← RTM-Suite home

Model comparison & verification

RTM-Suite ships every model twice — once in R, once in Python. This page runs independent simulations from shared, randomly-sampled Look-Up Tables through both language implementations and compares them directly, plus one section that validates a model against real measured data rather than the other language. Coverage spans leaf optics, canopy radiative transfer, soil (dry-soil mixing and wet-soil MARMIT), the full soil-plant-atmosphere chain (SPART), and the full coupled SCOPE 2.1 model — the same robustness checks any user could re-run themselves, with the scripts linked in each section.

R vs PythonLeaf × canopy models — all 15 combinations

Every leaf model (PROSPECT-D, PROSPECT-PRO, Liberty, Fluspect-B, Fluspect-B-Cx) run through every canopy model (fourSAIL, foursail2, INFORM) — 15 combinations, 100 simulations each from an independent randomly-sampled LUT per combination, full reflectance spectrum compared wavelength-by-wavelength. Liberty combinations show a small fraction of rows excluded on one side (89–93/100 valid) where an extreme random parameter draw pushed the iterative Liberty solver out of its valid domain in one language before the other — every row that did converge on both sides agrees to Liberty's own precision (RMSE ≈ 10^-12, vs ≈10^-16 for the other four leaf models).

R vs PythonSoil & atmosphere — SPART, MARMIT-1, MARMIT-2

Beyond leaf/canopy optics: the full soil-plant-atmosphere chain (SPART, TOC and TOA reflectance for a real sensor) and wet-soil reflectance (MARMIT, both published versions).

SPART — soil-plant-atmosphere, Sentinel-2A bands

100 simulations, one shared LUT (getLUT(inputsSPART), seed 21), fixed realistic sea-level clear-sky atmosphere (matching the tutorial's own choice to avoid getLUT()'s known unphysical-atmosphere edge case at its widest default ranges). Both TOC canopy-BRDF reflectance and TOA (after SMAC atmospheric correction) compared, all 13 Sentinel-2A bands.

MARMIT — wet-soil reflectance, both model versions

60 random draws (soil ID 1–17 from the bundled Bablet_2016 database, water-layer thickness L 0.01–0.15 cm, wet-surface fraction eps 0.1–0.9), run through both MARMIT-1 (Bablet et al. 2018, the simpler original model) and MARMIT-2 (adds soil-particle refractive-index/size effects) in both languages.

R vs PythonFull SCOPE 2.1

600 SCOPE simulations, one shared LUT (getLUT.SCOPE(), seed 7), default options — the full coupled soil/canopy/photosynthesis/energy-balance/fluorescence pipeline, not just radiative transfer. Top-of-canopy reflectance compared across 2,053 of 2,162 spectral-domain wavelengths (the remaining ~109 are a structural, non-contiguous gap matching atmospheric water-vapor absorption windows around 1400/1900/2500 nm — excluded from both sides rather than dropping otherwise-valid simulations).

Real data not R vs PythonPROSPECT-D vs. 308 real measured leaves (Angers database)

Everything above checks that R and Python agree with each other — strong evidence the Python port is faithful, but not by itself evidence that either implementation matches reality. This section instead validates PROSPECT-D against the bundled Angers leaf-optics database (308 real leaves with lab-measured chlorophyll a+b, carotenoids, anthocyanin, equivalent water thickness and leaf mass per area, paired with lab-measured reflectance and transmittance spectra) — the same kind of dataset PROSPECT itself was calibrated and validated against in the original literature. Measured biochemical traits go in as-is; the mesophyll structure parameter N (not independently measurable) is fit per leaf over a 1.0–3.0 grid, the standard PROSPECT calibration procedure.

Liberty, Fluspect and INFORM don't have an equivalent bundled field/lab dataset in this repo, so this real-data check currently covers PROSPECT-D only (the leaf model both the Angers database and the wider PROSPECT literature target directly).

R vs PythonPROSPECT-D + fourSAIL, 600-simulation deep dive

A larger single-combination run for extra statistical depth on the most commonly used pairing: 600 simulations, one shared LUT (getLUT(), seed 42), full 400–2500 nm reflectance spectrum compared wavelength-by-wavelength.

Reproduce this yourself

Each comparison is two scripts under Scripts/Comparison/: an R script builds the shared LUT and the R-side reference output, and a companion Python script reads that same LUT, runs the Python port, and produces the figure + error stats above.

This is a large-sample statistical comparison, complementary to the existing per-function precision verification (single test cases, checked to floating-point precision, 50 automated tests).