Source-linked AI summary
Estimating Crop Primary Productivity with Sentinel-2 and Landsat 8 using Machine Learning Methods Trained with Radiative Transfer Simulations
Aleksandra Wolanin, Gustau Camps-Valls, Luis Gómez-Chova, Gonzalo Mateo-García, Christiaan van der Tol, Yongguang Zhang, Luis Guanter
TL;DR
Existing GPP remote-sensing methods can oversimplify photosynthetic mechanisms or struggle with uncertain parameters and extrapolation across conditions. This paper combines SCOPE simulations, satellite and meteorological data, and machine learning to estimate crop GPP directly, achieving accurate results across C3 crops and environmental conditions. The approach supports high-spatiotemporal-resolution monitoring while remaining sensitive to clear-sky and preprocessing limitations.
Problem
Existing GPP-estimation approaches face oversimplified photosynthetic assumptions, uncertain process-model parameters, and limited extrapolation across conditions.
Method
The study trains machine-learning models on SCOPE radiative-transfer simulations and applies direct-GPP models to Sentinel-2, Landsat 8, and meteorological data.
Results
The final neural-network model estimated GPP at tested flux towers with r2 of 0.92 and RMSE of 1.38 gC d−1 m−2 across varied C3 crops and environmental conditions.
Takeaways & Limitations
The approach supports high-spatiotemporal-resolution crop monitoring with Sentinel-2 and Landsat 8 and can be extended to other satellites.
Takeaways & Limitations
Results are for clear-sky data, and clouds or cloud shadows not captured during training introduce errors in GPP models.
Abstract
from arXiv · showhide
Satellite remote sensing has been widely used in the last decades for agricultural applications, {both for assessing vegetation condition and for subsequent yield prediction.} Existing remote sensing-based methods to estimate gross primary productivity (GPP), which is an important variable to indicate crop photosynthetic function and stress, typically rely on empirical or semi-empirical approaches, which tend to over-simplify photosynthetic mechanisms. In this work, we take advantage of all parallel developments in mechanistic photosynthesis modeling and satellite data availability for advanced monitoring of crop productivity. In particular, we combine process-based modeling with the soil-canopy energy balance radiative transfer model (SCOPE) with Sentinel-2 {and Landsat 8} optical remote sensing data and machine learning methods in order to estimate crop GPP. Our model successfully estimates GPP across a variety of C3 crop types and environmental conditions even though it does not use any local information from the corresponding sites. This highlights its potential to map crop productivity from new satellite sensors at a global scale with the help of current Earth observation cloud computing platforms.
1. Introduction
The paper motivates a hybrid GPP-estimation approach because existing process-based, LUE, and statistical methods each face limitations in complexity, assumptions, extrapolation, or crop-specific performance. It combines radiative transfer modeling, satellite reflectance, meteorological information, and machine learning to estimate GPP directly across varied conditions.
- Motivation: GPP estimation supports monitoring crop status, potential yields, management practices, and geographic or meteorological effects on photosynthesis.GPP is defined as the amount of carbon fixed by plants through photosynthesis.
- Existing approaches: Process-based models describe photosynthetic mechanisms rigorously but require parameters that may be unavailable or highly uncertain.
- Existing approaches: LUE models simplify canopies and radiation–structure relationships, while their empirical relationships may not hold across broader scales or changing environmental conditions.
- Existing approaches: Statistical and machine-learning approaches are powerful but may lack capacity to extrapolate to different conditions and require sufficiently representative training data.
- Contribution: The proposed hybrid approach combines process-based radiative transfer models with Sentinel-2 reflectance and machine learning to predict GPP directly from spectral and meteorological information.Training uses modeled data rather than flux-tower GPP, enabling simulation across a broad range of conditions.
2. Materials and methods
The workflow creates a synthetic SCOPE dataset, evaluates alternative GPP-estimation strategies and machine-learning models, then applies a selected direct-GPP model to satellite and meteorological data.
- Workflow: The synthetic training dataset is generated with SCOPE before machine-learning models are trained.
- Candidate approaches: The study initially considers retrieving vegetation parameters and running SCOPE forward, retrieving fPAR for a LUE model, or estimating GPP directly.
- Workflow selection: Sensitivity analysis, relationships among fPAR, LUE, and GPP, and machine-learning retrieval tests guide selection of the final workflow.
- Application: The selected machine-learning GPP model is applied directly to satellite reflectance and meteorological data.
2.1. The SCOPE model
SCOPE is a one-dimensional radiative-transfer and energy-balance model that simulates canopy-observed spectra and water, heat, and carbon fluxes. The study uses its simulations to support alternative GPP-estimation workflows and machine-learning training.
- Model outputs: SCOPE calculates radiance from 0.4 to 50 µm and simulates water, heat, and carbon-dioxide fluxes above the canopy.
