🌿 ecology

Assessing the impact of data uncertainty on remote sensing estimation of plant functional diversity

Using simulations from the BOSSE experiment, this study reveals that random errors impact plant functional diversity estimates more than systematic errors, demonstrating that standardizing metric computation and quantifying uncertainties separately at the reflectance level are crucial for optimizing remote sensing assessments of plant functional diversity.

Original authors: Pacheco-Labrador, J., Rossi, C., Santos, M. J.

Published 2026-09-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pacheco-Labrador, J., Rossi, C., Santos, M. J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

To understand how healthy an ecosystem is, scientists often look beyond just counting the number of species. They examine the variety of jobs those species perform, such as how efficiently they capture sunlight, how much water they hold, or how they are built. This concept, known as plant functional diversity, acts as a vital link between the variety of life on Earth and the way ecosystems function, from storing carbon to regulating water. In recent years, satellites equipped with advanced cameras have begun to map these traits from space, offering a way to monitor the planet's biological health on a global scale. However, just like any measurement taken from a distance, these satellite observations are not perfect. They are subject to various types of noise and errors that can distort the picture. The critical question for scientists and policymakers is whether these imperfections in the data are significant enough to make the resulting maps of plant diversity unreliable.

A team of researchers set out to answer this question by simulating a controlled environment where they could introduce specific types of errors and watch exactly how they affected the final results. Using a sophisticated virtual generator, they created hundreds of synthetic landscapes, each populated with different combinations of plant species and traits, ranging from simple clusters to evenly distributed forests. These virtual scenes included detailed maps of plant characteristics, such as leaf thickness and pigment content, as well as the corresponding light signals that a satellite would see bouncing off the ground. To test the limits of accuracy, the researchers then deliberately added different kinds of "noise" to their data. They introduced systematic errors, which act like a consistent bias shifting every measurement in the same direction, and random errors, which act like static or fuzz that varies unpredictably from pixel to pixel. They also tested how these errors behaved when they were correlated, meaning the mistakes in one part of the image were linked to mistakes in neighboring areas or related plant traits.

The researchers then calculated the diversity of the plant communities in these simulated worlds using two different approaches. First, they calculated the diversity based on the "true" plant traits they had programmed into the simulation. Second, they calculated the diversity using the noisy, error-filled data that mimicked what a satellite would actually observe. They compared these two calculations to see how much the errors had changed the story the data was telling. The results revealed a clear pattern: random, unpredictable noise had a much stronger negative impact on the accuracy of the diversity estimates than consistent, systematic biases did. In fact, the random noise could distort the diversity estimates by up to three times more than the systematic errors. Interestingly, when the errors were correlated—meaning the mistakes were linked across space or between different plant traits—their damaging effect was actually reduced. This suggests that the way errors spread through the data can sometimes cancel each other out, rather than simply adding up to a larger mess.

The study also examined which types of satellite data were most robust against these errors. The researchers found that using the raw, detailed light signals directly from the satellite was more reliable than using simplified summaries or derived numbers, such as specific color combinations designed to highlight chlorophyll or water. As the data processing became more complex, moving from raw signals to calculated traits, the impact of uncertainty grew larger. However, the team discovered a powerful tool for managing these errors: standardization. By mathematically adjusting the data to remove consistent biases before calculating diversity, they could effectively neutralize the impact of systematic errors. This process made the satellite estimates much more comparable to the ground truth. Yet, a surprising twist emerged when they considered the errors in the ground measurements themselves. If the ground data also contained random noise, it sometimes appeared to improve the match between the satellite and ground estimates. The researchers caution that this is a deceptive improvement; both sets of data are actually biased, and the apparent agreement is a statistical illusion rather than a sign of true accuracy.

Ultimately, this work suggests that the way we handle uncertainty in plant diversity mapping needs to be different from how we handle it when measuring a single average trait. The study indicates that random noise is the primary enemy of accurate diversity maps, while consistent biases can often be corrected through standard mathematical adjustments. It also highlights that the more complex the satellite product becomes, the more vulnerable it is to these errors. For the future of monitoring biodiversity from space, the findings suggest that scientists must carefully quantify and separate different types of errors in their data. By understanding exactly how noise propagates through their calculations, they can design better methods and set clearer quality standards, ensuring that the maps used to guide conservation and policy are as trustworthy as possible.

Technical Summary: Assessing the Impact of Data Uncertainty on Remote Sensing Estimation of Plant Functional Diversity

Problem Statement
Remote sensing mapping of plant functional diversity (PFD) is increasingly utilized at the science-policy interface to link biodiversity with ecosystem functioning. However, rigorous assessments of uncertainty and quality standards for these products are rarely provided. While uncertainty in plant trait estimation is well-recognized, the specific impact of measurement uncertainty on PFD estimates—defined as the spatial variability of traits across pixels—remains poorly understood. Existing literature often fails to explicitly consider how uncertainties propagate through the processing chain (from digital numbers to orthorectification and atmospheric correction) into diversity metrics. This omission limits the ability to evaluate the robustness of remote sensing methods, select appropriate diversity measurements, and assess conformity to application requirements for policymaking. Specifically, it is unclear which remote sensing proxies are most robust to uncertainty, how joint uncertainties in remote sensing and in-situ data interact, and whether preprocessing can attenuate or inflate these effects.

