torchrtm 1.5.4

Torch-based Vegetation Radiative Transfer Model library (PROSPECT, SAIL, SMAC)

TorchRTM: A PyTorch-based Radiative Transfer Modeling Toolkit

TorchRTM is a GPU-accelerated, modular, and research-ready radiative transfer modeling (RTM) library built on top of PyTorch. It implements the full soil–leaf–canopy–atmosphere modelling chain (PROSPECT, 4SAIL, and SMAC) as native PyTorch tensor operations, enabling seamless integration with deep learning workflows.

This repository accompanies the paper:

TorchRTM: A high-performance vegetation RTM framework for deep learning integration

Installation

pip install torchrtm

Requires Python ≥ 3.9 and PyTorch ≥ 2.0.

Quick Start

Leaf Reflectance (PROSPECT)

import torch
from torchrtm.leaf.prospect import prospect5b, prospectd

# PROSPECT-5B: traits = [Cab, Car, Cbrown, Cw, Cm]
traits = torch.tensor([[40.0, 8.0, 0.0, 0.01, 0.009]])
N = torch.tensor([1.5])  # leaf structure parameter
rho, tau = prospect5b(traits, N)
# rho: leaf reflectance (1, 2101), tau: transmittance (1, 2101)

# PROSPECT-D: traits = [Cab, Car, Canth, Cbrown, Cw, Cm]
traits_d = torch.tensor([[40.0, 8.0, 0.0, 0.0, 0.01, 0.009]])
rho_d, tau_d = prospectd(traits_d, N)

Canopy Reflectance (PROSAIL)

from torchrtm.models.models import prosail

n = 100  # batch size
traits = torch.tensor([[40.0, 8.0, 0.0, 0.01, 0.009]] * n)
N       = torch.tensor([1.5] * n)
LIDFa   = torch.tensor([-0.35] * n)
LIDFb   = torch.tensor([-0.15] * n)
lai     = torch.tensor([3.0] * n)
q       = torch.tensor([0.01] * n)
tts     = torch.tensor([30.0] * n)  # solar zenith
tto     = torch.tensor([10.0] * n)  # view zenith
psi     = torch.tensor([0.0] * n)   # relative azimuth
alpha   = torch.tensor([40.0] * n)
psoil   = torch.tensor([1.0] * n)

result = prosail(traits, N, LIDFa, LIDFb, lai, q,
                 tts, tto, psi, alpha, psoil,
                 batch_size=n, prospect_type='prospect5b')
# result shape: (7, 2101, n) — 7 reflectance components

Atmospheric Correction (SMAC)

from torchrtm.atmosphere.smac import smac
from torchrtm.data_loader import load_smac_sensor

coefs, wl = load_smac_sensor("Sentinel2A-MSI")
tts = torch.tensor([30.0])
tto = torch.tensor([10.0])
psi = torch.tensor([0.0])
Ta_s, Ta_o, T_g, ra_dd, ra_so, ta_ss, ta_sd, ta_oo, ta_do = smac(tts, tto, psi, coefs)

LUT-based Retrieval

from torchrtm.retrival.fastLUT import Torchlut_pred

# xb: LUT spectra (N, D), y: LUT parameters (N, P), xq: query spectra (M, D)
preds = Torchlut_pred(xb, xq, y, k=5, device="cpu", agg="weighted")

Available Sensors (SMAC)

Sentinel2A-MSI, Sentinel2B-MSI, Sentinel3A-OLCI, Sentinel3B-OLCI, LANDSAT4-TM, LANDSAT5-TM, LANDSAT7-ETM, LANDSAT8-OLI, TerraAqua-MODIS

Testing

Install test dependencies and run:

pip install pytest
python -m pytest tests/test_torchrtm.py -v

The test suite covers data loaders, PROSPECT (5B/D/PRO), PROSAIL, SMAC, LUT retrieval, Inverse Net, math utilities, and full integration pipelines.

Supplementary Code

• Start_Tutorial/ — Reproduces the examples in the paper.

• Translation_Fidelity/ — Evaluates consistency of TorchRTM outputs against other RTM implementations.

Citation

If you use TorchRTM in your research, please cite:

Peng Sun, Peter van Bodegom, Marco Visser. TorchRTM: A high-performance vegetation RTM package for deep learning integration. Authorea. 15 January 2026.
DOI: https://doi.org/10.22541/au.176849838.80131044/v1

License

MIT