like(X_v)], dim=1)
    w = torch.linalg.lstsq(A, Z_v).solution

    Z_pred = A @ w
    return torch.sqrt(torch.mean((Z_v - Z_pred) ** 2)).item()