Abstract
International audience
Introduction
In precision agriculture and crop phenotyping, determining plant density—the number of plants per square meter—is a critical task for early-season management. It directly influences final yield and dictates decisions regarding fertilization and weeding. While high-resolution RGB imaging combined with deep learning is the current "gold standard," it is computationally intensive and requires extremely high spatial resolution to distinguish tiny seedlings from the soil. This study explores spectral reflectance as a high-throughput alternative for estimating the density of straw cereals at their earliest growth stages (1 to 3 leaves).
Methodology
The researchers implemented a two-step estimation framework to bridge the gap between light reflectance and plant count:
- Spectra to Green Fraction (GF): Gaussian Process Regression (GPR) was used to estimate the percentage of ground covered by green vegetation from spectral data.
- GF to Plant Density: A linear regression model converted the GF value into a final plant density estimate.
To optimize the model and prevent overfitting, the team performed forward band selection using the Akaike Information Criterion ($AIC_c$). This process identified the most parsimonious model—one that uses the minimum number of spectral bands required to maintain accuracy.
Results and Discussion
The study utilized a diverse dataset across six sites in France and China, covering species such as soft wheat, durum wheat, barley, and rye. Several key insights emerged:
- The 45° Advantage: Viewing the canopy at a 45° angle was significantly more effective than the nadir (0°) view. The oblique angle increases the "projected area" of small seedlings, making them easier to detect against the soil background.
- Optimal Band Selection: The $AIC_c$ analysis revealed that just four spectral bands (specifically around the Red and Red-edge/NIR regions, ~690nm and ~760nm) provided the best balance between simplicity and performance. Using more bands increased the risk of overfitting without substantial gains in accuracy.
- Soil Interference: Soil color and type vary significantly between locations, acting as "noise." Consequently, the researchers found that site-specific calibration is mandatory for accurate estimation.
Performance and Throughput
The best-performing models achieved a relative Root Mean Square Error (rRMSE) of 32.6% (RMSE of 82 plants/m²). While this is less precise than image-based methods, which can reach ~10% error, spectral sensors are less sensitive to pixel resolution. This makes them a promising tool for rapid scanning of large fields via drones or ground vehicles where high-resolution imagery may not be feasible.
Conclusion
While spectral measurements currently offer moderate accuracy compared to surgical imaging techniques, their potential for high-throughput application makes them a valuable asset for industrial-scale phenotyping. Future work will focus on refining high-spectral-resolution data mining to bridge the accuracy gap while maintaining the speed advantages of spectral sensing.
Summary updated