Continental-Scale Benchmark Transpiration Ratios Derived from Multi-Method Data Fusion across the NEON Network

Description This repository provides the data and Python code supporting an extended triple collocation (ETC)-based fusion of daily transpiration-to-evapotranspiration ratios (T/ET) across 47 sites of the National Ecological Observatory Network (NEON). Four categories of T/ET estimates are included: HFT: High-frequency Turbulence RSA: Remote Sensing Model GSP: Gridded Land Surface Product BIO: Biogeophysical Process Model ETC-based fusion framework The accompanying Python code implements an ETC framework to estimate the time-varying random errors of the four T/ET methods and combine them into a dynamically weighted fused product. The ETC system is reconstructed independently for each target day using a centered 30-day moving window, with at least 15 fully collocated daily observations required for estimation. This moving-window approach allows method-specific error variances, error covariance, fusion weights, fused T/ET, and fused uncertainty to vary with phenological and environmental conditions. HFT defines the common reference scale to which the other three estimates are rescaled, but it is not assumed to be error-free. The default ETC configuration permits correlated errors between HFT and BIO because both methods are derived partly from tower eddy-covariance observations and may share footprint and energy-balance-closure effects. The identifiability of the specified error-correlation structure is checked before the fusion is performed. For each valid moving window, the ETC equations are assembled as an overdetermined linear system and solved by ordinary least squares. If an ETC solution cannot be obtained, the program attempts a standard triple collocation solution without correlated errors. Method weights are subsequently calculated using best linear unbiased estimation based on the full error covariance matrix. Complete fusion dataset The principal file, NEON_TET_Fusion_47sites.csv, is a long-format dataset containing 85,654 daily site records and 16 variables. It includes: Daily fused T/ET; Estimated uncertainty of the fused T/ET; Daily T/ET estimates from HFT, RSA, GSP, and BIO; Dynamic fusion weights assigned to the four methods; and Method-specific random error variances. T/ET values, fusion weights, and fused uncertainties are dimensionless. Method-specific error variances have units of ((T/ET)^2). The four method weights sum to one within floating-point precision. Missing observations are represented by empty fields and were not filled during conversion of the fusion workbook to CSV. Figure-specific datasets The repository also contains processed source data used to reproduce Figures 2–7 and Figure 10 of the associated study: Figure2_Figure3_NEON_Site_Data.csv contains site-level error variances, error standard deviations, and mean fusion weights for all 47 sites. Error standard deviation is calculated as the square root of the temporally averaged error variance. Figure4_NEON_Fusion_Weights_by_Site_and_Vegetation.csv contains site coordinates, display coordinates, vegetation classifications, and temporally averaged method weights. Vegetation classes include evergreen forest, deciduous broadleaf forest, mixed forest, shrubland, grassland, and cropland. Figure5_Environmental_Drivers.csv contains site-level fusion weights and four environmental variables: mean annual precipitation, aridity index expressed as PET/P, mean air temperature, and leaf area index. These data support analysis of the environmental controls on method contributions. Figure6_TET_Consistency_and_Distributions.csv contains 85,654 time-resolved records used to compare the fused T/ET product with the four individual methods through scatter-density and probability-density distributions. Pairwise filtering provides 68,047 valid method–fusion observations for each comparison. Figure7_Data.csv contains 235 site–method records used to quantify reductions in error standard deviation and reproduce the Taylor diagram. The file includes absolute and relative error reductions, site-level and pooled correlations, normalized standard deviations, centered root-mean-square differences, and valid sample sizes. Figure10_Data.csv contains 1,527 records for three California sites—TEAK, SOAP, and SJER—used to evaluate T/ET responses during the 2021 drought. It includes raw and seven-day-smoothed daily trajectories, a non-drought baseline derived from 2019, 2020, 2022, and 2023, peak-season observations, and drought-suppression estimates. The corresponding peak-season suppression values are 8.9% for TEAK, 12.9% for SOAP, and −4.4% for SJER. Reproducibility and data handling The figure-specific datasets were recalculated directly from the complete fusion data rather than digitized from figure images or generated through simulation. Unless explicitly required to reproduce a figure, no temporal interpolation, smoothing, outlier removal, or missing-value imputation was applied. The centered seven-day moving averages are provided only for the Figure 10 daily trajectories. The Python implementation requires Python 3 together with numpy, pandas, scipy, and openpyxl. It also includes internal tests for ETC system structure, synthetic parameter recovery, and identifiability of the specified error-correlation structure. All CSV files are comma-delimited, use a period as the decimal separator, and retain missing numeric values as empty fields. Dates follow the ISO 8601 format (YYYY-MM-DD). Separate README files provide complete column definitions, calculation procedures, processing rules, figure-panel mappings, quality-control summaries, and instructions for running the ETC fusion code.