⚡ electrical engineering

Integrating Feature Selection and Machine Learning for Nitrogen Assessment in Grapevine Leaves using In-Field Hyperspectral Imaging

This study demonstrates that integrating in-field hyperspectral imaging with an ensemble feature selection framework enables accurate, cultivar-specific, and scale-consistent estimation of nitrogen status in grapevine leaves and canopies, offering a robust tool for optimizing vineyard fertilization.

Original authors: Atif Bilal Asad, Achyut Paudel, Safal Kshetri, Chenchen Kang, Salik Ram Khanal, Nataliya Shcherbatyuk, Pierre Davadant, R. Paul Schreiner, Santosh Kalauni, Manoj Karkee, Markus Keller

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Atif Bilal Asad, Achyut Paudel, Safal Kshetri, Chenchen Kang, Salik Ram Khanal, Nataliya Shcherbatyuk, Pierre Davadant, R. Paul Schreiner, Santosh Kalauni, Manoj Karkee, Markus Keller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a grape farmer. Your goal is to make the best wine possible. To do that, your grapevines need the perfect amount of Nitrogen (a vital nutrient, like vitamins for humans).

Too little Nitrogen: The grapes are small, the wine tastes weak, and the vines get sick easily.
Too much Nitrogen: The vines grow huge, leafy "jungle" canopies that shade the grapes. The grapes never ripen properly, turn green instead of red, and the wine lacks flavor. Plus, the extra nitrogen leaks into the soil and pollutes local rivers.

The Problem:
Traditionally, to check if a vine has enough nitrogen, a farmer has to walk through the field, pluck leaves, put them in a bag, drive them to a lab, wait days for chemical results, and then guess what to do. It's slow, expensive, and by the time you get the answer, the vine might have already changed its needs.

The Solution in This Paper:
This research team invented a "smart camera" system that acts like a super-powered stethoscope for grapevines. Instead of cutting leaves, they take high-tech photos of the vines and use a computer to instantly tell them exactly how much nitrogen is inside.

Here is how they did it, broken down into simple steps:

1. The "Super-Eye" Camera (Hyperspectral Imaging)

Imagine a regular camera sees the world in Red, Green, and Blue (like your phone). This special camera sees hundreds of colors that human eyes can't even see, stretching from deep reds to invisible infrared light.

When light hits a grape leaf, the chemicals inside (like chlorophyll and nitrogen) absorb and bounce back specific colors. It's like a fingerprint. A leaf with too much nitrogen bounces light differently than a leaf with too little. The camera captures this "light fingerprint" for every single leaf.

2. The "Noise-Canceling" Filter (Feature Selection)

The camera takes so many pictures (hundreds of color bands) that it creates a massive mountain of data. It's like trying to find a specific needle in a haystack, but the haystack is made of 1,000 different types of needles.

The researchers built a smart filter (an "Ensemble Feature Selection" framework). Think of this as a team of detectives working together:

Some detectives look for patterns in the red colors.
Others look at the invisible infrared colors.
They vote on which specific colors are the most important for guessing nitrogen levels.

They realized that white grapes (like Chardonnay) and red grapes (like Pinot Noir) are different.

White Grapes: Their "nitrogen fingerprint" is a mix of all colors (red, green, and invisible infrared).
Red Grapes: Because they have red pigments (anthocyanins), their fingerprint is dominated by the visible red and green colors. The red pigments hide some of the other signals, so the computer focuses on what it can see clearly.

3. The "Crystal Ball" (Machine Learning)

Once the team picked the best "fingerprint colors," they taught a computer (Machine Learning) to read them.

They fed the computer thousands of examples: "Here is a photo of a leaf, and here is the actual nitrogen level we measured in the lab."
The computer learned the rules: "If the leaf reflects this specific shade of red and this specific shade of infrared, it probably has 2.5% nitrogen."

4. The Big Test: From One Leaf to the Whole Bush

The hardest part of this study was a "leap of faith."

