Using a planted tree biodiversity experiment to evaluate imaging spectroscopy for species classification

This study evaluates airborne imaging spectroscopy for tree species classification using controlled biodiversity experiments in Germany, finding that while the method achieves high accuracy for up to four species, its effectiveness declines in more diverse plots, suggesting it may be better suited for monitoring functional diversity rather than distinguishing numerous individual species.

The Gist

Imagine you are a detective trying to identify different types of trees in a massive forest from a helicopter. You have a special high-tech camera that doesn't just take a photo; it sees the "light fingerprint" (spectra) of every leaf. The goal? To tell a Douglas Fir from a Beech tree just by looking at how they reflect light, without ever touching them.

This paper is the report card on how well that detective work actually works.

The Setup: The "Tree School" Experiment

In the real world, forests are messy. Trees grow in different soils, next to different neighbors, and at different ages. It's like trying to identify students in a crowded cafeteria where everyone is wearing different outfits and sitting in different lighting; it's hard to tell who is who.

To solve this, the researchers used two special "tree schools" in Germany called BIOTREE.

• The Concept: Instead of a wild forest, these are controlled experiments where scientists planted specific trees in neat, square or circular patches.
• Site A (Kaltenborn): A small school with only 4 types of trees.
• Site B (Bechstedt): A huge, crowded school with 16 different types of trees.

Because every tree in a patch was the same species and grew in the exact same soil, the researchers could be sure that any difference in the "light fingerprint" was actually due to the tree species, not the environment.

The Tools: The Super-Camera and The Brains

1. The Camera (HyPlant): They flew a plane over the sites with a sensor that sees 589 different colors of light (from visible violet to invisible infrared). It's like having a camera that sees 589 shades of gray instead of just red, green, and blue.
2. The Brains (Algorithms): They fed this data into two different computer "brains" to make the guesses:
   • LDA (Linear Discriminant Analysis): Think of this as a straightforward teacher. It looks for simple, straight-line differences between the groups. "If the light is this bright, it's a Pine; if it's that bright, it's an Oak."
   • SVM (Support Vector Machine): Think of this as a complex puzzle solver. It tries to find weird, curved, non-linear patterns to separate the groups. It's smarter but can sometimes get confused by trying too hard.

The Results: The Good, The Bad, and The Ugly

1. The "Easy" Test (Kaltenborn - 4 Species)

When the computer tried to identify the 4 tree types in the small school, it did amazingly well.

• Accuracy: Up to 83% for new, unseen trees.
• The Winner: The "straightforward teacher" (LDA) actually did better than the "complex puzzle solver" (SVM).
• The Lesson: When there are only a few options, simple rules work best. The trees had very distinct "light fingerprints."

2. The "Hard" Test (Bechstedt - 16 Species)

When they tried to identify 16 different tree types in the crowded school, the results dropped significantly.

• Accuracy: Only 31% to 49%. That's barely better than guessing!
• The Problem: With so many species, the "light fingerprints" started to blur together. A young Oak looks a lot like a young Beech when viewed from 680 meters up in the air. The computer got confused.
• The Twist: Interestingly, the "complex puzzle solver" (SVM) got worse at guessing new trees than the simple teacher. It had memorized the training data too perfectly (like a student who memorized the textbook but can't apply it to a new test).

The Big Takeaway: Why It Matters

The "Identity Crisis" of Trees
The paper reveals a hard truth: Airborne cameras are great at spotting broad categories (like "Conifers" vs. "Broadleaf trees") but struggle to tell specific species apart in a diverse forest.

Think of it like this:

• If you look at a crowd from a helicopter, you can easily tell the difference between a group of people wearing red shirts and a group wearing blue shirts.
• But if you have 16 different groups, each wearing slightly different shades of green, it becomes nearly impossible to tell them apart from the sky.

The Silver Lining
Even though the computer couldn't name every single tree species perfectly in the diverse forest, the data was still useful.

• Functional Diversity: Instead of asking "What species is that?", the technology is better at asking "How different are these trees from each other?"
• This helps scientists monitor forest health. If a forest has high "functional diversity" (lots of different types of trees doing different jobs), it's more resilient to droughts and diseases.

The Bottom Line

This study is a reality check for remote sensing.

• Can we map forests from the sky? Yes.
• Can we count every single tree species perfectly? Not yet, especially in complex, diverse forests.
• What should we do? Use these tools to monitor the variety and health of the forest (functional diversity) rather than trying to create a perfect species list. It's like judging a choir by how harmonious the sound is, rather than trying to identify every single singer's name from the back of the hall.

The researchers suggest that to get better results, we might need to combine these "light fingerprint" cameras with other tools, like LiDAR (which measures the 3D shape of the trees), to give the computer a better clue.

Technical Summary

1. Problem Statement

Forest biodiversity monitoring is critical for understanding ecosystem health, yet field-based surveys are often sparse due to logistical constraints. Remote sensing offers a scalable solution, but classifying tree species using airborne imaging spectroscopy in natural forests faces significant challenges:

• Spectral Confusion: Different species often share similar spectral properties due to shared traits, while individuals of the same species may exhibit different spectra due to microsite variations or neighbor effects.
• Environmental Covariation: In observational studies, species occurrence often correlates with environmental gradients, making it difficult to disentangle taxonomic signals from environmental noise.
• Accuracy Variability: Previous studies show variable classification accuracy, often failing in species-rich or mixed-canopy scenarios.

The authors aim to evaluate the potential of airborne imaging spectroscopy for species classification under controlled conditions where environmental variables and species arrangements are standardized, thereby isolating the spectral signal of species identity.

