Project aim

Forecast crop vigour at the level of individual fields, so farmers and advisers can see where growth is likely to fall behind before it shows in the field.

  • >0.50R² at parcel level
  • <0.07NDVI MAE
  • 10 mResolution
  • LSTMModel
Illustrative map for Agricultural Digital Twin Illustrative graphic

Overview

An LSTM forecasts 10 m Sentinel-2 NDVI from CHIRPS precipitation and ESA CCI soil moisture, with cloud gap-filling to keep the time series continuous. At field-parcel level it reaches R² above 0.50 and a mean absolute error below 0.07 under spatial cross-validation.

The challenge

Optical NDVI time series are full of cloud gaps in the tropics, and vegetation responds to rainfall with a lag that varies by crop and soil. A forecast has to work at parcel scale to be useful, not just as a regional average.

Data & inputs

  • Sentinel-2 NDVI (10 m)
  • CHIRPS precipitation
  • ESA CCI soil moisture
  • Field parcels

Approach

  1. 01

    Cloud gap-filling

    Sentinel-2 NDVI series are gap-filled to remove cloud-induced holes.

  2. 02

    Climate drivers

    CHIRPS precipitation and ESA CCI soil moisture are aligned to each parcel's time series.

  3. 03

    Sequence model

    An LSTM learns how NDVI responds to recent rainfall and moisture.

  4. 04

    Spatial cross-validation

    Evaluation holds out spatial blocks to avoid inflated scores.

Results

Forecasts reach R² above 0.50 and MAE below 0.07 at field-parcel level under spatial cross-validation.

Validation

Spatial cross-validation at field-parcel level.