Leaf Level: They first trained the computer on photos of single leaves. It was easy because the light was perfect, and the leaf was flat. The computer was very accurate here.
Canopy Level: Then, they asked the computer to look at the whole vine (a messy bush of 20–40 leaves, with shadows, different angles, and wind).

The Surprise: The computer didn't just work on single leaves; it successfully transferred its knowledge to the whole bush!

They took the "fingerprint rules" learned from a single Chardonnay leaf and applied them to a whole Chardonnay bush. It worked!
They took the rules from a Pinot Noir leaf and applied them to a Syrah bush (another red grape). It worked!

Why This Matters

This research is like giving farmers a magic wand that can scan a whole vineyard in minutes.

Speed: No more waiting days for lab results.
Precision: Farmers can treat every single vine exactly how it needs it (Variable Rate Application), saving money and protecting the environment.
Simplicity: They proved you don't need the entire mountain of data. You only need a few specific "magic colors" to get the job done. This means future sensors could be smaller, cheaper, and easier to use.

In a nutshell: The researchers taught a computer to "see" nitrogen in grape leaves by finding the specific colors that matter most. They proved that this trick works not just on single leaves, but on whole grapevines, and it works for both white and red grapes. It's a faster, smarter way to grow better wine while saving the planet.

1. Problem Statement

Nitrogen (N) is a critical nutrient for grapevine vigor, fruit composition, and wine quality. However, soil N availability varies spatially and temporally, making precise, plant-level management difficult.

Current Limitations: Traditional N assessment relies on labor-intensive, destructive leaf sampling and slow laboratory chemical analysis. This creates a time lag, lacks spatial resolution for individual vine management, and cannot support real-time decision-making.
Technical Challenges: While hyperspectral imaging (HSI) offers a non-destructive alternative, the high dimensionality of HSI data (hundreds of narrow bands) leads to data redundancy, the "Hughes effect" (overfitting due to limited samples vs. high features), and computational inefficiency. Existing methods often use the full spectrum or simple vegetation indices, failing to identify the most physiologically relevant, cultivar-specific spectral bands.

2. Methodology

The study employed a multi-stage pipeline integrating advanced feature selection with machine learning (ML) to estimate N concentration at both leaf and canopy levels.

A. Data Collection

Scope: Data was collected over two growing seasons (2022–2023) across two US regions (Washington and Oregon).
Cultivars: Four varieties: Chardonnay (white), Pinot Noir (red), Concord (red hybrid), and Syrah (red).
Stages: Bloom and Veraison.
Sensors: Ground-based VNIR hyperspectral camera (400–1000 nm, 274 bands).
Ground Truth:
- Leaf Level: Individual leaves sampled and analyzed for % Dry Weight (DW) N.
- Canopy Level: Composite samples (20–40 leaves per vine) analyzed for % DW N.
Preprocessing: Radiometric calibration, outlier removal (Median Absolute Deviation), Standard Normal Variate (SNV) transformation, and Savitzky-Golay smoothing.

B. Ensemble Feature Selection Framework

To address dimensionality and redundancy, the authors developed a novel ensemble approach:

Hierarchical Clustering: Reduced multicollinearity by grouping highly correlated bands. A representative band was selected from each cluster based on a Consensus Score (CS) integrating:
- Pearson correlation with N.
- Mutual Information (non-linear dependency).
- Centrality within the cluster.
Ensemble Ranking: Eight distinct ML algorithms (SelectKBest, Lasso, Ridge, Elastic Net, Random Forest, Extra Trees, Gradient Boosting, and a customized Random Frog algorithm) were used to rank features.
Stability Aggregation: Feature stability was quantified by tracking appearance frequency in top-k selections across 100 bootstrap iterations. Final scores combined ranking performance and stability.
Optimization: A Sequential Feature Optimization using Partial Least Squares Regression (PLSR) determined the optimal number of bands to maximize $R^2$ while minimizing RMSE, avoiding overfitting.

