Going further: growth-phase mapping with S1+S2 fusion

Why radar alone struggles with phenology β€” and how optical fusion helps

🌐 English Β· Bahasa Indonesia β†’

The limitation of VH-only phase

The cycle counter in the planting index chapter tells you how many cycles occur, but classifying the exact growth stage (flooding β†’ vegetative β†’ reproductive β†’ maturity) from VH alone is hard: reproductive stages overlap in backscatter, and rainfed fields without a clear flooding trough lose the phenological anchor. On field data, a VH-only threshold phase model reached only ~32% agreement.

Per-pixel growth-phase map, MT 2024/25 (VH-based estimate).

Adding optical: Sentinel-1 + Sentinel-2 fusion (MOGPR)

Sentinel-2 NDVI carries strong phenological information but is interrupted by clouds in the humid tropics. MOGPR (Multi-Output Gaussian Process Regression, from the open-source FuseTS toolkit) fuses the two: it uses the cross-correlation between VH and NDVI to fill gaps and produce a single, gap-free fused curve per pixel, which a time-series classifier (MiniROCKET + LightGBM) then labels by phase.

In the project’s evaluation (leave-one-region-out, drone-labelled points), adding optical via MOGPR raised generative-phase F1 from 0.52 (VH-only) β†’ ~0.67 (fused) β€” about +15 points. Optical, it turns out, is the determinant of phenological stage; radar provides the all-weather backbone.

NoteOpen-source, not proprietary

FuseTS/MOGPR is a free, reproducible alternative to proprietary fusion services (e.g. CropSAR) β€” important for a national agency on a public budget.

Where the fusion code lives

The fusion pipeline (data cubes, MOGPR, phase classifier, multi-year phase/IP/production maps) is in a separate repository:

It requires Sentinel-2 access (Microsoft Planetary Computer) and more compute than this laptop tutorial, which is why it is presented here as an extension rather than a runnable step.

Recap

You have now seen the whole system:

  1. Paddy map β€” VH phenology + SMOTE classifier + multi-year consensus.
  2. Planting index β€” data-driven cycle counting (1Γ—/2Γ—/3Γ—).
  3. Irrigation performance β€” Kc + water balance β†’ SI/CU/RI per tertiary block.
  4. Growth phase β€” S1+S2 MOGPR fusion for accurate stage mapping (extension).

Together they turn free radar imagery into decision-ready information for irrigation management. Thank you for following along β€” issues and contributions are welcome on GitHub.