Source-linked AI summary
Retrieval of aboveground crop nitrogen content with a hybrid machine learning method
Katja Berger, Jochem Verrelst, Jean-Baptiste Féret, Tobias Hank, Matthias Wocher, Wolfram Mauser, Gustau Camps-Valls
TL;DR
Methods had not yet been applied for nitrogen retrieval from imaging spectroscopy data. The study used a hybrid workflow with GP algorithms trained on a PROSAIL-PRO-simulated database; the heteroscedastic GP performed best, while GP-BAT identified valuable band combinations for aboveground N estimation.
Problem
Methods had not yet been applied for N retrieval from imaging spectroscopy data, despite the importance of proper N management for crop growth and quality.
Method
The study used a hybrid retrieval workflow with two GP algorithms trained on a database simulated with the PROSAIL-PRO model.
Results
The heteroscedastic GP performed best for estimating crop leaves plus stalks, while GP-BAT identified relevant band combinations for aboveground Narea estimation.
Takeaways & Limitations
GP-BAT provides valuable band combinations for aboveground Narea estimation, and the heteroscedastic GP is the strongest reported estimator for crop leaves plus stalks.
Takeaways & Limitations
The hybrid workflow does not alleviate the limitations of radiative transfer models, and mature growth stages can limit solar-radiation penetration.
Abstract
from arXiv · showhide
Hyperspectral acquisitions have proven to be the most informative Earth observation data source for the estimation of nitrogen (N) content, which is the main limiting nutrient for plant growth and thus agricultural production. In the past, empirical algorithms have been widely employed to retrieve information on this biochemical plant component from canopy reflectance. However, these approaches do not seek for a cause-effect relationship based on physical laws. Moreover, most studies solely relied on the correlation of chlorophyll content with nitrogen, and thus neglected the fact that most N is bound in proteins. Our study presents a hybrid retrieval method using a physically-based approach combined with machine learning regression to estimate crop N content. Within the workflow, the leaf optical properties model PROSPECT-PRO including the newly calibrated specific absorption coefficients (SAC) of proteins, was coupled with the canopy reflectance model 4SAIL to PROSAIL-PRO. The latter was then employed to generate a training database to be used for advanced probabilistic machine learning methods: a standard homoscedastic Gaussian process (GP) and a heteroscedastic GP regression that accounts for signal-to-noise relations. Both GP models have the property of providing confidence intervals for the estimates, which sets them apart from other machine learners. GP-based band analysis identified optimal spectral settings with ten bands mainly situated in the shortwave infrared (SWIR) spectral region. Use of well-known protein absorption bands from the literature showed comparative results. Finally, the heteroscedastic GP model was successfully applied on airborne hyperspectral data for N mapping. We conclude that GP algorithms, and in particular the heteroscedastic GP, should be implemented for global agricultural monitoring of aboveground N from future imaging spectroscopy data.
3 TETIS, INRAE, AgroParisTech, CIRAD, CNRS, Université Montpellier, Montpellier, France
The study validated Gaussian-process models for aboveground crop N retrieval using field and airborne hyperspectral data. Both models performed accurately, with slightly better heteroscedastic-GP results for leaves plus stalks, while fruit N reduced performance.
- Validation results: Both Gaussian-process models provided accurate aboveground N simulations from hyperspectral data.Validation used corn and winter wheat spectra, with destructive measurements collected separately for leaves, stalks, and fruits.
- Validation results: 2.1 g/m² RMSE was obtained for heteroscedastic-GP retrieval against in situ N measurements from leaves plus stalks.The heteroscedastic model performed slightly better in model testing and against the leaves-plus-stalks measurements.
- Limitations: Including fruit N content deteriorated retrieval results because radiation could not penetrate thick stalk, cob, and ear tissues.The limitation was especially relevant when estimating total crop N beyond leaves plus stalks.
- Spectral analysis: Ten optimal bands were identified mainly in the shortwave infrared spectral region.Protein absorption bands from the literature produced comparative results.
- Airborne application: The heteroscedastic Gaussian-process model was successfully applied to airborne hyperspectral data for N mapping.The study concludes that Gaussian-process algorithms, particularly the heteroscedastic model, should support future imaging-spectroscopy monitoring of aboveground N.
1. Introduction
The introduction motivates hyperspectral N monitoring beyond chlorophyll-based proxies and proposes a physically informed hybrid workflow. It also frames spectral-band reduction and probabilistic Gaussian-process regression as responses to hyperspectral dimensionality and transferability challenges.
