Wu et al. (2026) MA-UQNet: A multi-modal uncertainty quantification neural network for remote sensing-based wheat aboveground biomass estimation

Identification

• Journal: Artificial Intelligence in Agriculture
• Year: 2026
• Date: 2026-03-24
• Authors: Qiang Wu, Jiao Wang, Yuke Du, Dingyi Hou, Xiaoyu Cao, Xiaochun Wang, Hao Yang, Guijun Yang, Xinming Ma, Jinpeng Cheng
• DOI: 10.1016/j.aiia.2026.03.007

Research Groups

• State Key Laboratory of High-Efficiency Production of Wheat-Maize Double Cropping, College of Agronomy, Henan Agricultural University, Zhengzhou, China
• College of Information and Management Science, Henan Agricultural University, Zhengzhou, China
• Hohhot Agricultural and Animal Husbandry Technology Extension Center, Huhhot, China
• College of Life Sciences, Henan Agricultural University, Zhengzhou, China
• Key Laboratory of Quantitative Remote Sensing in Agriculture of Ministry of Agriculture and Rural Affairs, Information Technology Research Center, Beijing Academy of Agriculture and Forestry Sciences, Beijing, China
• College of Geological Engineering and Geomatics, Chang'an University, Xi'an, China

Short Summary

This study introduces MA-UQNet, a multi-modal deep learning framework for wheat aboveground biomass estimation that achieves superior prediction accuracy (R2 = 0.856) with well-calibrated uncertainty (97.18% coverage) by integrating adaptive multi-modal attention, growth stage-specific processing, and epistemic–aleatoric uncertainty decomposition.

Objective

• To develop a multi-modal data integration approach that leverages complementary information from hyperspectral remote sensing and environmental variables through adaptive fusion mechanisms.
• To systematically incorporate phenological knowledge into deep learning architectures to adapt feature extraction and integration strategies throughout wheat development stages.
• To quantify uncertainty by distinguishing between epistemic and aleatoric sources, enabling reliable operational predictions and actionable guidance for model improvement.

Study Configuration

• Spatial Scale: Field experiments at the National Precision Agriculture Research Center in Xiaotangshan, Changping District, Beijing, China (40.17°N, 116.43°E). Meteorological data from ERA5-Land reanalysis at 0.1° × 0.1° spatial resolution. Biomass sampled from 0.5 m² quadrats.
• Temporal Scale: Ten consecutive winter wheat growing seasons (2012–2022). Data collected at four key developmental stages: jointing, booting, flowering, and grain filling. Hourly meteorological data transformed into cumulative environmental variables.

Methodology and Data

• Models used:
  • MA-UQNet (Multi-modal Attention-based Uncertainty Quantification Network)
  • Components: Spectral Attention Module, Environmental Attention Module, Growth Stage-Specific Processing Module, Multi-Modal Fusion Module (bilinear attention), Uncertainty Quantification Module (heteroscedastic neural networks for aleatoric uncertainty, Monte Carlo Dropout for epistemic uncertainty).
  • Dimensionality reduction: Successive Projections Algorithm (SPA) for hyperspectral data.
  • Baselines for comparison: Random Forest, Bayesian Optimization-CNN-BiLSTM (BCBL), Multi-modal Gated Fusion (MMGF), TabPFN v2.5, Tabular Deep Learning model that makes Multiple predictions (TabM), Multimodal Transformer, Temporal Fusion Transformer (TFT), Deep Ensemble, Bayesian Neural Network, Quantile Regression.
• Data sources:
  • Remote Sensing: Hyperspectral reflectance (350–2500 nm) measured using an ASD FieldSpec Hi-Res spectroradiometer.
  • Environmental/Meteorological: ERA5-Land reanalysis dataset for hourly 2-meter air temperature (°C), total precipitation (mm), and surface solar radiation downward (J/m²).
  • Field Observations: Aboveground biomass (kg/ha) from destructive sampling.
  • Derived Variables: Effective Growing Degree Days (EGDD), Accumulated Precipitation (AP), Accumulated Solar Radiation (ASR), Photo-Thermal Units (PTU), Growth Days (GD).
  • Growth Stage Indicator: Zadoks (ZS) primary code (3, 4, 6, 7 for jointing, booting, flowering, grain filling).

Main Results

• MA-UQNet achieved superior performance on the independent test set with R2 = 0.856, RMSE = 1527.8 kg/ha, and a 95% prediction interval coverage rate of 97.18%.
• The model substantially outperformed Random Forest (R2 = 0.751, RMSE = 2008.0 kg/ha, coverage = 76.61%) and nine other representative baseline methods (R2 range: 0.763–0.821, coverage range: 18.95–94.76%).
• MA-UQNet demonstrated robust temporal generalization, with less than 4% degradation in R2 from training (0.888) to the independent test set (0.856) across a decade-spanning dataset.
• Uncertainty decomposition revealed that epistemic uncertainty contributed 53.27% of the total uncertainty, moderately dominating aleatoric uncertainty (46.73%), suggesting that strategic data collection is more impactful for uncertainty reduction than improving measurement precision.
• Ablation studies showed that the growth stage-specific processing module contributed the largest individual performance gain (7.7% R2 degradation upon removal), followed by spectral attention (5.8%) and environmental attention (4.8%).
• Multi-modal integration provided a 21.6% higher explained variance (R2) compared to spectral-only approaches (R2 = 0.704) and significantly outperformed environmental-only approaches (R2 = 0.441).
• Interpretability analysis highlighted key spectral bands (e.g., 2293 nm, 554 nm, 957 nm, 1098 nm, 1719 nm, 734 nm) and environmental variables (Effective Growing Degree Days, Accumulated Precipitation, Accumulated Solar Radiation) as primary drivers of biomass estimation. Spectral features accounted for 74.36% of total SHAP importance.

Contributions

• Development of MA-UQNet, a novel deep learning architecture that integrates multi-modal attention and growth stage-specific processing for enhanced wheat aboveground biomass estimation.
• Introduction of a hybrid uncertainty quantification framework that effectively decomposes total prediction uncertainty into epistemic and aleatoric components, providing well-calibrated confidence intervals (97.18% coverage).
• Demonstration of robust temporal generalization capabilities over a decade-spanning dataset, addressing a critical challenge for operational deployment of crop monitoring models.
• Insights into the relative importance of different data modalities and architectural components through comprehensive ablation studies, highlighting the critical role of phenological adaptation and multi-modal fusion.
• Agronomically validated interpretability and explainability analyses, linking model feature prioritization to established crop physiological principles, enhancing trustworthiness for precision agriculture applications.

Funding

• National Key Research and Development Program of China (2024YFD2301100)
• Special Project of Key Research and Development Program of Henan Province (251111111000)

Citation

@article{Wu2026MAUQNet,
  author = {Wu, Qiang and Wang, Jiao and Du, Yuke and Hou, Dingyi and Cao, Xiaoyu and Wang, Xiaochun and Yang, Hao and Yang, Guijun and Ma, Xinming and Cheng, Jinpeng},
  title = {MA-UQNet: A multi-modal uncertainty quantification neural network for remote sensing-based wheat aboveground biomass estimation},
  journal = {Artificial Intelligence in Agriculture},
  year = {2026},
  doi = {10.1016/j.aiia.2026.03.007},
  url = {https://doi.org/10.1016/j.aiia.2026.03.007}
}