[meta]: Data Quality, Sampling & Augmentation

[Dinghye]

Motivation

Label band analysis (LABEL_BAND_ANALYSIS.md) revealed several data-level issues that directly impact training quality — nodata contamination, extreme class imbalance, nDSM distribution skew, and non-standard patch shapes. The current baseline addresses none of these and uses no augmentation beyond random cropping. Since ~45% of the competition score depends on height RMSE and ~50% depends on Building, fixing data-level problems is likely the highest-ROI path before any architectural changes.

This meta-issue tracks all data-related improvements, from critical fixes to experimental augmentation.

Prior Analysis

• LABEL_BAND_ANALYSIS.md — full dataset statistics, class distribution, nodata characterization, nDSM verification
• BASELINE_REPORT.md — baseline results without any data fixes (AlphaEarth best: composite 0.741)

---

Work Plan

1. Nodata Masking in Loss (Critical)

Problem: 82.7% of samples contain nodata regions (all 4 bands = 0); 32.7% are more than half nodata. The baseline treats these as valid ground truth (predicting zero), injecting ~39% incorrect gradients.

• Implement global validity mask: valid = NOT (band0==0 AND band1==0 AND band2==0 AND band3==0) — exclude from all loss terms
• Implement nDSM-specific mask: additionally exclude pixels where nDSM==0 AND (building>0 OR vegetation>0 OR water>0) — 22.6% of samples have nDSM holes with valid land cover
• Propagate masks through all loss components (MAE, SSIM, Gradient, Tversky)
• Benchmark: expect direct RMSE improvement since ~39% of pixels currently contribute wrong signal

2. Non-standard Patch Shapes

Problem: 258 samples (12.7%) are 255×256 or 256×255 instead of 256×256.

• Verify current reflect-padding handles these correctly
• Ensure padded pixels are included in the nodata mask (not treated as valid ground truth)

3. nDSM Normalization & Outlier Handling

Problem: nDSM is heavily right-skewed (median 0.095 m vs. mean 3.87 m, max 209.9 m). Current baseline simply divides by 30.

• Clip extreme outliers — cap at ~45 m (P99.95) before normalization to prevent rare tall structures from distorting training
• Experiment with log(1+x) transform — may produce a more uniform distribution and improve gradient flow for the majority of near-ground pixels
• Compare /30 vs. log(1+x)/log(31) vs. clipped /30 on validation RMSE

4. Class Imbalance Mitigation

Problem: Building occupies 3.3% of pixels (82.6% of samples are >95% zero), Water 1.7% (93.3% >95% zero).

• Weighted sampling — oversample patches containing Building/Water during training (e.g., WeightedRandomSampler with inverse class-frequency weights)
• Loss reweighting — evaluate increasing loss weight for Building and Water channels beyond current Tversky α=0.3, β=0.7
• Focal loss variant — for sparse classes, down-weight easy negatives to focus on hard positives
• Verify train/val split has enough Building/Water samples for stable metric estimation

5. Geometric Augmentation

Problem: No augmentation beyond random crop. With only 2,024 samples, the model sees limited spatial variation.

• Implement unified spatial augmentation applying identical random transforms to embedding + label:
  • Random horizontal flip (p=0.5)
  • Random vertical flip (p=0.5)
  • Random 90° rotation (k ∈ {0,1,2,3})
• Must work for both PixelEmbeddingDataset (1:1) and LatentTokenDataset (16× multi-scale)
• Benchmark geometric aug vs. no-aug on AlphaEarth

6. Advanced Augmentation Experiments

These are experimental and should only be attempted after items 1–5 are resolved:

• CutMix — paste rectangular region from one sample's embedding+label onto another
• Copy-paste (class-aware) — extract Building/Water regions and paste onto background-dominated samples
• Mosaic — stitch 4 patches into 2×2 grid, crop to training size
• Gaussian noise on embeddings — test σ ∈ {0.01, 0.05, 0.1} relative to per-channel std
• Test-time augmentation (TTA) — predict on flipped/rotated variants and average; zero training cost

7. Ablation & Final Recipe

• Controlled ablation: add each improvement incrementally, measure individual contribution
• Select best recipe by validation composite score
• Verify recipe generalizes across embedding types (AlphaEarth, TerraMind S2, etc.)
• Document final data pipeline config and cumulative performance delta

---

Priority Order

Priority	Task	Expected Impact	Rationale
P0	Nodata masking	High	39% of pixels contribute wrong gradients; affects all metrics
P0	Non-standard shapes	Low effort	Quick fix, prevents edge-case corruption
P1	Class imbalance (sampling + loss)	Medium–High	Building = 50% of score but 3.3% of pixels
P1	nDSM normalization	Medium	Skewed distribution hurts height regression (45% of score)
P2	Geometric augmentation	Medium	Simple, low-risk, expands effective dataset 8×
P3	Advanced augmentation	Unknown	Experimental; depends on results from P0–P2

Success Criteria

• Composite validation score ≥ 0.76 (Δ ≥ +0.02 over baseline 0.741)
• No regression on any individual metric
• Training time overhead < 20%