Agro-Climatic Data Reconstruction with Machine Learning — BD-AgroClim
A reproducible research workflow comparing NASA POWER and ERA5-Land against 36 BMD stations and correcting daily agro-climatic reanalysis bias with machine learning.
The study evaluates spatial and seasonal reanalysis error, terrain controls, rainfall detection skill, and machine-learning residual correction under strict leave-one-station-out validation.
LightGBM was selected after model comparison, with SHAP used to interpret correction drivers and held-out island stations used to test spatial transferability.
What the project demonstrates.
NASA POWER + ERA5-Land evaluated against 36 BMD stations
Independently validated study period: 2000–2023
Maximum-temperature RMSE reduced from ~2.25°C to ~1.36°C under LOSO validation
SHAP interpretation and held-out island transferability checks
How the work was approached.
The challenge
Make coarse reanalysis products more useful at individual Bangladesh meteorological stations without overstating geographic transfer.
Technical approach
- Compare ERA5-Land and NASA POWER with meteorological reference observations.
- Model residual weather biases and evaluate with spatially separated validation.
Evidence & outcomes
- 36 BMD station locations studied.
- Reported maximum-temperature reduction from about 2.25°C to 1.36°C in the project evaluation.
Scope note: Numerical scores depend on the named validation configuration; new under-review manuscript experiments are separate.
Want to inspect the implementation? Explore the source repository ↗