Zhao et al. (2025) Super-resolve satellite imagery to perform on par with UAV-borne hyperspectral imagery in predicting spring wheat physiological parameters using transformer models

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-11-17
• Authors: Jiangsan Zhao, Jakob Geipel
• DOI: 10.1016/j.compag.2025.111204

Research Groups

• Department of Agricultural Technology, Norwegian Institute of Bioeconomy Research (NIBIO), Ås, Norway

Short Summary

This study developed a deep learning-based super-resolution model to fuse UAV-borne RGB imagery with Sentinel-2 satellite data, creating high spatial and spectral resolution (HRS2S) images. These HRS2S images, combined with a novel transformer model (ResTrans21), accurately predicted spring wheat physiological parameters (dry matter, nitrogen content, nitrogen uptake), demonstrating performance comparable to costly UAV-borne hyperspectral imagery and superior to classical machine learning.

Objective

• To overcome spatial resolution limitations of Sentinel-2 satellite imagery for small-plot agricultural trials by generating plot-level satellite reflectance data through super-resolution (SR) reconstruction.
• To design and train a deep learning-based SR model to fuse high spatial resolution UAV-borne RGB images with low spatial resolution multispectral Sentinel-2 satellite images, producing high spatial and spectral resolution Sentinel-2 satellite (HRS2S) images.
• To develop and validate a composite transformer model, ResTrans21, for predicting spring wheat physiological parameters (dry matter, nitrogen content, and nitrogen uptake) using the generated HRS2S imagery.
• To compare the performance of ResTrans21 models based on HRS2S imagery against models trained on UAV-borne hyperspectral imagery and a classical Gaussian Process Regression (GPR) baseline.

Study Configuration

• Spatial Scale: Field trial area of 120 m × 100 m, divided into 60 plots of 20 m × 10 m. Remote sensing data were processed to a mutual ground sample distance of 0.1 m for analysis.
• Temporal Scale: Spring wheat sown on 18 May 2018. Sentinel-2 image acquired on 25 June 2018. UAV-borne imaging campaign and biomass sampling performed on 27 June 2018, corresponding to crop growth stages between BBCH 37 and BBCH 55 (mostly BBCH 39).

Methodology and Data

• Models used:
  • Super-resolution (SR) model: Deep learning-based, adapted from an image decomposition model, incorporating Dirichlet representation and non-negativity regularization.
  • Prediction model: ResTrans21, a composite transformer model combining a 2D Transformer (SimpleViT) and a custom 1DTransformer with skip connections, batch normalization, and dropout regularization.
  • Classical baseline: Gaussian Process Regression (GPR) with a squared exponential covariance.
• Data sources:
  • Satellite Imagery: Sentinel-2 Level 2A (L2A) (bottom of atmosphere reflectance), 8 VIS-NIR bands (B2, B3, B4, B5, B6, B7, B8, B8A), original 10 m spatial resolution.
  • UAV-borne Imagery:
    • Sony α5100 RGB camera (3 bands, 410–660 nm), original 0.01 m spatial resolution.
    • Rikola HSI (29 wavebands, 460–790 nm), original 0.04 m spatial resolution.
    • HySpex Mjolnir V-1240 (200 wavebands, 400–1000 nm), original 0.04 m spatial resolution.
  • Ground Truth Data: Dry matter (DM), nitrogen content (NC), and nitrogen uptake (NU) of spring wheat, obtained from 180 samples (three 1.5 m strips per sub-plot) harvested from 60 plots. DM determined gravimetrically, NC from crude protein content via NIR spectroscopy.
  • Ancillary Data: ASD FieldSpec 3 spectroradiometer for radiometric calibration, 0.5 m × 0.5 m 50% reflectance panel for validation.

Main Results

• Super-resolution (SR) Model Performance:
  • Generated HRS2S imagery preserved the spatial resolution of UAV-borne RGB (0.1 m) and spectral resolution of Sentinel-2.
  • Evaluation metrics for reconstructed Sentinel-2 (down-sampled to original resolution): Mean Absolute Difference (MAD) = 0.0049, Spectral Angle Mapper (SAM) = 0.77.
  • Band-wise linear regression (reconstructed vs. original Sentinel-2): Average R² = 0.82 and RMSE = 0.00 in the VIS region (B2-B4); R² = 0.62 and RMSE = 0.01 in the NIR region (B5-B8A).
• ResTrans21 Prediction Performance for Physiological Parameters:
  • Dry Matter (DM): HRS2S (R² = 0.84, RMSE = 11.51 kg ha⁻¹, RPD = 2.45), Rikola HSI (R² = 0.86, RMSE = 11.24 kg ha⁻¹, RPD = 2.51), Mjolnir V-1240 (R² = 0.89, RMSE = 10.00 kg ha⁻¹, RPD = 2.80).
  • Nitrogen Content (NC): HRS2S (R² = 0.85, RMSE = 0.13 %, RPD = 2.46), Rikola HSI (R² = 0.86, RMSE = 0.13 %, RPD = 2.46), Mjolnir V-1240 (R² = 0.85, RMSE = 0.13 %, RPD = 2.40).
  • Nitrogen Uptake (NU): HRS2S (R² = 0.86, RMSE = 0.48 kg ha⁻¹, RPD = 2.65), Rikola HSI (R² = 0.90, RMSE = 0.44 kg ha⁻¹, RPD = 2.89), Mjolnir V-1240 (R² = 0.92, RMSE = 0.35 kg ha⁻¹, RPD = 3.60).
• Comparison with Classical Baseline (GPR): ResTrans21 models consistently showed higher R² (0.07 to 0.22 higher) and RPD (0.41 to 1.03 higher) values compared to GPR for predictions using hyperspectral imagery. RMSE values were comparable, with GPR slightly lower for DM and NU, and ResTrans21 slightly lower for NC.
• Regularization Impact: Batch normalization and dropout regularizations significantly improved ResTrans21 model performance across all parameters and image sources.

Contributions

• First study to apply transformer models for crop growth status monitoring using super-resolved satellite imagery, demonstrating their effectiveness in precision agriculture.
• Developed a novel deep learning pipeline that fuses readily available, low-cost UAV-borne RGB imagery with free Sentinel-2 satellite data to generate high spatial and spectral resolution (HRS2S) imagery, enabling plot-level analysis previously limited by satellite resolution.
• Introduced ResTrans21, a unique composite transformer architecture capable of directly processing 2D image inputs and effectively extracting spatial-spectral features, outperforming classical machine learning (GPR) and traditional vegetation index-based methods.
• Demonstrated that the proposed approach provides a cost-effective and highly accurate alternative to expensive UAV-borne hyperspectral imagery for monitoring key physiological parameters in spring wheat.

Funding

• Norwegian Space Agency, grant number 74CO2210 (“Bruk av Copernicus/Sentinel-data til jordbruksfaglige problemstillinger”).

Citation

@article{Zhao2025Superresolve,
  author = {Zhao, Jiangsan and Geipel, Jakob},
  title = {Super-resolve satellite imagery to perform on par with UAV-borne hyperspectral imagery in predicting spring wheat physiological parameters using transformer models},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111204},
  url = {https://doi.org/10.1016/j.compag.2025.111204}
}