C. Machine Learning Modeling

Models: Elastic Net (EN), Support Vector Regression (SVR-RBF), XGBoost, and Gaussian Process Regression (GPR).
Inputs: Optimized spectral bands + one-hot encoded growth stage (Bloom/Veraison).
Validation: Nested cross-validation with GroupKFold (grouped by growth stage and year) to prevent data leakage and ensure robust generalizability.

3. Key Contributions

Novel Ensemble Framework: Introduced a robust feature selection pipeline that combines hierarchical clustering with multiple complementary ML selectors to identify stable, physiologically meaningful spectral bands.
Cultivar-Specific Insights: Demonstrated that optimal spectral bands differ significantly between white and red cultivars due to pigment interactions (anthocyanin vs. chlorophyll).
Cross-Scale Transferability: Successfully tested and validated the transfer of leaf-level selected bands to canopy-level predictions, proving that N-sensitive features remain robust across measurement scales.
Ground-Based HSI Validation: Provided a comprehensive assessment of ground-based hyperspectral imaging for vineyard N monitoring, filling a gap left by studies focused primarily on aerial (UAS) platforms.

4. Key Results

A. Feature Selection Patterns

White Cultivars (Chardonnay): Selected bands were balanced across Visible (VIS), Red Edge, and Near-Infrared (NIR) regions.
Red Cultivars (Pinot Noir, Syrah, Concord): Selected bands were heavily concentrated in the Visible region (400–670 nm). This is attributed to anthocyanin-chlorophyll interactions, where N status alters pigment ratios more visibly than in white grapes.
Optimal Band Count: Reduced feature sets (5–14 bands) outperformed or matched full-spectrum models (274 bands), confirming that dimensionality reduction minimizes noise.

B. Prediction Performance

Leaf Level:
- Chardonnay: SVR-RBF achieved the highest accuracy ( $R^2 = 0.82$ , RMSE = 0.19% DW).
- Pinot Noir: GPR performed best ( $R^2 = 0.69$ , RMSE = 0.20% DW).
- Leaf-level models generally outperformed canopy models due to reduced environmental noise (lighting, leaf angle).
Canopy Level:
- Chardonnay: GPR ( $R^2 = 0.65$ , RMSE = 0.17% DW).
- Concord: GPR/XGBoost ( $R^2 = 0.72$ , RMSE = 0.12% DW).
- Syrah: Elastic Net ( $R^2 = 0.70$ , RMSE = 0.16% DW).
Transferability: Models trained on leaf-level selected bands successfully predicted canopy-level N with comparable accuracy (e.g., Chardonnay leaf-trained model on canopy data: $R^2 = 0.67$ vs. canopy-trained model: $R^2 = 0.65$ ). This confirms the robustness of the selected spectral features across scales.

C. Seasonal Dynamics

Models successfully captured the biological trend of higher N concentrations during the Bloom stage compared to Veraison, validating the model's ability to learn phenological shifts.

5. Significance and Implications

Precision Agriculture: The study enables rapid, non-destructive, vine-specific N monitoring, facilitating Variable Rate Application (VRA) of fertilizers to optimize yield and quality while reducing environmental runoff.
Sensor Development: The identification of compact, cultivar-specific band sets (e.g., specific VIS bands for red grapes, balanced bands for white grapes) provides a blueprint for designing cost-effective, customized multispectral sensors for field deployment, eliminating the need for expensive full-spectrum hyperspectral cameras.
Scalability: The demonstrated transferability of leaf-level features to canopy levels suggests that models trained on detailed leaf data can be effectively scaled up for whole-vine monitoring, bridging the gap between physiological research and practical vineyard management.

In conclusion, this research establishes a robust, transferable framework for nitrogen assessment in vineyards, proving that integrating ensemble feature selection with machine learning can overcome the challenges of hyperspectral data complexity and deliver actionable insights for precision viticulture.