Evaluating EC site representativeness through GPP reconstruction

This archive is a minimal working example (MWE) accompanying our study on deep-learning "bridge-model" reconstruction of terrestrial Gross Primary Productivity (GPP). The bridge model reconstructs a full spatial GPP field over the 24-state U.S. "Heartland" domain (0.05° resolution, 472 × 712 grid, 8-day time step) from the GPP observed at a single flux-tower pixel. Driving the model with one tower and evaluating the reconstruction at other locations quantifies how far that tower represents its surrounding region; the area reconstructed with R² > 0.95 defines the tower's representativeness footprint. To keep the archive compact while remaining fully reproducible, it provides one complete, runnable example — flux site US-MBP (obs000) with the MOD17 GPP product — including the trained model, the training/validation/test data, and the training and inference code. The full study covers 105 flux sites × 4 GPP products (MOD17, FluxSat, GOSIF, BESSv2); every other combination uses the identical code and network and differs only in the product field and the single-pixel site mask (masks for all 105 sites are included). Contents code/ — bridge-model definition, training (train_obs.py), single-site inference (infer_obs.py), and the preprocessing script that builds the model inputs. model/ — the trained bridge model for US-MBP × MOD17. data/ — the MOD17 inputs: normalized GPP field, valid-pixel mask, train/validation/test split, grid coordinates, and study-domain mask. obs_maps/ — single-pixel masks for all 105 sites. expected_output/ — a reference skill report to validate a run. Running python code/infer_obs.py 000 MOD17 reproduces the reported skill for this example (overall R² = 0.84; R² > 0.95 over 5.0 % of the domain). See README.md for full instructions and the software environment (environment.yml / requirements.txt). Input datasets (MOD17 GPP, FLUXNET/AmeriFlux tower GPP, NLDAS-2 meteorology, USDA Cropland Data Layer) are described and cited in the accompanying paper.