Preprocessing & augmentation

Pipeline

flowchart TB
  subgraph Offline["Offline · prepare_data.py"]
    direction LR
    A[Raw scenes + labels] --> B[Orthorectify to label grid] --> C[Pack 512–1024 px tiles] --> D[Per-tile 2/98 % bounds]
  end
  subgraph Online["Online · data loader"]
    direction LR
    E[Pick store by weight] --> F[GSD jitter 0.30–1.20 m] --> G[Crop in source px → 512 px] --> H[D4 flip / rotate]
  end
  subgraph GPU["On GPU"]
    direction LR
    I[Photometric jitter] --> J[Pan-sharpen · grey] --> K[Normalise mean/std]
  end
  Offline --> Online --> GPU

Radiometric normalisation

Each scene is contrast-stretched between its 2nd and 98th percentiles, per band, then normalised with the encoder’s statistics (mean 0.430 / 0.411 / 0.296, std 0.213 / 0.156 / 0.143). Stretch bounds are computed once per scene, and per tile at packing time, so training and inference see the same radiometry.

This fixed a cross-sensor failure. Without the per-scene stretch, an image from a different sensor gave a “crumpled mountain range”: only 28 % of pixels below 1 m where 56 % was expected (Design findings).

Scale augmentation without padding

Real inputs range from 0.3 m aerial imagery to 1.0 m satellite imagery, so each training crop picks a random target GSD in [0.30, 1.20] m (p = 0.9). The crop is chosen in source pixels, so that resampling it to 512 px yields exactly the target GSD. It is never padded.

Note

In v2 this crop was padded to 512 px. 76.5 % of samples were padded, 49.1 % were more than half padding, and about 41 % of supervised pixels were affected. The network was learning that black borders are ground. v3 removed padding entirely. Note that the 1.2 m upper end is never reached in practice, because the coarsest source tops out at about 0.66 m.

Augmentations

Augmentation	Setting	Where
D4 flips / rotations	8 symmetries, uniform	loader
Brightness · contrast	±0.25 each	GPU, p 0.9
Saturation · gamma	0.35 each	GPU, p 0.9
Per-channel gain	0.12	GPU, p 0.9
Gaussian blur	p 0.2 (0.3 in the final run)	GPU
Gaussian noise	std 0.015	GPU
Pan-sharpen simulation (v5)	chroma downsampled 2.0–3.3×, p 0.5	GPU
Greyscale (v5)	p 0.1	GPU

The two v5 augmentations target Cartosat products. Cartosat MX (1.6 m) is fused with PAN (0.6 m), so colour is much blurrier than luminance, and some inputs are PAN-only.

Sampling

Stores are drawn according to sampler_weights. The v5 final run used:

Store	Weight
GAMUS	2.5
NEON	2.0
MVS3DM	1.5
US3D	1.0
SynRS3D g05	1.0
SynRS3D g1	0.5

Within each store, forested tiles are drawn 2× and sparse tiles 1.5× as often as urban ones (landscape_sampler_boost). An epoch is a fixed number of random crops (default 12,000; 36,000 on the final Modal run), not a pass over the dataset.