2. Methodology

Study Sites and Experimental Design

The study utilized two sites from the BIOTREE biodiversity experiment in Germany, featuring planted trees in monospecific patches to eliminate crown overlap and microsite heterogeneity:

• Kaltenborn Site: Contains 16 plots with varying species richness (1 to 4 species) drawn from a pool of 4 species (Pseudotsuga menziesii, Picea abies, Fagus sylvatica, Quercus petraea). Patches were $8 \times 8$ m.
• Bechstedt Site: Contains 25 hexagonal plots manipulating functional diversity using a pool of 16 species (including various Acer, Betula, Fagus, Pinus, etc.). Patches were circular with a 3.5 m radius.

Data Acquisition

• Sensor: HyPlant imaging spectrometer mounted on a Cessna 208B aircraft.
• Spectral Range: 400–2500 nm with 589 quasi-continuous bands (3 nm resolution in VIS-NIR; 10 nm in SWIR).
• Spatial Resolution: 1 m ground sampling distance.
• Timing: Flown on June 17, 2021, during peak growing season.
• Preprocessing: Data converted to top-of-canopy reflectance using ATCOR-4 atmospheric correction.

Data Processing and Analysis

• Pixel Selection: To minimize shadow effects, the analysis focused on the three brightest pixels within the 25 central pixels of each monospecific patch.
• Statistical Analysis:
  • ANOVA: General linear models were fitted across 589 wavelengths to quantify the contribution of taxonomic groups (class, order, species), diversity levels, and their interactions to spectral variance.
  • Classification: Two methods were compared:
    1. Linear Discriminant Analysis (LDA): Linear separation.
    2. Support Vector Machine (SVM): Non-linear separation using radial basis function kernels.
  • Input Features: Models were trained using varying numbers of similarly-spaced wavelengths or Principal Components (PCs) derived from the full spectrum.
• Validation: Accuracy was assessed via:
  • In-sample (training on all data).
  • Within-sample prediction (random splits).
  • Out-of-sample prediction (training on specific plots/diversity levels and predicting spatially distinct patches).

3. Key Results

Spectral Variation

• Taxonomic Signal: Significant spectral variation was detected between taxonomic classes (Gymnosperms vs. Angiosperms), orders, and individual species across the spectrum.
• Wavelength Importance: Discriminatory power was generally higher in the NIR and SWIR regions (influenced by leaf structure, water content, and non-pigment constituents) compared to the Visible (VIS) range.
• Diversity Interactions: Significant interactions between taxonomic identity and diversity levels were found, though their effect sizes were an order of magnitude smaller than the main taxonomic differences.

Classification Accuracy

• Training Performance: Both LDA and SVM achieved near-perfect accuracy (up to 100%) in training datasets when using sufficient spectral dimensions.
• Kaltenborn (4 Species):
  • Out-of-sample accuracy: 77%–83% using LDA; slightly lower for SVM.
  • LDA performed better than SVM in prediction tasks.
  • Using 20–25 wavelengths or PCs was sufficient to reach peak accuracy.
• Bechstedt (16 Species):
  • Out-of-sample accuracy: Significantly lower, ranging from 31% to 49%.
  • LDA (49%) outperformed SVM (31–36%).
  • Specific species like Fagus sylvatica were poorly classified (<25% accuracy), while others like *Ulmus glabra* were well classified (>80%).
• Cross-Site Prediction: Predicting Fagus and Quercus in Bechstedt using Kaltenborn as training data yielded moderate success (~78% for LDA), though Fagus remained difficult to distinguish.

Methodological Comparisons

• LDA vs. SVM: LDA consistently provided more robust generalization for out-of-sample predictions. SVM tended to overfit the training data, achieving 100% accuracy even with randomized species labels when using many PCs, indicating it was separating individuals based on noise rather than species identity.
• Wavelengths vs. PCs: For LDA, using original wavelengths was as effective as using PCs. For SVM, PCs significantly improved performance over raw wavelengths.

4. Key Contributions

1. Controlled Benchmarking: Provided a rigorous evaluation of imaging spectroscopy using a "ground truth" dataset where environmental and structural confounders were minimized, establishing a baseline for species discrimination potential.
2. Algorithmic Insight: Demonstrated that linear models (LDA) often generalize better than non-linear models (SVM) for species classification in this context, as SVMs can overfit to individual tree variations rather than species-level spectral signatures.
3. Spectral Range Identification: Confirmed that NIR and SWIR regions are more critical for distinguishing tree species than the VIS range, likely due to structural and biochemical differences (water, cellulose, polyphenolics).
4. Scalability Limit: Identified a clear threshold where classification accuracy drops significantly as species richness increases (from 4 to 16 species), suggesting current optical remote sensing is better suited for low-diversity or functional diversity monitoring rather than fine-scale species mapping in complex forests.

5. Significance and Implications

• Functional Diversity Monitoring: While species-level classification in high-diversity forests remains challenging, the study suggests that imaging spectroscopy is highly effective for monitoring functional diversity and broad taxonomic groups (e.g., Gymnosperms vs. Angiosperms).
• Management Applications: The methodology is viable for mapping species composition in managed forests with lower diversity or known species pools (e.g., plantations, reforestation sites).
• Future Directions: The authors suggest that combining imaging spectroscopy with LiDAR (to capture structural traits) and field data is necessary to overcome the inherent information limits of reflectance spectra alone, particularly for distinguishing closely related species in complex, heterogeneous canopies.
• Data Availability: The study offers a unique, open-access dataset for further method development in ecological remote sensing.

In conclusion, the paper establishes that while airborne imaging spectroscopy can effectively distinguish up to four tree species under controlled conditions, its utility for species-level classification diminishes in species-rich environments. In such cases, the technology is more valuable for assessing functional diversity and broad community composition.