Agricultural monitoring at fine field scales has gained substantial predictive reliability through a remote sensing framework that blends physical crop canopy simulations with aerial multispectral imagery. Conducted across dedicated experimental fields in Tongliao City within the western Songliao Plain of Inner Mongolia, the investigation demonstrated that calibrating the PROSAIL radiative transfer model for distinct maize developmental phases and accounting for unmanned aerial vehicle sensor bandpass profiles substantially elevated the precision of leaf area index retrieval. Leaf area index serves as a fundamental biophysical parameter for characterizing crop canopy architecture, tracking physiological development, and guiding precision management decisions throughout the growing season. While conventional in situ measurements remain severely constrained by intensive labor demands and limited spatial coverage, standard satellite observations frequently lack the spatial and temporal resolution required to manage individual plot variations. High-resolution multispectral imagery captured by unmanned aerial vehicles offers an adaptable alternative, yet empirical retrieval models consistently struggle when sparse ground sampling fails to capture dynamic structural and biochemical transitions between early vegetative growth and mature reproductive stages.
The central difficulty in retrieving maize canopy metrics across complete growing seasons stems from dramatic temporal shifts in crop geometry, leaf biochemical constituents, and underlying soil exposure. Data-driven machine learning models frequently establish statistical relationships linking ground observations to image-derived spectral reflectance and vegetation indices. Although these empirical techniques can achieve high accuracy when provided with extensive training samples, pooling all seasonal observations into a single overarching model inevitably obscures critical biophysical divergences that distinguish vegetative foliage from reproductive canopies. Conversely, attempting to train separate empirical models for individual measurement intervals often leads to severe sample deficits, as gathering sufficient field measurements at every key developmental phase is logistically and economically prohibitive. When training data lack adequate representativeness across varying cultivars, planting densities, and water or nutrient treatments, standard statistical models fail to generalize reliably.
Among available physical canopy formulations, the coupled PROSAIL model integrates the PROSPECT leaf optical model with the SAIL canopy radiative transfer model to simulate directional canopy reflectance across a continuous spectrum. By manipulating structural parameters, leaf chlorophyll and carotenoid concentrations, dry matter content, water thickness, observation geometry, and soil background conditions, PROSAIL generates vast synthetic datasets suitable for training machine learning algorithms. Nevertheless, the real-world utility of synthetic training spectra hinges entirely on how closely simulated conditions match actual field realities. Applying excessively broad, unconstrained parameter boundaries frequently produces physiologically implausible parameter combinations that depart from real field conditions, thereby widening discrepancies between simulated and measured spectra. Furthermore, because maize foliage properties and structural alignments shift markedly between early expansion and grain filling, applying static parameter ranges across the entire season undermines model fidelity. Constraining PROSAIL parameter spaces through Bayesian numerical optimization tailored to specific phenological stages provides a rigorous mechanism for preserving parameter interdependence while improving the representativeness of synthetic spectra under limited field sampling.
Physical consistency between simulated spectral libraries and real aerial observations also requires careful consideration of sensor hardware mechanics. Radiative transfer models typically calculate continuous canopy reflectance across a smooth wavelength spectrum. In contrast, aerial multispectral instruments record discrete spectral bands governed by specific internal optical filters and detector sensitivities. Standard modeling workflows often approximate sensor band reflectance by simply extracting simulated values at the nominal center wavelength of each band. This common simplification completely neglects spectral response variability across the bandpass width, introducing systematic discrepancies that degrade retrieval performance. To eliminate this source of physical mismatch, researchers incorporated the exact spectral response functions of the aerial sensor, convolving continuous simulated spectra across each bandpass profile. Weighted integration based on actual sensor response curves ensures that simulated band reflectances directly match the physical radiation captured by the multispectral camera, establishing an integrated simulation pipeline that bridges theoretical radiative transfer modeling and real-world unmanned aerial vehicle observations.
The empirical research evaluated this integrated approach using comprehensive agronomic field trials established on May 10, 2021, in Tongliao City. The experimental layout comprised an irrigation management trial alongside an independent nitrogen fertilizer optimization trial managed under local production practices with wide-narrow row spacing of 0.8 meters and 0.4 meters. The water management experiment included twenty-four distinct treatment combinations testing two maize varieties, Dika 159 and Zhengdan 958, planted at densities of six and nine plants per square meter across six irrigation gradients ranging from zero to 0.45 cubic meters per square meter in seventy-two square meter plots. The nitrogen optimization trial utilized cultivar Dika 159 at nine plants per square meter across fifteen nitrogen application rates spanning zero to 675 kilograms per hectare in 240 square meter plots. Ground measurements and aerial surveys were synchronized at four distinct developmental intervals: the sixth-leaf stage, twelfth-leaf stage, silking stage, and milk stage, capturing representative canopy transformations from early vegetative development through reproductive grain formation.
Aerial imagery was gathered using a DJI Inspire 2 unmanned aerial vehicle flown at seventy meters above ground level with forward and side overlaps fixed at eighty-five percent. The aircraft carried a Parrot Sequoia multispectral camera capturing green, red, red-edge, and near-infrared bands, accompanied by a Zenmuse X5S camera. Standardized radiometric calibration panels were imaged prior to each mission, and real-time kinematic global positioning established precise ground control coordinates across all plots. Orthomosaics produced via PhotoScan software were aggregated using circular sampling supports measuring 2.5 meters in diameter centered on georeferenced ground measurement locations to minimize positional sensitivity to individual plant rows and inter-row soil gaps. On the same dates, field crews quantified leaf area index using a LI-COR LAI-2200C plant canopy analyzer operated during early morning, sunset, or overcast conditions. A 270-degree view cap restricted surveyor interference during later growth stages, combining one above-canopy reading with four below-canopy measurements across wide and narrow rows to yield 276 total verified ground samples.
