A Python-only workflow that downscales NASA ECOSTRESS land surface temperature (LST) data from 70 m to 10 m resolution using a Random Forest model. Sentinel-2 spectral bands and Copernicus DEM terrain variables serve as predictors, with residual correction applied to tie the sharpened output back to the original ECOSTRESS values. The pipeline has been tested across seven cities — Paris, Tunis, Kitchener-Waterloo, Toronto, Cairo, Madrid, and Vancouver — to check that it generalizes across different urban forms and climates.
This project is based on this NASA ARSET training module
For the complete project write-up, see: Full write-up
- Preprocessing — ECOSTRESS LST tiles are pulled via
earthaccess, masked for quality using bit-level QC decoding, water-masked, and cloud masked. Where a scene spans multiple MGRS tiles, mosaics are built with feathered blending and bias correction. - Predictors — Sentinel-2 surface reflectance bands and Copernicus DEM are resampled to match the target 10 m grid.
- Model — A Random Forest Regressor (
n_estimators=100,max_depth=15,min_samples_leaf=5) is trained to predict LST from the predictor stack. - Residual correction — Model output is reconciled with the coarse ECOSTRESS observations, correcting systematic bias introduced by the resolution gap.
- Evaluation — Feature importance (MDI and permutation-based) and diurnal temperature range (DTR) analysis are used to sanity-check what the model is learning and how it performs across sites.
The diagram below summarizes a simplified end-to-end downscaling pipeline, from ECOSTRESS data preprocessing and predictor preparation to model inference and heatmap outputs.
The figure below shows an example of LST sharpening for Paris during June 2026, illustrating the enhanced 10 m thermal detail produced by the workflow.
