LESS 3.0

GPU + CPU accelerated 3D radiative transfer model for remote sensing simulation

LESS 3.0 is a complete reimplementation of the LESS model — a widely-used 3D radiative transfer framework in quantitative remote sensing research. Built around NVIDIA OptiX, cross-vendor Vulkan Ray Query, Intel Embree, and a modern Python API, LESS 3.0 inherits the physical foundations of the original LESS while adding hardware acceleration, a redesigned scripting interface, and full physical coupling of energy balance, photosynthesis, and solar-induced chlorophyll fluorescence (SIF).

The original LESS remains available at lessrt.org. LESS 3.0 is the recommended path for new projects.

Key Features

• Three compute backends — OptiX is the high-performance NVIDIA path, Vulkan Ray Query provides experimental cross-vendor GPU coverage, and Embree supplies portable CPU workflows.
• Clean Python API — Designed for remote sensing scientists. Build a scene in 10 lines of code.
• Digital twin architecture — Build the scene once, simulate many times. Update illumination or surface properties without rebuilding geometry.
• Multi-sensor support — Optical, thermal infrared, LiDAR point cloud/waveform, SIF fluorescence, and passive/active microwave sensors.
• Full physical coupling — 3D radiation field → energy balance → photosynthesis → SIF, all tightly coupled to the Monte Carlo ray tracing engine.
• Interactive 3D viewer — scene.show() opens a browser-based viewer to inspect the scene and preview simulation results.

Installation

pip install less3d

Pre-built wheels for Windows and Linux are provided for Python 3.10–3.13. No compiler, CUDA toolkit, or additional runtime installation required.

Feature coverage differs by backend. Use less.can_use("feature") or scene.can_use("feature") for the final answer in the current installation.

Backend feature matrix

OK means the workflow is implemented by a non-experimental backend. Experimental means it is implemented and covered by cross-backend regression tests, but the Vulkan backend is still being optimized and may be substantially slower than OptiX. -- means that backend does not implement the workflow.

The table describes code coverage, not the hardware in a particular computer. For the single user-facing answer on the current machine, call for example:

less.can_use("lidar_waveform", backend="vulkan")
scene.can_use("energy_balance")

On NVIDIA hardware, use OptiX for production simulations. Vulkan is currently intended for cross-vendor access, validation, and backend development; it is not yet a performance replacement for the CUDA kernels.

Quick Start

import less

# Create a 10 m × 10 m bare-soil scene
scene = less.Scene()
scene.size = 10.0
scene.terrain = less.Terrain(property=less.Lambertian(reflectance=0.3))
scene.illumination = less.DirectDiffuse(sun_zenith=30, sun_azimuth=150)
scene.build()

# Simulate a nadir RGB image
image = scene.simulate(less.OpticalImager(
    less.Orthographic(image_size=256),
    bands=[650, 550, 450],   # R, G, B (nm)
    quality=128,
))
image.save("output.png")

Thanks to the digital twin architecture, updating scene parameters does not require a rebuild:

# Change sun angle and re-simulate — no rebuild needed
scene.illumination = less.DirectDiffuse(sun_zenith=60, sun_azimuth=90)
image2 = scene.simulate(sensor)

Simulation Capabilities

Forest Scene Example

import less
import numpy as np

scene = less.Scene()
scene.size = 30.0
scene.terrain = less.Terrain(property=less.Lambertian(reflectance=0.12))
scene.illumination = less.DirectDiffuse(sun_zenith=35, sun_azimuth=225)

# Broadleaf tree with PROSPECT leaf optical properties
birch = less.Object("birch", mesh="birch.obj")
birch.set_property("leaves", less.Prospect(cab=40, car=8, cw=0.012, cm=0.009, N=1.5))

# Random planting positions
xy = np.random.rand(50, 2) * 30
positions = np.c_[xy, np.zeros(50)]
scene.add(birch, positions=positions)
scene.build()

# Multispectral image (green, red, NIR)
image = scene.simulate(less.OpticalImager(
    less.Orthographic(image_size=512),
    bands=[550, 670, 800],
    quality=256,
))
image.save("forest.tif")   # GeoTIFF preserves physical radiance values

Built-in Examples and Viewer

import less

less.examples.list()                  # list all built-in scenes
less.examples.run("autumn_forest")    # launch interactive viewer
less.examples.run("maize_field")

scene.show()   # open 3D viewer in browser (Scene / Simulate / Products tabs)

Validation

LESS 3.0 is cross-validated against the original LESS on reference scenes:

Full validation reports are regenerated on each release. See tests/validation/.

Backend LiDAR throughput and CPU/GPU crossover can be reproduced with:

python benchmarks/benchmark_lidar_scaling.py

The benchmark reports fast point-cloud and full-waveform throughput at 1 thousand through 1 million pulses and checks numerical agreement against the Embree backend.

Documentation and Tutorials

Full documentation with step-by-step tutorials:

Topics covered: coordinate system · terrain and surface properties · illumination models · optical imaging · BRDF/BRF sampling · thermal infrared · LiDAR · energy balance · photosynthesis · SIF · visualization · scene serialization

Citing

If you use LESS 3.0 in your research, please cite the original LESS paper:

Qi, J., Xie, D., Yin, T., Yan, G., Gastellu-Etchegorry, J.P., Li, L., Zhang, W., Mu, X., Norford, L.K. (2019). LESS: LargE-Scale remote sensing data and image Simulation framework over heterogeneous 3D scenes. Remote Sensing of Environment, 221, 695–706.