FForge Outcome Delivery
SAR Multi-Crop Acreage Estimator

Outcome Contract

← Back to outcome

Outcome Contract — SAR Multi-Crop Acreage Estimator (Round 1)

Request

Build the Round-1 solution for a SAR-based agricultural-intelligence challenge: *Sown Area Progression & Multi-Crop Acreage Estimation*. Using multi-temporal X-band SAR imagery across the Kharif season, estimate the cultivated area (in hectares) of five crops — Rice, Cotton, Maize, Bajra, Groundnut — for every village in a test set. Submissions are scored by Mean Squared Error (MSE) across all crop-area columns and all villages. Output schema:

ID,Rice_ha,Cotton_ha,Maize_ha,Bajra_ha,Groundnut_ha

Scope & honest constraint

The official competition datasets (multi-temporal X-band SAR tiles, village boundary polygons, and the training/label tables) are **not present in this workspace**, and cannot be retrieved from the competition platform from here. Therefore the contracted, verifiable outcome is a **complete, runnable Round-1 pipeline** with a documented seam for the real data — not a leaderboard number.

What is delivered LIVE and verified:

  • A reusable SAR preprocessing workflow (calibration domain handling,

multi-temporal speckle filtering, multi-temporal stacking).

  • Agricultural-extent extraction from temporal backscatter dynamics.
  • A physically-matched multi-crop acreage estimator (ridge linear unmixing,

with an optional random-forest residual model).

  • Exact submission-file generation.
  • A real-data ingestion seam (src/ingest.mjs) that turns a per-village zonal

statistics table into model-ready stacks, plus the documented raster recipe.

  • End-to-end verification on a physically-motivated synthetic Kharif benchmark.

Success criteria (MUST PASS)

  1. node run.mjs trains and writes a valid data/submission.csv with the exact

required header and one row per village, all values numeric and ≥ 0.

  1. node verify.mjs passes 100% of its checks, asserting:
  • calibration (dB ↔ linear) integrity;
  • the temporal speckle filter reduces variance while preserving level;
  • the agricultural-extent score is monotone in temporal variability;
  • the model beats a naive column-mean baseline by a clear MSE margin;
  • all predictions obey physical area constraints (0 ≤ area ≤ village area);
  • the result is reproducible (seeded);
  • the submission schema is exact;
  • the real-data ingestion seam round-trips.
  1. Deterministic, dependency-free Node (no network, no native deps) so it runs

anywhere and is honestly re-runnable.

Out of scope (Round 1)

  • Crop condition, yield, and production forecasting (later rounds).
  • The GDAL/rasterio raster→zonal-stats step itself (documented as a seam; it is

the only stage requiring a geospatial stack).

Acceptance

Outcome is accepted when both MUST_PASS commands succeed on a clean checkout and the delivery package documents exactly how to swap in the official data to produce a real submission.