To establish a rigorous modeling foundation, the researchers conducted global Sobol sensitivity analysis using the Saltelli sampler within the SALib library, generating thousands of model evaluations to track parameter importance across developmental periods. The sensitivity results revealed striking differences: leaf area index dominated canopy reflectance variations in the red, red-edge, and near-infrared bands during vegetative growth, whereas leaf structural parameters and chlorophyll content exerted substantially greater relative influence during the reproductive phase. Based on these findings, key PROSAIL variables including the leaf structure coefficient, chlorophyll content, carotenoid content, equivalent water thickness, and dry matter content were calibrated using Markov chain Monte Carlo sampling implemented through the Differential Evolution Adaptive Metropolis algorithm in SPOTPY. Calibration utilized eighty percent of measured data at each stage while withholding twenty percent for independent model evaluation. Four parallel Markov chains achieved convergence with Gelman-Rubin diagnostics remaining below 1.10, yielding well-defined posterior parameter distributions across all four developmental stages.
The calibrated posterior parameter distributions revealed pronounced, stage-dependent trajectories reflecting the biological maturation of the maize crop. The posterior mean of the leaf structure parameter increased from 0.917 at the sixth-leaf stage to 1.692 at the twelfth-leaf stage, dropped slightly to 1.529 at silking, and culminated at 2.023 at the milk stage. Carotenoid content exhibited a continuous decline from 9.44 to 7.65 across the four growth periods, while dry matter content dropped from an initial 0.00745 to 0.00297 at the twelfth-leaf stage before stabilizing near 0.0045 in reproductive stages. Posterior correlation analyses demonstrated that most parameter couplings remained weak to moderate, with effective sample sizes ranging between 81.4 and 201.2. A moderate compensatory correlation emerged between leaf structure and dry matter at the silking stage, confirming that calibrated values represent effective parameter combinations that reproduce observed spectra rather than isolated physical measurements.
Synthetic training libraries generated by sampling joint posterior parameter distributions were subsequently merged with ground observations to train random forest, support vector regression, and back propagation neural network models. During the vegetative stage, models trained solely on simulated spectra achieved coefficients of determination between 0.86 and 0.93, while combining ground measurements with simulated data propelled the random forest model to an exceptional coefficient of determination of 0.97, alongside a root mean square error of 0.22. During the reproductive phase, when severe canopy closure and reflectance saturation typically degrade empirical retrieval, ground-only models plateaued at a coefficient of determination of 0.45. Introducing Markov chain Monte Carlo calibrated simulations increased the random forest coefficient of determination from 0.43 to 0.65 while reducing root mean square error from 0.49 to 0.38. Integrating sensor spectral response functions proved decisive, driving reproductive combined-dataset performance from 0.50 under simple center-wavelength sampling to 0.65, demonstrating the immense value of physical bandpass weighting under high canopy density.
Nevertheless, the study highlighted important methodological constraints inherent in current radiative transfer assumptions. The standard SAIL model presumes a turbid, horizontally homogeneous vegetation layer, which cannot fully replicate the distinct clumped row structure and anisotropic soil exposure present during early maize emergence. Furthermore, the absence of site-specific soil reflectance measurements required the use of standard library soil spectra, introducing potential background uncertainty. The authors emphasized that because both experimental trials were situated within a single geographic region under consistent meteorological conditions, extensive future testing across diverse soil types, climatic zones, multi-year timelines, and varied crop cultivars will be essential to establish the broad generalizability of sensor-weighted, stage-calibrated radiative transfer inversion workflows in precision agricultural management.
Subject of Research: Maize leaf area index estimation using stage-calibrated radiative transfer modeling and drone multispectral imagery
Article Title: Maize leaf area index retrieval using MCMC calibrated PROSAIL with UAV sensor spectral response weighting
Article References: Gao, X., Wang, Y., Wang, H., Yao, Y., Yang, S., Zhang, X., & Liu, Z. (2026). Maize leaf area index retrieval using MCMCcalibrated PROSAIL with UAV sensor spectral response weighting. Smart Agricultural Technology, 15, Article 102748. https://doi.org/10.1016/j.atech.2026.102748
Image Credits: AI Generated
DOI: 10.1016/j.atech.2026.102748
Keywords: Maize Leaf Area Index, PROSAIL Model, Multispectral Remote Sensing, Unmanned Aerial Vehicles, Markov Chain Monte Carlo, DREAM Algorithm, Spectral Response Functions, Sobol Sensitivity Analysis, Parrot Sequoia, Random Forest Regression, Crop Canopy Modeling, Precision Agriculture, Nitrogen Fertilizer Optimization, Agricultural Water Management, Vegetation Indices, Leaf Optical Properties, Radiative Transfer Inversion
News Source: Alan Morgan. (November 18, 2026). Sensor Weighted Crop Simulations Significantly Improve Drone Maize Leaf Area Retrieval. Scienmag.



