Release Notes

v1.0.0 — SAR Multi-Crop Acreage Estimator (Round 1)

First release of the Round-1 pipeline for SAR-based multi-crop acreage estimation (Rice, Cotton, Maize, Bajra, Groundnut).

Added

• SAR preprocessing (src/sar.mjs): dB ↔ linear-power calibration handling,

multi-temporal Lee speckle filter, temporal stacking, and a rich per-village feature set (temporal statistics, phenology windows, cross/co ratios, and absolute backscattered-power features).

• Agricultural-extent extraction: temporal-variability cropland indicator.
• Estimator: physically-matched multi-output ridge linear unmixing

(closed-form, fast, deterministic), with an optional random-forest residual model (src/forest.mjs) for non-linear gains.

• Pipeline (src/pipeline.mjs): training, prediction, total-area

reconciliation, MSE/R²/baseline metrics, and exact submission generation.

• Synthetic Kharif benchmark (src/synth.mjs): per-crop seasonal

signatures, area-weighted linear-power mixing, realistic noise.

• Real-data ingestion seam (src/ingest.mjs): long-format zonal-stats CSV →

model-ready stacks; run.mjs --zonal … --labels ….

• Verification (verify.mjs): 11 screening checks, all passing.
• Season Explorer (public/tool.html): interactive forward/inverse SAR

visualizer.

Verified results (synthetic benchmark)

• MSE 15,434 ha² vs 37,725 ha² naive baseline (59.1% skill), pooled R² 0.589.
• Total cultivated-area recovery r = 0.897.
• Trains in ~56 ms; fully deterministic; zero runtime dependencies.

Known limitations

• No official leaderboard submission produced — competition data unavailable in

this workspace (see certification-report.md).

• The raster → zonal-statistics stage is documented but not bundled (needs a

geospatial stack).

• Crop condition / yield / production forecasting are later rounds, out of scope.