- Motivation: Crop nitrogen is important for plant growth, yield, quality, and sustainable soil–plant N management.The introduction presents N deficiency as reducing photosynthetic assimilation and crop yield per area.
- Motivation: Chlorophyll–N relationships are moderate across species and growth stages, with reported Pearson correlation r = 0.65 + 0.15.N translocation between vegetative and reproductive organs further weakens the relationship after vegetative growth.
- Motivation: Protein-sensitive hyperspectral measurements offer an alternative to chlorophyll-based N monitoring because protein signatures mainly occur in the SWIR region.The introduction emphasizes continuous hyperspectral coverage for capturing subtle protein-related signatures.
- Challenges: Hyperspectral data create collinearity, overfitting, computational-cost, and out-of-range prediction problems that motivate reducing the spectral data space.The text describes dimensionality reduction through feature transformation, selection, or extraction.
- Proposed approach: Hybrid retrieval combines radiative-transfer models with advanced regression methods so physical information and statistical learning are used together.The study applies PROSPECT-PRO coupled with 4SAIL and Gaussian processes to retrieve aboveground N independently from purely chlorophyll-based relations.
- Research gap: The study addresses the limited operational use of hybrid methods for N retrieval from imaging spectroscopy data.It compares spectral sampling strategies and explores a first step toward protein-related N sensing.
2.1. Study areas
The study combined field campaigns in Munich with airborne hyperspectral data from Barrax to characterize crop N across plant organs and growth stages. Measurements included canopy spectra, destructive biomass sampling, and laboratory nitrogen analysis.
- Munich-North-Isar study area: Field campaigns were conducted in 2017 and 2018 at the Munich-North-Isar test site in southern Bavaria.The campaigns covered winter wheat and corn during their growing seasons.
- Destructive measurements: Plant samples were separated into leaves, stalks, and fruits, oven-dried, and analyzed for nitrogen concentration using the Dumas combustion method.Organ-specific nitrogen concentration was converted to aboveground Narea by multiplying it by dry mass per unit ground area.
- Nitrogen variable: The study focused on aboveground Narea rather than leaf nitrogen concentration alone because upscaling requires dry-matter information.Narea is also described as correlated with leaf photosynthetic capacity and carbon fixation.
- Growth-stage dynamics: Nitrogen shifted from leaves and stalks toward wheat ears and corn cobs during senescence and flowering.Winter-wheat leaf and stalk Narea began decreasing at senescence, while corn N was transferred to cobs around flowering.
- Airborne data: Airborne HyMap data from Barrax, Spain, were used to illustrate N mapping over an agricultural area.The 126-band data covered 438–2483 nm and were resampled to the EnMAP spectral configuration.
2.2 Hybrid retrieval workflow for the estimation of aboveground N
The workflow combines PROSAIL-PRO radiative-transfer simulations with Gaussian-process regression to retrieve aboveground Narea from imaging spectroscopy, while providing predictive means and uncertainties. It uses GP-based band selection and supports application to airborne and satellite top-of-canopy reflectance data.
- Application: The retrieval workflow is designed for top-of-canopy imaging spectroscopy data, including airborne and satellite observations, and outputs Narea with predictive uncertainty.The reported workflow produces a predictive mean and predictive variance for each pixel.
- Optimal spectral band analysis: The GP-based band analysis tool selected optimal spectral subsets, with key protein-related bands concentrated in the SWIR region.Literature protein absorption bands provided a comparison for the selected spectral settings.
- Radiative transfer modeling: PROSPECT-PRO was coupled with 4SAIL to form PROSAIL-PRO, simulating canopy reflectance from biophysical and biochemical parameters.The modeled parameters include LAI, average leaf inclination angle, Cab, Cp, and leaf equivalent water thickness.
- Training database: A lookup table of 1’000 randomly generated PROSAIL-PRO parameter combinations supplied simulated EnMAP reflectances for GP training.Parameters were sampled from uniform distributions, except LAI and Cp, which used Gaussian distributions to produce a more realistic Narea distribution.
- Training database: Aboveground Narea was calculated from modeled protein content using a 4.43 conversion factor, then scaled by LAI from leaf to canopy ground-surface units.The workflow converts Cp to leaf Narea and multiplies by LAI and 10’000 for conversion from g/cm² to g/m².
- Machine learning regression: Standard homoscedastic and heteroscedastic GP models represent constant versus observation-specific noise variance, respectively.Both models infer predictive means and variances for each pixel, enabling spatially explicit estimates and uncertainty maps.
