Validation: Sensor convolution

Recorded evidence: 5 variations; 5 passed, 0 failed, and 0 skipped (100.0% pass rate over all recorded variations).

Evidence boundary

The current checked-in campaign uses small synthetic or already-present inputs and does not contact NEON. It validates software contracts and diagnostics, not real-flightline scientific accuracy.

What this module test exercises

Compare spectral convolution with an independently calculated weighted average while varying source and target band counts.

Implementation exercised: resample_chunk_to_sensor

Inputs varied

Field	Why it is recorded
input_shape_y_x_b	Varies spatial size and source wavelength count.
target_band_count	Varies the number of target spectral responses.

Checks and how to interpret them

Check	Question	PASS means	If it does not pass
output_band_count_correct	Does output contain one band per supplied response function?	The final axis equals the target response count.	Inspect response iteration and output allocation.
weighted_average_matches_reference	Does convolution match an independent normalized-weight calculation?	Maximum absolute difference is at most 2e-7.	Inspect response normalization, wavelength alignment, and axis order.
dtype_is_float32	Does convolution retain the expected compact datatype?	Output dtype is float32.	Review NumPy promotion and memory cost.

Diagnostics recorded for every variation

Field	Why it is recorded
output_shape	Observed target cube dimensions.
max_absolute_error	Difference from the independent reference.
output_min	Minimum convolved reflectance in the fixture.
output_max	Maximum convolved reflectance in the fixture.

Input variations and results

On narrow screens, scroll the table horizontally to see every diagnostic and check.

Variation	Input variation	Result	Diagnostics	Explicit checks
sensor_convolution-001 Resample 4 source bands into 1 target bands.	input_shape_y_x_b=[3,4,4]; target_band_count=1	PASS	max_absolute_error=2.98023e-08; output_max=0.793138; output_min=0.253127; output_shape=[3,4,1]	dtype_is_float32=✓; output_band_count_correct=✓; weighted_average_matches_reference=✓
sensor_convolution-002 Resample 5 source bands into 2 target bands.	input_shape_y_x_b=[4,5,5]; target_band_count=2	PASS	max_absolute_error=5.96046e-08; output_max=0.824112; output_min=0.138205; output_shape=[4,5,2]	dtype_is_float32=✓; output_band_count_correct=✓; weighted_average_matches_reference=✓
sensor_convolution-003 Resample 6 source bands into 3 target bands.	input_shape_y_x_b=[5,4,6]; target_band_count=3	PASS	max_absolute_error=1.19209e-07; output_max=0.71096; output_min=0.121563; output_shape=[5,4,3]	dtype_is_float32=✓; output_band_count_correct=✓; weighted_average_matches_reference=✓
sensor_convolution-004 Resample 7 source bands into 4 target bands.	input_shape_y_x_b=[3,5,7]; target_band_count=4	PASS	max_absolute_error=1.19209e-07; output_max=0.764092; output_min=0.219115; output_shape=[3,5,4]	dtype_is_float32=✓; output_band_count_correct=✓; weighted_average_matches_reference=✓
sensor_convolution-005 Resample 8 source bands into 5 target bands.	input_shape_y_x_b=[4,4,8]; target_band_count=5	PASS	max_absolute_error=5.96046e-08; output_max=0.768413; output_min=0.313292; output_shape=[4,4,5]	dtype_is_float32=✓; output_band_count_correct=✓; weighted_average_matches_reference=✓

What a passing result establishes

Band-count, weighting, precision, and dtype contracts.

What it does not establish

Scientific adequacy of a particular sensor response curve.

The matching real stage checks are explained in the stage QA test guide.

Example from the real R10C test run

The real stage uses the same spatial/spectral support diagnostics as other reflectance products.

Configured and fitted brightness adjustments overlap; this verifies application, not scientific optimality of the coefficients.

The figure is evidence from one completed flightline, not a replacement for the variation table above. Open the real flightline walkthrough for exact values and limitations.

Expansion to 100 real variations

The repository includes a live 100-flightline campaign specification. It requires a pinned inventory of real flightline IDs plus an explicit compute, storage, and network allocation. Live results must be stored as a new campaign record; they must not overwrite this offline baseline.

Reproduce or expand this module

# Fast local evidence matrix (five variations per module)
python scripts/run_validation_campaign.py --iterations-per-module 5

# Exercise 100 deterministic small-data variations per module
python scripts/run_validation_campaign.py --iterations-per-module 100 \
  --output validation/results/offline-contract-100.json

python scripts/generate_validation_docs.py

The 100-case offline command scales contract variation and randomized synthetic inputs. It does not substitute for 100 distinct NEON downloads.

Last updated: 2026-08-14