Installs into .claude/skills of the current project.
Are you the author of Harvest Detect?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/adilmunawar-harvest-detect)
---
name: harvest-detect
disable-model-invocation: true
argument-hint: "<timeseries.parquet> [crop]"
description: Detect harvest dates per parcel from NDVI (and S1) drops; produce a harvest layer and summary.
---
Input: parcel × date index table from `ndvi-timeseries`.
1. Per parcel on the smoothed series: find peak (mature plateau ≥ `ndvi_mature`, default 0.55 for sugarcane) sustained ≥ 20 days.
2. Harvest = first date after the plateau where NDVI drops by ≥ `drop` (default 0.25) within ≤ 15 days *and* stays below `post_max` (0.35) for ≥ 10 days. Optional S1: VH backscatter drop ≥ 2 dB in the same window raises confidence.
3. Confidence: high (both sensors), medium (NDVI only, clean series), low (gap > 20 days around the event). Flag parcels with no event as `unharvested` or `no-data`.
4. Outputs: `harvest_<aoi>_<season>.gpkg` (parcel id, harvest_date, confidence, peak_ndvi, drop), summary by district/week (count, area ha), and a Mermaid/Markdown table; hand to `zd-reports` for the client report.
5. Validate on any ground-truth dates available; report MAE in days.
