3 Feature Engineering
The pipeline assembles 591 features per segment. Features fall into two categories: pixel features (computed at the raster level, then summarised per segment as mean + std) and segment-only features (computed directly from segment geometry or label raster).
3.1 Feature summary
| Group | Features | Type |
|---|---|---|
| Spectral bands (8 bands × 10 epochs) | 80 | Pixel |
| Legacy indices (NDVI/NDWI/NDBI/EVI × 10 epochs) | 40 | Pixel |
| Temporal summaries (max/min/std/amp × 4 indices) | 16 | Pixel |
| Tree height ETH (mean + std within 30 m) | 2 | Pixel |
| New per-epoch indices (10 indices × 10 epochs) | 100 | Pixel |
| Temporal percentiles (p10/p50/p90 × 4 indices) | 12 | Pixel |
| Harmonic fit (amp/phase/offset × 4 indices) | 12 | Pixel |
| Year-over-year delta (March 2024 → March 2026 × 4 indices) | 4 | Pixel |
| Texture (entropy, range, CV, MAD of NDVI/NIR) | 6 | Segment |
| Shape (area, perimeter, compactness, elongation, fill) | 5 | Segment |
| Sentinel-1 SAR (19 bands × mean + std) | 38 | Pixel |
| PALSAR-2 (5 bands × mean + std) | 10 | Pixel |
| Meta v2 canopy (4 statistics × mean + std) | 8 | Pixel |
| Total | 591 |
3.2 Legacy pixel cache (138 features)
Built by planetscope_10epoch_local.py. For each of the 10 epochs:
- 8 raw spectral bands (
{epoch}_b1…{epoch}_b8) - 4 spectral indices: NDVI, NDWI, NDBI, EVI
Plus temporal summary statistics across all 10 epochs (4 indices × 4 stats = 16 features):
max{index}, min{index}, std{index}, amp{index}And 2 ETH canopy height features: tree_height_mean, tree_height_std.
Band mapping (legacy): the legacy pixel script treats b2=Blue, b3=Green_I, b4=Red, b6=Red, b8=NIR. The v3 extension uses the full SuperDove mapping for new indices only; legacy features are unchanged to preserve reproducibility.
3.3 New per-epoch indices (100 features)
Computed by planetscope_10epoch_obia_v3.py using the full SuperDove band set. 10 new indices per epoch × 10 epochs = 100 features:
| Index | Formula | Sensitivity |
|---|---|---|
| NDRE | (NIR − RedEdge) / (NIR + RedEdge) | Canopy chlorophyll |
| CIre | NIR / RedEdge − 1 | Chlorophyll index |
| RENDVI | (RedEdge − Red) / (RedEdge + Red) | Early stress detection |
| SAVI | (NIR − Red)(1 + L) / (NIR + Red + L), L=0.5 | Bare-soil adjusted |
| OSAVI | (NIR − Red) / (NIR + Red + 0.16) | Open canopy |
| MSAVI2 | 0.5 × (2NIR + 1 − √((2NIR+1)² − 8(NIR−Red))) | Minimised soil effect |
| GNDVI | (NIR − Green) / (NIR + Green) | Canopy greenness |
| ARVI | (NIR − (2Red − Blue)) / (NIR + (2Red − Blue)) | Atmospheric resistant |
| VARI | (Green − Red) / (Green + Red − Blue) | Visible atmosphere resistant |
| BSI | ((Yellow + Red) − (NIR + Blue)) / ((Yellow + Red) + (NIR + Blue)) | Bare soil (no SWIR; Yellow substitutes) |
Results are cached as single-band GeoTIFFs in outputs_10epoch_obia_v3/feature_cache_v3/. First run: ~45 sec/epoch. Subsequent runs: instant cache hit.
3.4 Temporal features (28 features)
Derived across the 10-epoch time series for four curated indices: NDVI, EVI, NDRE, GNDVI.
3.4.1 Percentiles (12 features)
Per-pixel p10, p50, p90 across the 10 epochs. Robust to outliers from cloud shadow or missing acquisitions:
p10{index}, p50{index}, p90{index}3.4.2 Harmonic fit (12 features)
Fits a one-term Fourier model per pixel:
y = a0 + a1·cos(2π·DOY/365) + b1·sin(2π·DOY/365)Three derived features per index: amplitude (√(a1² + b1²)), phase (arctan2(b1, a1)), and offset (a0). The amplitude captures the magnitude of seasonal greenness variation; the offset is the mean greenness level.
harmAmp{index}, harmPhase{index}, harmOffset{index}harmOffsetNDVI ranked #10 and ampNDVI ranked #4 in the top-15 feature importance list — showing that seasonality structure is more informative than any single-date observation.
3.4.3 Year-over-year delta (4 features)
March 2024 → March 2026 pixel-wise difference for each curated index. Captures land-cover change and phenological shift over the two-year observation window:
yoy{index}_mar26_mar243.5 Texture features (6 segment-level)
Computed on the temporal-median NDVI and NIR images (not per-epoch, to avoid dominance by any single date):
| Feature | Description |
|---|---|
entropyNDVI__seg |
Shannon entropy of the NDVI histogram within the segment |
entropyNIR__seg |
Shannon entropy of the NIR histogram within the segment |
rangeNDVI__seg |
p90 − p10 range of NDVI pixels within the segment |
rangeNIR__seg |
p90 − p10 range of NIR pixels within the segment |
cvNDVI__seg |
Coefficient of variation of NDVI within the segment |
madNDVI__seg |
Mean absolute deviation of NDVI within the segment |
High entropy and range indicate heterogeneous vegetation (e.g. agroforest edges), while low values indicate uniform canopy (e.g. paddy, water).
3.5.1 Note on Haralick GLCM textures
Haralick GLCM textures (OTB otbcli_HaralickTextureExtraction) were tested on 2026-05-03 and disabled. The hypothesis was that planting-row periodicity in production forest would appear as high ASM/energy at a 1-pixel (~3 m) offset. In practice, the temporal-median composite smoothed away the spatial pattern, and the 32 extra features dropped L2 CV OA from 0.65 to 0.61 (curse of dimensionality at 97 training rows). Re-enabling is gated by cfg.haralick_enable = True.
3.6 Shape features (5 segment-level)
Derived from the segment label raster:
| Feature | Description |
|---|---|
area_m2__seg |
Segment area in square metres |
perimeter_m__seg |
Perimeter length (4-connected edge counting) |
compactness__seg |
4π × area / perimeter² |
elongation__seg |
Bounding-box major / minor axis ratio |
fill_ratio__seg |
Area / bounding-box area |
3.7 Zonal statistics
All pixel features are summarised per segment via zonal mean and standard deviation — two segment features per pixel feature. The standard deviation captures within-segment variability (e.g. a heterogeneous segment with mixed vegetation will have high std across spectral bands).
Segments with fewer than 4 in-AOI pixels are flagged as boundary slivers and dropped from training.
3.7.1 Median imputation
Some PlanetScope epochs have a 1–2 pixel ragged edge in certain bands (observed: sep25_b1, nov25_b1). Rather than dropping all segments that touch this edge, v3 median-imputes partial-NaN features:
non_sliver = ~seg_df[feature_cols].isna().all(axis=1)
medians = seg_df.loc[non_sliver, feature_cols].median()
seg_df.loc[non_sliver, feature_cols] = (
seg_df.loc[non_sliver, feature_cols].fillna(medians)
)This recovered 13 labeled segments in the 2026-05-03 run, including 6 of 6 newly-added Production Forest samples that sat on the edge.