Leaf and cluster spectral signatures reveal trait-dependent prediction performance for grapevine cluster architecture and juice quality

This study demonstrates that hyperspectral reflectance from grapevine clusters, rather than leaves, combined with trait-specific data partitioning strategies, significantly improves the prediction accuracy of cluster architecture and juice quality traits, offering optimized non-destructive phenotyping approaches for grapevine breeding.

The Gist

Imagine you are a grapevine breeder. Your goal is to find the perfect grapevine clones that produce juicy, high-quality wine and resist diseases. Traditionally, to check if a grape cluster is good, you have to pick it, crush it, and run expensive lab tests. It's slow, destructive, and you can't do it for every single vine in a massive vineyard.

This paper is about a new, faster way to "read" the grapes using light. Think of it like a "super-vision" that can tell you what's happening inside a grape just by looking at how it bounces light back to a camera.

Here is the story of what the researchers discovered, broken down into simple concepts:

1. The Problem: Guessing the Inside from the Outside

Grapes are tricky. A grapevine has two main parts: the leaves (the solar panels) and the clusters (the fruit).

• The Old Way: Scientists mostly looked at the leaves to guess how the fruit was doing. It's like trying to guess how a cake is baking by only looking at the oven door. Sometimes it works, but often the oven door doesn't tell you if the cake is burnt or undercooked.
• The New Idea: Why not look at the fruit itself? The researchers asked: "If we scan the leaves, do we get a better picture of the grapes? Or if we scan the grapes directly, do we get a better picture?"

2. The Experiment: The "Light Scanner"

The team went to a vineyard in Germany with two famous grape types: Riesling and Pinot. They had hundreds of different clones (like different family branches of the same grape).

They used a special handheld scanner that shoots light at the grapes and leaves and reads the "echo" (reflectance). It's like a bat using sonar, but with light. They scanned:

• Dry Leaves: Taken earlier in the season.
• Fresh Clusters: Taken right at harvest.

Then, they used a smart computer algorithm (a digital detective) to try and predict two things:

1. Architecture: How many berries are in the bunch? Are they tight or loose? How big are they?
2. Quality: How sweet is the juice? How acidic is it? What is the pH?

3. The Big Discovery: "Go to the Source!"

The results were like finding a missing piece of a puzzle.

• The Leaf vs. The Cluster:

  • When they tried to predict the size and shape of the berries (like how many berries are in a bunch), the cluster scans were the clear winners.
  • Analogy: If you want to know how many people are in a room, it's better to look at the people directly (the cluster) than to look at the shadows they cast on the wall from the hallway (the leaves).
  • However, for juice pH (acidity), the leaves actually did a surprisingly good job. This suggests that the leaves and the fruit are talking to each other chemically, like a phone call between a parent and a child.

• The "Loose" vs. "Tight" Bunches:

  • They also tried grouping the data by how "tight" the grape bunches were (loose, medium, or compact).
  • Analogy: Imagine trying to teach a robot to recognize dogs. If you only show it Golden Retrievers, it might fail when it sees a Poodle. But if you show it all types of dogs mixed together, it learns the general concept of "dog" better.
  • The researchers found that mixing the different types of bunches together in their training data actually helped the computer learn better, rather than separating them too strictly.

4. The "Secret Code" of Light

The computer didn't need the whole rainbow of light to make these predictions. It found a "secret code" in specific colors:

• Visible Light (Red/Blue): Tells the computer about the pigments (like the red skin of a grape).
• Red-Edge: A special transition zone between red and infrared that is very sensitive to plant health.
• Near-Infrared: Tells the computer about water content and sugar.

It's like a chef who doesn't need to taste every single ingredient to know the soup is ready; they just need to smell the specific aroma of the herbs to know the dish is perfect.

5. Why This Matters for the Future

This study is a game-changer for grape breeding.

• Speed: Instead of crushing thousands of grapes in a lab, breeders can walk through a vineyard with a scanner and instantly know which vines have the best fruit structure and juice quality.
• Precision: They learned that you can't use a "one-size-fits-all" approach. If you want to know about the fruit's shape, scan the fruit. If you want to know about acidity, the leaves might actually help.
• Scalability: This technology can be put on drones or robots. Soon, we might have flying robots that scan entire vineyards, telling farmers exactly which vines to keep and which to cut, all without touching a single grape.

The Bottom Line

The paper teaches us that context is everything. To understand a grapevine, you need to look at the right part of the plant at the right time. By using light as a tool, we can finally "see" the invisible qualities of grapes, making it easier to breed better, tastier, and more resilient wines for the future.

Technical Summary

1. Problem Statement

Grapevine breeding programs face significant challenges in selecting for cluster architecture (e.g., berry number, diameter, compactness) and juice quality traits. These traits are critical for yield stability, disease resistance (specifically Botrytis cinerea susceptibility in compact clusters), and winemaking quality. However, traditional phenotyping is:

• Destructive and labor-intensive: Requiring manual harvesting and chemical analysis.
• Low-throughput: Limiting the scale of evaluation in breeding pipelines.
• Complex: The relationship between canopy morphology and biochemical quality is non-linear and influenced by genetic background.

While High-Throughput Phenotyping (HTP) using hyperspectral reflectance offers a non-destructive alternative, it remains unclear how the source of the spectral data (different plant organs like leaves vs. clusters) and data partitioning strategies (grouping by population vs. cluster type) affect the accuracy and generalizability of predictive models.