- Model structure: SCOPE integrates radiative transfer and energy-balance calculations for individual leaves and canopies.Leaf optics use Fluspect, canopy radiative transfer uses 4SAIL, and biochemical processes include separate C3 and C4 formulations.
- GPP workflows: The study uses SCOPE simulations to compare three GPP-estimation workflows.
- Simulation inputs: Input parameters for SCOPE simulations are varied under an assumed uniform distribution.
- Model adaptations: The model is customized to output GPP and to calculate Vcmax at 25°C as a function of chlorophyll concentration Cab.
2.2. Sentinel-2 data
Sentinel-2 provides multispectral observations spanning visible, near-infrared, and shortwave-infrared wavelengths, including red-edge bands relevant to chlorophyll retrieval. The study uses atmospherically corrected, visually selected relatively cloud-free images.
- Sensor characteristics: Sentinel-2 samples 13 spectral bands from the visible and near infrared through the shortwave infrared.
- Sensor characteristics: Its red-edge bands at 705 nm and 740 nm are important for chlorophyll-content retrieval.
- Observation data: Sentinel-2 spatial resolution ranges from 10 m to 60 m, with a 290 km field of view.
- Preprocessing: The study uses visually selected relatively cloud-free images from 2016–2017 and applies Sen2Cor atmospheric correction.
2.3. Landsat 8 data
Landsat 8/OLI provides atmospherically corrected, masked surface reflectance with 30 m resolution and a 16-day repeat cycle. The study uses these data alongside GLDAS 2.1 meteorological inputs for crop-productivity analysis.
- Landsat 8 characteristics: Landsat 8/OLI has spectral coverage broadly similar to Sentinel-2, shares six bands, but lacks equally dense vegetation red-edge coverage.The study uses Landsat 8 bands common with Sentinel-2 for model application.
- Landsat 8 characteristics: Landsat 8 images have a 16-day repeat cycle and 30 m resolution for the bands used in this study.
- Data preprocessing: The analysis uses atmospherically corrected Landsat 8/OLI surface reflectance from Google Earth Engine with cloud, shadow, water, snow, and saturation masking.
- Meteorological inputs: GLDAS 2.1 supplies meteorological land-surface states and fluxes and was selected partly because it is available through Google Earth Engine.
2.5. Flux tower sites
The feasibility test uses flux-tower GPP observations from four United States sites and one German site, focusing on soybeans and other C3 crops. Vegetation-index light-use-efficiency models provide comparison baselines.
- Flux-tower data: The feasibility test uses four United States flux-tower sites and one German site representing soybeans and other C3 crops.
- Flux-tower data: Site GPP observations were obtained for dates with Sentinel-2 observations during 2016 and 2017.
- Comparison models: Vegetation-index light-use-efficiency models using incident PAR were applied as comparison methods, selecting equations previously reported to perform best.
2.7. Global sensitivity analysis
The study uses global sensitivity analysis to identify which model inputs most influence numerical outputs. It applies PAWN, which evaluates the full output distribution rather than only output variance.
- Purpose: Global sensitivity analysis attributes variation in a numerical model’s output to variation in its inputs.It can evaluate relative input importance and identify influential variables affecting model outputs.
- PAWN analysis: The PAWN method quantifies parameter sensitivity using the entire model-output distribution through its cumulative distribution function.
- PAWN analysis: PAWN is applicable independently of output-distribution shape, unlike variance-based analysis that uses only output variance.This distinction matters when output distributions are multimodal or highly skewed.
2.8. Machine learning models
The machine-learning workflow trains regressions on SCOPE-generated synthetic data and applies them to Sentinel-2 or Landsat 8 reflectances with meteorological inputs. It evaluates indirect parameter-retrieval and direct-GPP strategies at field and daily scales.
- Model training: Machine-learning models learn flexible nonlinear mappings from inputs such as reflectances to outputs such as GPP using SCOPE-generated training data.
- Algorithm selection: Neural networks and random forests were chosen over Gaussian-process regression because their training and testing are more computationally efficient for global application.
- Training setups: Two training setups retrieve canopy or leaf parameters and fPAR from reflectances and solar zenith angle, or retrieve GPP directly.
- Training setups: Direct GPP retrieval uses reflectances, observation solar zenith angle, meteorological conditions, and the modeling-time solar zenith angle.
- Satellite inputs: Models use either all ten Sentinel-2 spectral land bands or the subset shared with Landsat 8.
- Satellite application: The trained GPP models are applied to satellite reflectances and GLDAS 2.1 meteorology, producing daily values by integrating four daytime points.
- Satellite application: Daily field GPP is calculated by averaging satellite pixels within each field for comparison with flux-tower measurements.
2.9. Modeling set-up
The modeling setup uses SCOPE simulations with selected vegetation, soil, and observation parameters to support global-sensitivity analysis and machine-learning training. Sampling assumptions, geometric simplifications, and additional low-LAI simulations define the training dataset.
