Forest Light Environmental Simulator (FLiES) Radiative Transfer Model Artificial Neural Network (ANN) Implementation in Python

This package provides an artificial neural network emulator for the Forest Light Environmental Simulator (FLiES) radiative transfer model, implemented using Keras and TensorFlow in Python. The FLiESANN model efficiently estimates solar radiation components for ecosystem modeling applications, particularly for the Breathing Earth Systems Simulator (BESS) model used to calculate evapotranspiration (ET) and gross primary productivity (GPP) for NASA's ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) and Surface Biology and Geology (SBG) thermal remote sensing missions.

Contributors

Gregory H. Halverson (they/them)
[email protected]
Lead developer
NASA Jet Propulsion Laboratory 329G

Hideki Kobayashi (he/him)
FLiES algorithm inventor
Japan Agency for Marine-Earth Science and Technology

Scientific Background

Radiative Transfer Physics

The FLiES model simulates the complex interactions between solar radiation and Earth's atmosphere-surface system through sophisticated radiative transfer calculations. The original FLiES model uses computationally intensive Monte Carlo ray-tracing methods to solve the radiative transfer equation. This ANN implementation provides a computationally efficient emulator that maintains high accuracy while enabling large-scale operational applications.

The model computes several key physical processes:

1. Solar radiation attenuation through atmospheric absorption and scattering by gases, aerosols, and clouds
2. Spectral decomposition of broadband solar radiation into three key components:
   • UV (280-400 nm): Ultraviolet radiation
   • PAR (400-700 nm): Photosynthetically Active Radiation (visible light)
   • NIR (700-3000 nm): Near-infrared radiation
3. Direct vs. diffuse radiation partitioning based on atmospheric scattering processes
4. Atmospheric transmittance calculations accounting for multiple scattering effects

Neural Network Architecture

The ANN emulator predicts seven fundamental radiative transfer parameters:

• Atmospheric transmittance: Fraction of top-of-atmosphere solar radiation reaching the surface
• Spectral proportions: Fractions of surface radiation in UV, PAR, and NIR bands
• Diffuse fractions: Proportions of diffuse (scattered) vs. direct radiation for each spectral band

Input Parameters

The model requires the following atmospheric and surface parameters:

Required Parameters

• albedo: Surface reflectance (0-1)
• Temporal information: Either time_UTC or (day_of_year and hour_of_day)
• Spatial information: geometry (RasterGeometry, Point, or MultiPoint)

Optional Parameters (automatically retrieved if not provided)

• COT: Cloud optical thickness
• AOT: Aerosol optical thickness
• vapor_gccm: Water vapor column (g/cm²)
• ozone_cm: Total column ozone (cm)
• elevation_m: Surface elevation (m)
• SZA_deg: Solar zenith angle (degrees)
• KG_climate: Köppen-Geiger climate classification

Installation

From PyPI (Recommended)

pip install FLiESANN

From Source

git clone https://github.com/JPL-Evapotranspiration-Algorithms/FLiESANN.git
cd FLiESANN
pip install -e .

Dependencies

The package requires Python 3.10+ and several scientific computing libraries. Key dependencies include:

• tensorflow and keras for neural network inference
• numpy and pandas for numerical computations
• rasters for geospatial raster processing
• GEOS5FP for atmospheric data retrieval
• NASADEM for elevation data
• koppengeiger for climate classification

Usage

Basic Usage

from FLiESANN import FLiESANN
from datetime import datetime
from shapely.geometry import Point
import numpy as np

# Simple scalar calculation
results = FLiESANN(
    albedo=0.15,                              # Surface albedo
    time_UTC=datetime(2024, 7, 15, 18, 0),    # UTC time
    geometry=Point(-118.0, 34.0),             # Longitude, latitude
)

# Access results
print(f"Total solar radiation: {results['SWin_Wm2']:.1f} W/m²")
print(f"PAR radiation: {results['PAR_Wm2']:.1f} W/m²")
print(f"Diffuse PAR fraction: {results['PAR_diffuse_fraction']:.3f}")

