The method

NDVI, stacked, smoothed, and asked one question.

Healthy paddy has a seasonal rhythm: flooded planting, vegetative growth, mature canopy, harvest. The Normalised Difference Vegetation Index reads that rhythm as a curve. Abandoned paddy goes flat. It is the absence of the curve, not its presence, that we look for.

1. 01Acquire

   Six years of satellite imagery, masked to the jelapang boundary.

   Monthly maximum-value NDVI composites are computed in Google Earth Engine from Sentinel-2 Level-2A scenes between 2019 and 2024. Each composite reduces the month down to the cleanest, highest-vegetation pixel: a robust proxy that survives Malaysia's cloud cover better than single-date imagery.

   The boundary mask is the official Jelapang Padi shapefile published by the Ministry of Agriculture. Anything outside the recognised jelapang is excluded; we are not estimating national paddy, we are measuring designated national rice infrastructure.

   Source

   COPERNICUS/S2_SR_HARMONIZED

   Reducer

   Monthly maximum NDVI

   Cloud handling

   SCL + QA60 bit-mask

   Boundary

   MoA Jelapang recognised polygons
2. 02Reduce

   One curve per cultivation unit, gap-filled and smoothed.

   The seventy-two composites collapse into a per-pixel time-series. Cloud gaps are filled with Savitzky–Golay smoothing across the temporal axis, which preserves the shape of the curve without inventing growth where none existed.

   The output of this stage is not a map. It is a vector for every pixel: a small EKG that says this is what this hectare looked like over six years. Maps come later; rhythm comes first.

   Smoothing

   Savitzky-Golay, window 5

   Gap-fill

   Temporal interpolation

   Output

   72-dim vector per pixel

   Storage

   Cloud-Optimised GeoTIFF
3. 03Classify

   K-Means separates the flat from the seasonal.

   Unsupervised K-Means clustering with k=2 is run on the per-pixel time-series. The classifier is given no prior labels; it is asked only to find the two natural shapes inside the data. The result is consistent: one cluster has the expected seasonal rhythm of active paddy, the other does not.

   The boundary between the two clusters is where policy lives. A pixel that drifts from active to inactive across the six-year window is the signal that triggers field validation.

   Algorithm

   K-Means, k=2, scikit-learn

   Distance

   Euclidean on z-normalised curves

   Cross-check

   Silhouette + Davies-Bouldin

   Stability

   Seed-averaged across 20 runs
4. 04Validate

   Field photos close the loop. Two-way by design.

   The web platform that hosts these maps is also a collection tool. MADA, KADA, and IADA officers, accredited enumerators, and trained farmers can submit geo-tagged photographs of any pixel the model flags. Each submission carries GPS, timestamp, observation type, and an optional camera bearing that becomes a directional arrow on the map.

   Validations enter an admin queue, are reviewed for quality, and are then fed back into the model as ground truth for the next iteration. The classifier is never finished; it learns from every visit.

   Frontend

   Leaflet + PMTiles, mobile-first

   Validators

   MADA · KADA · IADA · enumerators

   Quality gate

   Admin review queue

   Feedback loop

   Quarterly retraining