Agricultural Digital Twin
LSTM forecasting of 10 m Sentinel-2 NDVI from CHIRPS rainfall and ESA CCI soil moisture; R² above 0.50 and MAE below 0.07 at parcel level.
Overview
I apply Earth observation to concrete environmental questions: what is growing where, how vegetation responds to rainfall, where forest carbon is heading and how fast a coastline is retreating. Every answer is fused from several sensors and validated before it is reported.
What you get
- Crop-type and crop-condition maps at 10 m
- Field-level vegetation forecasts driven by rainfall and soil moisture
- Forest biomass and carbon forecasts under climate scenarios for REDD+
- Shoreline-change forecasts with uncertainty
- Land-degradation and mining-impact assessments with clear, quotable figures
- Reproducible notebooks and technical reports
Who this helps
Agricultural programmes & supply chains
Including cocoa and smallholder systems in West Africa.
Forest-carbon & REDD+ projects
Needing biomass trends and forward-looking scenarios.
Coastal authorities
Monitoring erosion hotspots before infrastructure is lost.
NGOs proving impact
Monitoring land and forests with satellite data and participatory ground truth.
How I approach it
-
01
Frame the environmental indicator
Agree what is measured (NDVI trend, crop type, biomass, shoreline position), at what scale and for which decision.
-
02
Fuse complementary sources
Sentinel-1 and -2, Landsat, GEDI LiDAR, SMAP, ERA5, CHIRPS, ESA CCI soil moisture and UAV orthomosaics, each covering another's blind spots.
-
03
Ground it in the field
Participatory plot-level ground truth, UAV-derived labels and DSAS transects anchor the models to reality.
-
04
Model change over time
LSTMs, Bayesian LSTMs and temporal graph neural networks capture lagged responses and trends; scenario runs extend them under RCP pathways.
-
05
Validate spatially and quantify uncertainty
Spatial-block cross-validation and held-out areas, with uncertainty reported alongside every forecast.
-
06
Translate into decisions
Clear figures and hotspot maps, such as "32.0% of the district in measurable decline", that stakeholders can use directly.
Evidence from my work
- 32.0%Of Obuasi district in measurable vegetation decline (2018–2024)
- R² > 0.50Parcel-level NDVI forecasts, MAE below 0.07
- < 8 mMAE for 3-month shoreline forecasts
- 333.9 km²Under automated change detection for Rainforest Alliance
Figures come from the Obuasi, Agricultural Digital Twin and Coastal Erosion projects and from Rainforest Alliance work in 2024.
Selected work
View case study
SMART-FOREST
Fuses Sentinel-1/2, GEDI LiDAR, SMAP and ERA5 in object-based and temporal GNNs to forecast above-ground biomass over 10-year horizons under RCP scenarios.
Coastal Erosion Digital Twin
3-month shoreline forecasts (MAE below 8 m) from Sentinel-1/2, Landsat, UAV orthomosaics and wave climatology, validated against DSAS transects.
Obuasi Vegetation Decline
Boundary-clipped USGS Landsat NDVI trend: 32.0% of the district in measurable decline (9.2% strong, 22.8% moderate), concentrated around mining.
What sets my work apart
- 1
Multi-sensor by default
Radar, optical, LiDAR and climate data together, rather than a single index.
- 2
Local landscapes, real examples
Built on West African systems such as cocoa farms, the Atewa range and Lake Volta shorelines.
- 3
Forecasts with error bars
Forward-looking products always carry their uncertainty.