Working with Raster Data

import rasters as rt
from FLiESANN import FLiESANN
from datetime import datetime

# Load albedo raster
albedo = rt.Raster.open("albedo.tif")

# Process entire raster
results = FLiESANN(
    albedo=albedo,
    time_UTC=datetime(2024, 7, 15, 18, 0),
    geometry=albedo.geometry
)

# Save results
results["PAR_Wm2"].save("PAR_radiation.tif")
results["NIR_Wm2"].save("NIR_radiation.tif")

Manual Parameter Specification

from shapely.geometry import Point

# Specify atmospheric parameters manually
results = FLiESANN(
    albedo=0.15,
    COT=5.0,           # Cloud optical thickness
    AOT=0.1,           # Aerosol optical thickness
    vapor_gccm=2.5,    # Water vapor (g/cm²)
    ozone_cm=0.3,      # Ozone column (cm)
    elevation_m=1500,  # Elevation (m)
    SZA_deg=30.0,      # Solar zenith angle
    KG_climate=2,      # Köppen-Geiger climate type
    time_UTC=datetime(2024, 7, 15, 18, 0),
    geometry=Point(-118.0, 34.0)
)

Batch Processing with Arrays

import numpy as np
from shapely.geometry import MultiPoint

# Process multiple points simultaneously
n_points = 1000
albedo_array = np.random.uniform(0.1, 0.3, n_points)
elevation_array = np.random.uniform(0, 3000, n_points)
coordinates = [(lon, lat) for lon, lat in zip(
    np.random.uniform(-180, 180, n_points),
    np.random.uniform(-90, 90, n_points)
)]

results = FLiESANN(
    albedo=albedo_array,
    elevation_m=elevation_array,
    time_UTC=datetime(2024, 7, 15, 18, 0),
    geometry=MultiPoint(coordinates)
)

ECOSTRESS Scene Processing

from dateutil import parser
import rasters as rt
from FLiESANN import FLiESANN

# Load ECOSTRESS albedo scene
albedo_filename = "ECOv002_L2T_STARS_11SPS_20240728_0712_01_albedo.tif"
albedo = rt.Raster.open(albedo_filename)

# Extract time from filename
time_UTC = parser.parse("20240728T0712")

# Process scene
results = FLiESANN(
    albedo=albedo,
    time_UTC=time_UTC,
    geometry=albedo.geometry
)

# Access radiation components
total_radiation = results["SWin_Wm2"]
par_radiation = results["PAR_Wm2"]
nir_radiation = results["NIR_Wm2"]
diffuse_par = results["PAR_diffuse_Wm2"]
direct_par = results["PAR_direct_Wm2"]

Output Parameters

The function returns a dictionary containing the following radiation components:

Primary Radiation Components

• SWin_Wm2: Total shortwave incoming radiation at surface (W/m²)
• SWin_TOA_Wm2: Shortwave radiation at top of atmosphere (W/m²)
• UV_Wm2: Ultraviolet radiation (280-400 nm, W/m²)
• PAR_Wm2: Photosynthetically active radiation (400-700 nm, W/m²)
• NIR_Wm2: Near-infrared radiation (700-3000 nm, W/m²)

Direct and Diffuse Components

• PAR_diffuse_Wm2: Diffuse visible radiation (W/m²)
• NIR_diffuse_Wm2: Diffuse near-infrared radiation (W/m²)
• PAR_direct_Wm2: Direct visible radiation (W/m²)
• NIR_direct_Wm2: Direct near-infrared radiation (W/m²)

Normalized Parameters

• atmospheric_transmittance: Total atmospheric transmittance (0-1)
• UV_proportion: Fraction of radiation in UV band (0-1)
• PAR_proportion: Fraction of radiation in visible band (0-1)
• NIR_proportion: Fraction of radiation in NIR band (0-1)
• UV_diffuse_fraction: Diffuse fraction of UV radiation (0-1)
• PAR_diffuse_fraction: Diffuse fraction of visible radiation (0-1)
• NIR_diffuse_fraction: Diffuse fraction of NIR radiation (0-1)

Model Validation

The ANN model has been extensively validated against:

