Discrimination of Annonaceae using herbarium leaf reflectance spectra under limited sample size conditions

This study demonstrates that despite variations in preservation history and limited sample sizes, herbarium leaf reflectance spectra can reliably identify Annonaceae species using supervised classification models, though accuracy depends on taxon-specific inter-individual variability.

The Gist

Imagine a massive, dusty library where the books aren't made of paper, but are actually dried-up leaves pressed between cardboard sheets. This is a herbarium. For centuries, botanists have used these "leaf books" to identify plants, looking at the shape of a leaf or the color of a flower. But what if we could ask these dried leaves a different question? What if we could ask them, "Who are you?" by shining a special light on them and listening to how they bounce that light back?

That is exactly what this paper by Boughalmi and his team set out to do. They wanted to see if dried, old museum leaves could still tell us their species identity using spectroscopy (a fancy way of saying "reading the light signature").

Here is the story of their experiment, broken down into simple concepts:

1. The Big Question: Are Old Leaves "Deaf" to Light?

Think of a fresh leaf like a crisp, colorful apple. It has a very clear "flavor" (or in this case, a light signature). But a herbarium leaf is like an apple that has been dried out, stored in a box for 50 years, maybe dipped in alcohol to stop it from rotting, and glued to a card.

The scientists worried: Has all that time and those chemical treatments wiped out the leaf's unique "voice"? If the leaf has changed too much, the computer might not be able to tell a "Mango" leaf from a "Papaya" leaf anymore.

2. The Experiment: The "Taste Test"

To test this, the team grabbed two groups of leaves from the tropical plant family Annonaceae (which includes the famous Custard Apple and Soursop).

• Group A (The Museum Giants): They went to the famous Natural History Museum in Paris. They scanned 14 different species of leaves. These leaves were old—some were nearly 200 years old! They didn't know exactly how these leaves were treated when they were collected, which made them a perfect test for "real-world" chaos.
• Group B (The Controlled Lab): They went to a forest in Ecuador. They collected fresh leaves and treated them in two ways: some were just dried, and others were soaked in alcohol (a common practice in humid jungles to stop mold) before drying. This let them see if the alcohol "ruined" the leaf's light signature.

They used a special scanner (like a high-tech barcode reader) to read the light bouncing off every leaf. Then, they fed this data into five different AI brain models (computer programs designed to learn patterns) to see if the computers could guess the species.

3. The Results: The Leaves Still Have a Voice!

The results were surprisingly good. It's like finding out that even though an old letter has yellowed and the ink has faded, you can still read the handwriting perfectly.

• The "Group Hug" Success: When the computers were trained with several leaves from each species (like having a few examples of every student in a class), they got it right 80% to 90% of the time.
• The "One Leaf" Challenge: What if you only have one leaf to represent a whole species? (This is common in museums; sometimes there's only one specimen). The computers got a bit confused, but they still did better than random guessing. Some species were easy to identify (like a distinct fingerprint), while others were tricky twins.
• The Alcohol Test: Did the alcohol soak ruin the data? Nope. The leaves soaked in alcohol still sounded exactly like their non-alcohol cousins to the computer. The "voice" was still there, just maybe a little quieter.

4. The "Twins" Problem

The study found that while the method works great for most plants, it struggles with cousins.

• Imagine trying to tell apart two identical twins who wear the same clothes. The computer sometimes mixed up two very similar species of a plant called Hexalobus.
• However, for plants that are more different from each other, the computer was a master detective.

5. Why This Matters (The "So What?")

This is a game-changer for science.

• Non-Destructive: You don't have to cut the leaf or destroy the specimen to study it. You just shine a light on it.
• Unlocking History: We have millions of these dried leaves sitting in museums. This study proves we can turn those silent, dried leaves into a massive digital database of plant traits.
• Future Proofing: Even if a species goes extinct, if we have a dried leaf in a museum, we can still study its chemical makeup and identity decades later.

The Bottom Line

Think of this study as proving that old, dried-out leaves still have a secret identity card hidden inside them. Even after decades of storage, alcohol dips, and glue, their "light signature" remains strong enough for a computer to say, "Ah, I know you! You are a Uvaria grandiflora!"

This opens the door to digitizing the world's botanical history, turning dusty museum shelves into a powerful, high-tech library of life on Earth.

Technical Summary

1. Problem Statement

Herbarium collections represent a vast, irreplaceable archive of plant biodiversity, yet their utility for species identification via spectral data is constrained by uncertainties regarding preservation histories. While Near-Infrared Spectroscopy (NIRS) has proven effective for identifying fresh, "pressed" specimens, it remains unclear if these signals persist in historical herbarium specimens subjected to:

• Variable preservation methods: Including pre-drying alcohol treatments, glue application, and diverse drying protocols.
• Temporal degradation: Specimens ranging from decades to centuries old.
• Geographic and environmental heterogeneity: Specimens collected across different climates and biogeographic regions.
• Limited sample sizes: The frequent scarcity of specimens per species in historical collections, which challenges the training of robust machine learning models.

The study aims to determine if herbarium leaf reflectance spectra retain a stable "Species Spectral Signature" (SSS) sufficient for accurate species-level discrimination under these realistic, noisy conditions.