Methodology
The study utilizes the Biodiversity Observing System Simulation Experiment (BOSSE) to generate synthetic datasets that encompass plant traits and remote sensing proxies. The workflow involves four main stages:

Simulation: 180 synthetic scenes (120x120 pixels) were generated across four climatic zones and three spatial patterns (clustered, intermediate, even) using four dates per scene (totaling 720 pairs of trait maps and hyperspectral reflectance images). These scenes simulate 12 plant traits (e.g., leaf chlorophyll, water content, LAI) and corresponding reflectance factors (R) based on the EnMAP spectral configuration (197 bands).
Uncertainty Injection: Eight distinct types of uncertainty were simulated and applied to both the "in-situ" plant trait (PT) maps and the reflectance factors (R). These types were categorized by:
- Error Nature: Systematic (bias) vs. Random (noise).
- Magnitude Scaling: Constant (independent of value) vs. Scaled (proportional to value).
- Correlation: Decorrelated, trait-correlated (t-correlated), and spatially correlated (x,y-correlated).
  Uncertainties were generated using Monte Carlo approaches with Uniform distributions for systematic errors and Gaussian distributions for random errors, with a standard deviation set to 5% relative uncertainty.
Derivation of Proxies: From the perturbed reflectance factors, secondary proxies were computed: Spectral Vegetation Indices (NDVI, NIRv, and a cube of three indices, SVI3) and Optical Traits (OT), which are reflectance-based estimates of plant traits retrieved via a neural network.
Diversity Estimation and Evaluation: Rao's quadratic entropy (QRao) was computed for all variables (PT, R, SVIs, OT) using a 3x3 moving window. The impact of uncertainty was evaluated by regressing QRao derived from remote sensing proxies against QRao derived from the "true" (uncertainty-free) plant traits. Performance was assessed using $R^2$ , Pearson correlation ( $r_{Pearson}$ ), relative root mean squared error (RRMSE), mean absolute error (MAE), slope, and bias. Evaluations were conducted at two scales: "between sites" (regional/global means) and "within sites" (local spatial variability).

Key Contributions and Results

Dominance of Random Errors: Random errors (noise) significantly impact PFD estimation more than systematic errors (bias). Systematic uncertainties had negligible effects on PFD estimates, largely because the standardization and dimensionality reduction (PCA) applied during QRao calculation effectively cancel out mean biases.
Impact of Proxy Complexity and Uncertainty:
- Performance under No Uncertainty: Both Reflectance factors (R) and Optical Traits (OT) achieved the best estimates of PFD when no uncertainty was present.
- Sensitivity to Uncertainty: Despite their high baseline performance, OTs (along with SVI3) were the most affected by the introduction of uncertainty. Random errors, particularly decorrelated ones, led to the largest overestimation for OTs across all remote sensing proxies.
- Spectral Indices: Unidimensional indices like NDVI and NIRv underestimated PFD but were less affected by uncertainty types than complex proxies; however, random errors partially compensated for their inherent underestimation.
Role of Correlation: Introducing correlation in random errors (both trait-correlated and spatially correlated) reduced the overestimation of PFD caused by decorrelated white noise. This suggests that error correlation can attenuate the propagation of uncertainty.
Spurious Error Cancellation: When uncertainty was introduced into both the in-situ plant traits and the remote sensing proxies, the resulting statistical metrics (e.g., $R^2$ , RMSE) sometimes improved. The authors identify this as a "spurious" effect where uncertainties in the ground reference data coincidentally counteract biases in the remote sensing estimates, masking the true performance degradation.
Preprocessing Importance: Standardization and dimensionality reduction are critical not only for linking field and remote sensing data but also for mitigating systematic uncertainties. Omitting these steps significantly reduced performance, especially for high-dimensional data like hyperspectral reflectance.

Significance and Claims
The paper claims to provide a new understanding of how uncertainty propagates into PFD estimates, distinguishing this process from traditional mean trait estimation. The authors assert that:

Quality Requirements: Random and systematic uncertainties must be quantified separately at the reflectance factor level to guide their use and interpretation in PFD analysis.
Methodological Optimization: The findings suggest that standardization is a necessary step to reduce systematic bias in PFD estimation, potentially reducing the need for uniform image-correction processes (like empirical line corrections) in some contexts.
Proxy Selection: While hyperspectral reflectance (R) and Optical Traits (OT) offer the best baseline performance, the trade-off involves computational cost and sensitivity to uncertainty. The study highlights that derived proxies (indices, optical traits) accumulate uncertainty, with complexity acting as a multiplier for error impact, particularly for OTs which showed the greatest variability across bootstrap realizations.
Validation Challenges: The study underscores the difficulty in validating remote sensing PFD estimates when ground data also contains uncertainty, as this can lead to misleadingly positive statistical correlations.

The authors conclude that while simulations offer valuable insights, future work requires extensive observational datasets with explicit uncertainty budgets (separating systematic and random components) to fully define the limitations and quality requirements of remote sensing products for plant functional diversity.

More like this