- Parameter sampling: Fifteen leaf, canopy, and soil parameters were varied under independent uniform distributions, while other parameters remained constant.The uniform sampling was chosen for simplicity, generalization, and even performance across the input range.
- Simulation conditions: V 25 cmax was set to 100 µmol m−2 s−1 or varied as a function of Cab, while meteorological conditions remained constant for the sensitivity analysis.The constant value was selected based on prior soybean simulations and common estimates for C3 crops.
- Observation geometry: Observation geometry used constant nadir-like viewing assumptions, with a 0° observation zenith angle and 90° solar-observation azimuthal difference.The setup reflects the relatively narrow fields of view and quasi-nadir observations of Sentinel-2 and Landsat 8.
- Parameter sampling: LIDFa and LIDFb were represented through their sum and difference rather than sampled independently because their values are constrained by their absolute-value sum.This parameterization enforces the model’s geometric constraint on leaf-angle-distribution parameters.
- Training-data construction: Nonconvergent energy-balance simulations were excluded from machine-learning training, and an additional 10,700-simulation subset with LAI = 0.001 represented nearly bare-canopy conditions.The added subset addressed underrepresentation of cases with very small LAI.
3. Results and Discussion
SCOPE simulations show that GPP depends on multiple vegetation and canopy parameters, especially at high productivity, while direct neural-network GPP retrieval performs better than several intermediate-parameter retrievals. The model captures seasonal GPP dynamics and supports satellite-based estimation, but its evaluation is limited by uncertain parameter assumptions and clear-sky data.
- 3.1. Predictor variables of GPP: LAI was the most influential input for GPP, followed by Cab, while other parameters became more influential at high GPP.LAI had a total PAWN index of 0.97 and Cab 0.87 over the whole output range.
- 3.1. Predictor variables of GPP: For high GPP, canopy height, pigments, dry matter, and canopy geometry increasingly contributed to GPP variability.Canopy geometry alters sunlit and shaded leaf relationships, while pigments compete for absorbed light.
- 3.2. Relationships between APAR and GPP: SCOPE relationships between GPP and APAR, APARCab, or canopy chlorophyll were nonlinear and scattered across broad parameter combinations, but became more linear when variability was constrained.The result supports the conclusion that LUE changes with leaf and canopy properties rather than depending only on meteorological conditions.
- 3.3. Training of vegetation parameters and GPP models: The GPP neural network had a mean error of 0.2 µCO2 s−1 m−2 across the complete GPP range and performed better than LAI and canopy chlorophyll retrievals.Its performance was similar to fPARCab retrieval, despite the difficulty of accurately retrieving several vegetation parameters.
- 3.4. Application to Sentinel-2: The approach is designed for global application because models can be retrained for new spectral settings and applied across satellites without local empirical information.The authors emphasize portability across instruments rather than direct superiority over empirical vegetation-index models.
- 3.4. Application to Sentinel-2: The model captured seasonal GPP magnitude and phenology across seasons, with r2 = 0.92, including emergence and senescence or harvest.However, these results were obtained for clear-sky data, and clouds were not accounted for in statistical training.
4. Conclusions and Future Work
The study developed a hybrid approach combining satellite remote sensing, machine learning, and detailed photosynthetic process modeling to monitor crop GPP. The approach produced accurate estimates across crop types and environmental conditions without site-specific information, supporting broader crop-productivity monitoring.
- The final neural-network model estimated GPP at tested flux towers with r2 of 0.92 and RMSE of 1.38 gC d−1 m−2.The model was evaluated across the tested flux towers.
- The proposed approach estimated GPP across varied crop types, environmental conditions, and periods without vegetation, without using additional local site information.This supports applying the method beyond the tested sites.
- The method enabled high spatio-temporal resolution crop monitoring with Sentinel-2 and Landsat 8 and could be extended to other satellite sensors.The authors identify global application through cloud-computing platforms as a future direction.
Appendix A.1. RF
The appendix compares random-forest GPP modeling with neural-network estimates and flux-tower observations. It evaluates time-series agreement, including dates affected by thin clouds, and summarizes RF performance with scatter-plot comparisons against daily flux-tower GPP.
- Random-forest GPP models were tested with settings for maximal depth, minimal samples per leaf, and sample weighting for small-LAI scenarios.Five RF models were selected for comparison with flux-tower data after synthetic-data testing.
- Figure A.12 compares flux-tower GPP with RF and neural-network time series, distinguishing clear-sky dates from dates affected by thin clouds.Straight lines denote clear-sky dates, while dotted lines denote thin-cloud dates not removed during visual checking.
- Figure A.13 plots daily flux-tower GPP against RF-estimated GPP, with a straight line marking the 1:1 relationship.The plot provides a direct visual comparison between measured and modeled daily GPP.