3. Results
GP-BAT identified a compact set of informative bands for aboveground Narea retrieval, concentrated mainly in the SWIR. Heteroscedastic GP retrievals performed best with leaves plus stalks validation, while organ-specific mismatches appeared for leaves-only and fruit-inclusive references.
- Spectral optimization: R^2_cv remained stable at 0.98 with SD 0.006 until fewer than 12 bands were used.
- Spectral optimization: Two selected bands at 1834 and 1854 nm were discarded because atmospheric water-vapor absorption prevents reliable hyperspectral measurement there.
- Spectral optimization: Using only one GP run ranked the 1762 nm band fifth of 235 but the 786 nm band 83rd, supporting iterative SBBR rather than single-run selection.
- Spectral optimization: GP-BAT selected most bands from SWIR 2, followed by SWIR 1, NIR, VIS, and red edge, which contributed least to Narea retrieval.
- Organ-specific retrieval: Leaves-plus-stalks validation performed well, whereas leaves-only validation overestimated Narea and fruit-inclusive validation underestimated it, especially at mature stages.
- Uncertainty and mapping: The heteroscedastic GP produced tighter confidence intervals for low estimates and wider intervals for high estimates, unlike the standard GP’s more uniform uncertainty.
- Uncertainty and mapping: The heteroscedastic GP was applied to a resampled EnMAP scene to produce aboveground Narea estimates and associated SD uncertainties, but validation was qualitative without flight-parallel N measurements.
4. Discussion
The discussion finds that both GP models learned the simulated-to-measured Narea relationship, with the heteroscedastic GP offering more differentiated uncertainty estimates. It also identifies organ visibility, simulated-data realism, and radiative-transfer assumptions as important boundaries on interpretation.
- Model performance: The heteroscedastic GP slightly outperformed the standard GP on independent test validation and gave the most stable in situ N validation.
- Model uncertainty: The heteroscedastic GP produced tighter low-value confidence intervals and a more differentiated predictive uncertainty pattern than the homoscedastic GP.
- Model performance: Both GP models showed close correspondence between model testing and in situ validation errors, supporting avoidance of under- and overfitting.
- Model performance: In situ validation achieved RMSE ≅2.2 g/m², compared with model-testing RMSE = 4.5 g/m², partly because simulated and measured Narea ranges differed.
- Organ-specific retrieval: Including fruits led to underestimation, with RMSE = 8.3–9 g/m², because thick cobs and ears contain nitrogen that reflected hyperspectral signals do not fully detect.
- Organ-specific retrieval: Leaves-plus-stalks validation produced RMSE = 2.1-2.3 g/m², while leaves-only validation produced RMSE = 5.5-5.8 g/m² with strong overestimation.
- Mapping and scope: The method’s mapped Narea values reached 18 g/m², but may mainly represent leaves plus stalks rather than total crop nitrogen.
- Hybrid-method limitations: The hybrid workflow retains radiative-transfer limitations, including the ill-posed inverse problem and constrained reproduction of measured canopy spectra.
5. Conclusions
The study presents a hybrid Gaussian-process and radiative-transfer workflow for retrieving aboveground crop nitrogen from imaging spectroscopy, with emphasis on uncertainty, computational efficiency, and transferable band selection. The authors recommend combining variable-sensitive bands with fast GP algorithms and RTMs for operational agricultural processing.
- Conclusions: The workflow is designed to support future imaging-spectroscopy applications requiring per-pixel uncertainty and high computational speed.These requirements are linked to assessing transferability across locations and times and to repetitive image processing.
- Conclusions: The hybrid workflow combines PROSAIL-PRO-simulated training data with two GP algorithms to retrieve aboveground crop N.The heteroscedastic GP performed best for estimating crop leaves-plus-stalks in situ N data.
- Conclusions: GP-based band analysis rapidly identifies relevant band combinations for aboveground Narea estimation.The selected settings improve estimation accuracy and increase processing speed relative to using all available bands.
- Conclusions: More comprehensive validation should apply models trained for aboveground Narea to real imaging-spectroscopy scenes across acquisition geometries, locations, crop types, and growth stages.Validation exercises should also include in situ data from different plant organs.
- Conclusions: The study recommends combining flexible GP algorithms with RTMs and variable-specific sensitive band settings for vegetation-property retrieval.The authors frame this combination as relevant to operational agricultural imaging-spectroscopy processing schemes.