2. Methodology

Study Organisms and Datasets

The researchers focused on the pantropical family Annonaceae (approx. 2,500 species), chosen for its taxonomic complexity and ecological diversity. Two distinct datasets were generated:

1. ID_PARIS (Historical Herbarium): 14 species scanned from the Muséum national d'Histoire naturelle (Paris). Specimens ranged from 9 to 201 years old (mean: 69 years). Detailed pre-drying treatment history was generally unavailable.
2. ID_YASUNÍ (Controlled Experiment): 9 species collected in the Yasuní Forest (Ecuador). This dataset specifically tested the impact of alcohol preservation. Specimens were either dried directly or sprayed with 70–80% alcohol for three days prior to drying.

Data Acquisition

• Instrumentation: ASD LabSpec® instruments were used.
  • ID_PARIS: Rapid Analysis Probe with fiber optic illuminator (20 readings/scan, 20 scans/specimen).
  • ID_YASUNÍ: 6 mm pen probe (30 readings/scan, 50 scans/specimen).
• Protocol: Spectra (350–2500 nm) were collected from the adaxial side of mature leaves, avoiding midribs and damaged areas.
• Preprocessing:
  • Standard Normal Variate (SNV): To normalize mean and variance, removing noise from instrument settings and sample thickness.
  • Savitzky-Golay Filter: Applied to compute the first derivative (polynomial order 2, window size 7) to capture spectral slope variations and smooth data.

Analytical Framework

Five supervised classification models were evaluated for high-dimensional spectral data:

1. Partial Least Squares Discriminant Analysis (PLS-DA)
2. Support Vector Machine (SVM)
3. K-Nearest Neighbors (KNN)
4. Linear Discriminant Analysis (LDA)
5. Random Forest (RF)

Experimental Design:

• Stratified Resampling: To prevent data leakage, specimens were strictly separated into tuning, calibration, and independent prediction subsets.
• Robustness Testing: Models were tested under varying conditions:
  • Using all spectra vs. averaged spectra per specimen.
  • Varying training set sizes (from 1 specimen to 8 specimens per species).
  • Randomly mixing alcohol-treated and non-treated specimens in training/testing sets to simulate unknown preservation histories.
• Validation: 100 random iterations were performed to ensure statistical stability.

3. Key Contributions

• Validation of Historical Spectra: Demonstrated that even specimens over 200 years old retain sufficient chemical/structural information for species discrimination.
• Alcohol Preservation Impact: Provided empirical evidence that short-term alcohol treatment (a common field practice in humid tropics) does not significantly degrade the taxonomic signal, as alcohol-treated specimens clustered closely with non-treated ones in PCA space.
• Sample Size Optimization: Quantified the relationship between training set size and classification accuracy, identifying a "sweet spot" for model performance.
• Model Benchmarking: Systematically compared five standard machine learning algorithms specifically for herbarium spectral data, identifying LDA as the most robust performer in this context.

4. Results

Overall Classification Accuracy

• High Accuracy: All models achieved high accuracy (>80%) when trained on multiple specimens.
  • LDA performed best (87.48% accuracy using averaged spectra).
  • SVM (86.35%) and RF (84.96%) followed closely.
  • KNN was the least accurate (79.23%).
• Averaging Effect: Using a single averaged spectrum per specimen (rather than 20 individual scans) consistently improved accuracy across all models.

Impact of Sample Size (Limited Data)

• Single Specimen: Accuracy varied drastically by species.
  • High performers: Anaxagorea dolichocarpa (85%), Ambavia gerrardii (83%).
  • Low performers: Uvaria grandiflora (38%), Hexalobus species (43–49%).
• Convergence: Accuracy stabilized near 90% when training sets reached 5 specimens per species.
• LDA Superiority: LDA consistently outperformed other models, achieving >70% accuracy with just one specimen and >90% with three.

Alcohol Treatment Effects

• In the ID_YASUNÍ dataset, models achieved >90% accuracy (reaching >99% for SVM, PLS-DA, and LDA) regardless of whether alcohol-treated specimens were in the training or test sets.
• Alcohol-treated specimens did not form distinct clusters in PCA space; they remained within the species-specific spectral envelope.

Taxonomic Challenges

• Congeneric Confusion: Closely related species (e.g., Hexalobus crispiflorus vs. H. monopetalus) showed higher misclassification rates, often confusing one for the other.
• Intraspecific Variability: Some species (e.g., Uvaria grandiflora) showed poor classification despite being the sole representative of their genus, suggesting that intraspecific variability or specimen condition (glue, curling) can be more detrimental than taxonomic proximity.

5. Significance and Implications

• Non-Destructive Taxonomy: Confirms that herbarium spectroscopy is a viable, non-destructive tool for species identification, particularly valuable for rare, type, or extinct specimens where morphological features (flowers/fruits) are missing.
• Digitization of Biodiversity: Supports the integration of historical collections into "high-dimensional trait space," allowing researchers to extract functional trait data from millions of existing specimens.
• Operational Feasibility: The workflow is efficient (~10 minutes per specimen), suggesting that large-scale digitization of herbarium spectral data is operationally feasible.
• Methodological Guidance:
  • Sample Size: Researchers can achieve robust results with as few as 5 specimens per species.
  • Preprocessing: Standard techniques (SNV, derivatives) effectively mitigate noise from preservation artifacts.
  • Protocol Harmonization: The study highlights the need for standardized scanning protocols (e.g., IHerbSpec) to ensure interoperability across different herbaria and instruments.

Conclusion: The study concludes that despite variations in age, geography, and preservation history, herbarium leaf spectra retain a strong, coherent taxonomic signal. This validates the use of machine learning on historical collections for biodiversity monitoring, trait estimation, and the detection of cryptic diversity.