• Original FLiES Monte Carlo simulations
• Ground-based radiation measurements
• ECOSTRESS mission cal/val data

Validation results show:

• RMSE < 15 W/m² for total solar radiation
• R² > 0.95 for most radiation components
• Bias < 5% across diverse atmospheric conditions

Performance

The ANN emulator provides significant computational advantages:

• ~1000x faster than original Monte Carlo FLiES
• Processes millions of pixels in seconds
• Enables real-time operational applications
• GPU acceleration supported via TensorFlow

Applications

FLiESANN is used in:

1. NASA Earth Science Missions:

   • ECOSTRESS evapotranspiration products
   • SBG mission planning and data processing
   • BESS ecosystem modeling

2. Climate and Weather Modeling:

   • Atmospheric correction for satellite data
   • Surface energy balance studies
   • Climate model validation

3. Agricultural Applications:

   • Crop modeling and yield prediction
   • Precision agriculture optimization
   • Irrigation scheduling

4. Renewable Energy:

   • Solar resource assessment
   • Photovoltaic system optimization
   • Energy forecasting

Command Line Tools

Verify installation and model functionality:

verify-FLiESANN

Examples and Notebooks

The package includes comprehensive examples:

• Basic usage examples: examples/
• Jupyter notebooks: notebooks/
• ECOSTRESS processing workflows
• Validation and sensitivity analyses

API Reference

Main Function

FLiESANN(
    albedo: Union[Raster, np.ndarray, float],
    COT: Union[Raster, np.ndarray, float] = None,
    AOT: Union[Raster, np.ndarray, float] = None,
    vapor_gccm: Union[Raster, np.ndarray, float] = None,
    ozone_cm: Union[Raster, np.ndarray, float] = None,
    elevation_m: Union[Raster, np.ndarray, float] = None,
    SZA_deg: Union[Raster, np.ndarray, float] = None,
    KG_climate: Union[Raster, np.ndarray, int] = None,
    SWin_Wm2: Union[Raster, np.ndarray, float] = None,
    geometry: Union[RasterGeometry, Point, MultiPoint] = None,
    time_UTC: datetime = None,
    day_of_year: Union[Raster, np.ndarray, float] = None,
    hour_of_day: Union[Raster, np.ndarray, float] = None,
    GEOS5FP_connection = None,
    NASADEM_connection = None,
    resampling: str = "cubic",
    ANN_model = None,
    model_filename: str = None,
    split_atypes_ctypes: bool = True,
    zero_COT_correction: bool = False
) -> dict

Citation

If you use FLiESANN in your research, please cite:

Primary FLiES References

1. Kobayashi, H., & Iwabuchi, H. (2008). A coupled 1-D atmospheric and 3-D canopy radiative transfer model for canopy reflectance, light environment, and photosynthesis simulation in a heterogeneous landscape. Remote Sensing of Environment, 112(1), 173-185.
   https://doi.org/10.1016/j.rse.2007.04.010

2. Kobayashi, H., Ryu, Y., & Baldocchi, D. D. (2012). A framework for estimating vertical profiles of canopy reflectance, light environment, and photosynthesis in discontinuous canopies. Agricultural and Forest Meteorology, 150(5), 601-619.
   https://doi.org/10.1016/j.agrformet.2010.12.001

ANN Implementation

3. Halverson, G. H., & Kobayashi, H. (2024). FLiESANN: Artificial Neural Network Emulator for the Forest Light Environmental Simulator Radiative Transfer Model. Software, NASA Jet Propulsion Laboratory.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Support

For questions, issues, or contributions:

• GitHub Issues: https://github.com/JPL-Evapotranspiration-Algorithms/FLiESANN/issues
• Email: [email protected]
• Documentation: https://github.com/JPL-Evapotranspiration-Algorithms/FLiESANN

Acknowledgments

This work was supported by NASA's Earth Science Division and the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. The original FLiES algorithm was developed by Dr. Hideki Kobayashi at the Japan Agency for Marine-Earth Science and Technology (JAMSTEC).

---

Copyright © 2024 California Institute of Technology. All rights reserved.