TL;DR
This study compares RGB imaging, active multispectral sensors (Greenseeker), and passive hyperspectral sensors for assessing early plant vigor in winter wheat. RGB imaging proved to be a cost-effective and accurate method for very early growth stages, while hyperspectral sensors with red-edge bands became more effective as the crop developed.
What Problem Does This Paper Address?
Accurately assessing early plant vigor in crops is crucial for optimizing nutrient management, predicting yield, and identifying promising breeding material. However, traditional spectral sensing methods can be hampered by soil background reflection, especially in the early stages when the canopy is sparse. This makes it challenging to obtain reliable measurements of plant traits like biomass and nitrogen uptake. The paper aims to evaluate different remote sensing technologies, including low-cost RGB imaging, to overcome these limitations and provide effective early-stage crop monitoring.
Methods
The researchers employed a split-split plot experimental design with winter wheat cultivars, varying nitrogen regimes, and sowing densities. They collected data over four early growth stages (Zadok's stages 14-32).
- RGB Imaging: A consumer-grade Nikon D5100 camera was used to capture RGB images. Image segmentation in the HSV color space was applied to extract canopy cover (CC).
- Active Multispectral Sensing: A Greenseeker handheld crop sensor was used to measure the Normalized Difference Vegetation Index (NDVI).
- Passive Hyperspectral Sensing: A Handyspec spectrometer captured spectral readings from 300 nm to 1140 nm. Various established vegetation indices (e.g., NDVI, NDRE, MCARI, MTCI, PRI, REIP, WBI) were calculated, along with spectral unmixing using two endmembers (bare soil and dense vegetation). Additionally, a "contour map analysis" was performed to identify optimal narrow-band combinations for estimating traits.
Correlations between sensor-derived metrics and ground-truth measurements (dry weight, nitrogen content, nitrogen uptake) were analyzed using squared Pearson correlation coefficients (R²).
Datasets
The study utilized data collected from a field experiment with winter wheat cultivars (Anapolis, Mulan, and Hybery) grown under different nitrogen fertilization rates and sowing densities at the Dürnast research station in Germany. Ground-truth data included dry weight (DW), nitrogen content (NC), and nitrogen uptake (Nup), measured at four sampling dates corresponding to Zadok's growth stages 12-32.
Key Insights and Ideas
- Early Stage Dominance of RGB: For the very early growth stages (until early tillering), RGB imaging-based canopy cover provided surprisingly strong correlations with dry weight and nitrogen uptake, outperforming spectral methods. This highlights the potential of low-cost RGB imaging when soil background is a significant factor.
- Soil Background Limitation: Spectral sensing, particularly in the early stages, was negatively impacted by soil background reflection, leading to weaker correlations compared to RGB imaging.
- Red Edge Importance for Later Stages: As canopy cover increased, hyperspectral sensors, especially those utilizing red-edge bands (e.g., NDRE), became more effective for estimating dry weight and nitrogen uptake.
- Contour Map Analysis for Optimization: The contour map analysis, which explores all possible narrow-band combinations, revealed optimal spectral indices that often outperformed established ones, particularly for early growth stages and specific traits. This suggests that custom index development can yield significant improvements.
- Greenseeker's Consistent Performance: The active Greenseeker sensor showed consistent and often competitive performance, sometimes even outperforming hyperspectral indices in the early stages, especially for canopy cover estimation.
- N Content Challenges: Estimating nitrogen content (NC) proved more challenging than DW or Nup, with weaker and less consistent correlations across all sensor methods.
Key Results
- Early Growth (Zadok's stages 14-20): RGB-based canopy cover achieved R² values greater than 0.60 for dry weight and nitrogen uptake, significantly outperforming other methods.
- Later Growth (from early tillering onwards): Red-edge based indices like NDRE and a newly developed normalized difference index (736 nm / 794 nm) showed the best correlations for dry weight and nitrogen uptake.
- Canopy Cover Estimation: Greenseeker and spectral unmixing correlated best with canopy cover in later stages.
- Contour Map Indices: Optimized spectral indices derived from contour map analysis showed potential for improved correlations, with specific band combinations (e.g., 792 nm / 736 nm for DW and Nup) outperforming established indices.
- Overall: The study demonstrated that RGB imaging is a viable low-cost option for very early growth stages, while multispectral and hyperspectral sensors with red-edge bands are crucial for later stages.
Limitations and Future Work
- Ground Resolution: The authors note that for UAV-based applications, better cameras might be needed to achieve sufficient ground resolution for accurate measurements.
- Lighting Conditions and Vibrations: Uncertainties related to less favorable lighting conditions and vibrations from mobile sensor platforms need to be addressed for practical applications.
- Robust Segmentation: More robust segmentation methods for RGB images are suggested.
- Non-linear Models: The paper hints that non-linear models might further improve results, particularly for spectral unmixing.
- Nitrogen Content: The difficulty in accurately estimating nitrogen content (NC) suggests that further research is needed in this area, potentially exploring different spectral regions or indices.
- Optimal Indices: The study identified promising narrow-band combinations, suggesting that further validation and refinement of these custom indices could be beneficial.
Summary updated