2. Methodology

Study Design & Populations:

• Location: Experimental vineyards of Hochschule Geisenheim University, Germany (2024 harvest).
• Populations: Two extensive clonal populations of Vitis vinifera:
  • Riesling: 220 clones (217 White, 3 Red).
  • Pinot: 240 clones (214 Noir, 7 Precoce, 15 Blanc, 4 Gris).
• Phenotyping:
  • Cluster Architecture: Measured using a handheld 3D scanner (Artec 3D Spider) to extract berry number, diameter, volume, cluster volume, weight, and convex hull.
  • Juice Quality: Berries were pressed, homogenized, and analyzed via Fourier-transformed infrared (FTIR) spectroscopy for pH, nitrogen (NOPA), density, total sugars, and organic acids (tartaric, malic).
  • Spectral Data Collection:
    • Cluster Reflectance: Measured immediately after 3D scanning on fresh berries (350–2500 nm).
    • Leaf Reflectance: Measured on dried leaves collected at flowering (BBCH 65) and oven-dried (60°C).
  • Preprocessing: Savitzky-Golay filtering (window=11, order=3) and spectral range restriction (400–2400 nm).

Statistical Modeling:

• Algorithm: Partial Least Squares Regression (PLSR) was used to model the relationship between hyperspectral predictors ($X$) and traits ($Y$).
• Spectral Inputs: Six distinct predictor matrices were tested:
  1. Dry leaf reflectance.
  2. Cluster reflectance.
  3. Averaged organ reflectance.
  4. Combined organ reflectance.
  5. Weighted 70% cluster / 30% leaf.
  6. Weighted 30% cluster / 70% leaf.
• Data Partitioning Strategies: Two schemes were compared to test model generalization:
  1. Cluster Architecture-based: Grouped by OIV descriptors (Compact, Intermediate, Loose).
  2. Population-based: Grouped by cultivar (Riesling vs. Pinot).
• Validation: 80% training / 20% independent validation split. Model performance assessed via $R^2$, RMSE, and bias. Optimal latent components selected via cross-validation.
• Feature Importance: Variable Importance in Projection (VIP) scores identified key spectral wavelengths.

3. Key Results

Phenotypic Variation:

• Riesling vs. Pinot: Riesling clusters were significantly larger in volume, berry count, and berry diameter, with lower pH and higher total acidity. Pinot had higher NOPA and total sugars.
• Cluster Types: Compact clusters generally had higher berry numbers and weights. Riesling showed greater variability in acidity across cluster types compared to Pinot.

Predictive Accuracy (Organ Specificity):

• Cluster Architecture:
  • Cluster reflectance significantly outperformed leaf reflectance for berry-specific traits.
  • Berry Diameter: High accuracy ($R^2 = 0.79$).
  • Berry Number: Moderate accuracy ($R^2 = 0.53$).
  • Cluster-level traits (Weight, Volume, Convex Hull) were poorly predicted ($R^2 < 0.10$) across all spectral sources, likely due to occlusion and structural complexity.
• Juice Quality:
  • pH: Best predicted using dry leaf reflectance ($R^2 = 0.63$), suggesting a physiological coupling between leaf ion regulation and berry pH.
  • Total Acidity: Best predicted using cluster reflectance ($R^2 = 0.48$).
  • Total Sugars: Best predicted using cluster reflectance.
  • Organic Acids (Malic/Tartaric): Generally low prediction accuracy across all models.

Data Partitioning Impact:

• Partitioning by cluster architecture type (Compact/Intermediate/Loose) generally yielded slightly higher $R^2$ and lower RMSE than population-based partitioning.
• Population-based models suffered from reduced phenotypic variance in the training sets, limiting generalization.

Key Spectral Regions (VIP Analysis):

• Leaf Models: Critical wavelengths were in the Visible (VIS), Red-Edge, and NIR regions.
• Cluster Models: Critical wavelengths were concentrated in the VIS (400–778 nm) and Red-Edge, with additional contributions from NIR/SWIR water absorption bands (950–1200 nm).

4. Key Contributions

1. Organ-Specific Predictive Power: The study definitively establishes that the choice of plant organ for spectral measurement is trait-dependent. Cluster spectra are superior for morphological berry traits, while leaf spectra can surprisingly outperform cluster spectra for specific chemical traits like pH.
2. Limitations of Combined Spectra: Creating synthetic "combined" spectral profiles (mixing leaf and cluster data) did not improve accuracy and often reduced it, suggesting that mixing distinct biochemical signals dilutes trait-specific associations.
3. Data Partitioning Strategy: Demonstrated that grouping data by phenotypic cluster type rather than genetic population improves model calibration for architecture traits, likely by capturing more relevant variance within the training set.
4. Identification of Key Wavelengths: Pinpointed specific spectral bands (VIS, Red-Edge, and water absorption regions) that drive prediction, providing a roadmap for developing cost-effective multispectral sensors for breeding programs.

5. Significance and Implications

• Breeding Efficiency: This research enables the development of scalable, non-destructive proxies for complex traits. Breeders can now select for specific cluster architectures (e.g., loose bunches to reduce disease risk) and juice quality parameters using rapid spectral scanning.
• Sensor Optimization: The identification of key spectral regions allows for the design of targeted multispectral cameras (rather than expensive hyperspectral systems) for UAVs or field gantries, making HTP more accessible.
• Model Generalization: The findings highlight that phenomic models must be calibrated with diverse germplasm and appropriate data partitioning to ensure transferability across different environments and genetic backgrounds.
• Physiological Insight: The success of leaf spectra in predicting juice pH suggests a strong physiological link between leaf ion status and berry chemistry, offering new avenues for understanding source-sink dynamics in grapevines.

Conclusion:
The study concludes that while hyperspectral phenomics is a powerful tool for grapevine breeding, its success relies heavily on matching the spectral source (organ) to the target trait and employing strategic data partitioning. Future work should focus on integrating multi-organ data across different phenological stages to build robust, generalizable models for perennial crops.