🧬 biology

The predictive power of new remote sensing data for biological distributions

This study demonstrates that integrating novel remote sensing data, such as GEDI LiDAR-derived canopy height and hyperspectral foliar traits, significantly enhances the accuracy and interpretability of models predicting bird vocal activity across California's Sierra Nevada by capturing nuanced, species-specific habitat requirements.

Original authors: Laura Berman, Fabian D. Schneider, Ryan P. Pavlick, M. Zachariah Peery, Connor M. Wood, Ting Zheng, Ethan Shafron, Zhiwei Ye, Natalie Queally, Jason M. Winiarski, Anu Kramer, Philip A. Townsend

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Laura Berman, Fabian D. Schneider, Ryan P. Pavlick, M. Zachariah Peery, Connor M. Wood, Ting Zheng, Ethan Shafron, Zhiwei Ye, Natalie Queally, Jason M. Winiarski, Anu Kramer, Philip A. Townsend

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 where wildlife lives and how many individuals are present, scientists have long relied on a combination of weather patterns, the shape of the land, and maps of what plants grow where. These factors create the stage upon which animals perform their daily lives. A bird needs a specific temperature range to survive, a certain type of terrain to build a nest, and a particular kind of vegetation to find food. For decades, researchers have used satellite images to track these broad environmental conditions, creating a global picture of where species might exist. However, these traditional maps often describe the climate or the general type of forest, missing the finer details of what the plants actually look like and how they are built. As the world faces a rapid loss of biodiversity, knowing exactly where animals are and why they choose those spots has become urgent, yet gathering this information by walking through every forest is impossible.

A team of researchers set out to see if new, high-tech ways of looking at the Earth could fill in these missing details. They focused on the Sierra Nevada mountains in California, a region with a dramatic rise in elevation that hosts a wide variety of bird species. The team collected sound recordings from 553 locations across this landscape, listening for the songs and calls of 94 different bird species. Instead of just asking where a bird was found, they measured how active the birds were at each spot, using the frequency of their calls as a proxy for how much they liked the habitat. To explain these patterns, they built computer models using 129 different pieces of data about the environment. These data points fell into six groups: the climate, the shape of the land, human activity, the timing of plant growth, the physical structure of the forest, and the chemical makeup of the leaves. The researchers wanted to know if adding the new, detailed data about forest structure and leaf chemistry would help them predict bird activity better than the older, more general data alone.

The study found that the new data did indeed make a significant difference. While climate and terrain were important for all the birds, the addition of information about forest structure and leaf traits improved the accuracy of the models for dozens of species. The researchers discovered that these new data streams provided unique information that the older maps simply could not offer. For instance, two areas might have the same temperature and rainfall, but if one has a tall, dense canopy and the other has short, scrubby bushes, they will attract completely different groups of birds. The new data allowed the models to see these differences. Specifically, measurements of how tall the trees were and the chemical composition of their leaves helped explain why certain birds preferred one spot over another. This was particularly true for species that rely on specific types of trees or need particular structural features, like tall trunks for nesting or specific leaf nutrients for the insects they eat.

The researchers also found that no single type of data worked best for every bird. Each species had its own unique set of requirements, meaning the most important factor for one bird might be irrelevant for another. For example, the Red-breasted Nuthatch, a bird that lives in mature pine forests, was strongly predicted by data showing tall canopies and specific leaf chemicals found in conifers. In contrast, birds that prefer open meadows or scrublands were better predicted by data showing shorter vegetation. This highlights that to understand the complex web of life, scientists cannot rely on just one kind of map. They need a broad toolkit that includes weather, land shape, and now, detailed views of the plants themselves. The study suggests that as satellites become capable of measuring these fine details of vegetation across the entire globe, scientists will be able to create much more accurate maps of where wildlife lives. This improved understanding is essential for protecting habitats and managing land in a way that supports biodiversity, ensuring that the right environments remain available for the species that depend on them.

Technical Summary: The Predictive Power of New Remote Sensing Data for Biological Distributions

Problem Statement
Addressing the global biodiversity crisis requires accurate, scalable data on species distributions and abundances. However, direct field collection of such data is often prohibitively expensive and logistically difficult, particularly across large spatial scales or over time. While remote sensing has long been used to map climate, terrain, and broad land cover, the ecological community is increasingly asking whether emerging remote sensing datasets—specifically those providing quantitative measures of vegetation structure and physiology—offer genuinely new predictive power for modeling species distributions. A critical uncertainty remains regarding the degree of redundancy between these novel data streams (e.g., GEDI LiDAR canopy height, hyperspectral foliar traits) and established variables (climate, terrain, phenology). It is unclear if these new datasets capture unique ecological dimensions or merely reiterate information already contained in traditional layers.

Methodology
The study was conducted across the central and southern Sierra Nevada, California, a region spanning a pronounced elevational gradient from ~500 m to 3,000 m. The research integrated three primary data components:

Biological Data: Vocal activity rates for 94 bird species were derived from passive acoustic monitoring at 553 sites. Recordings were collected during dawn and dusk periods from 2021–2023 using autonomous recording units (ARUs). Species detections were processed using a customized BirdNET implementation, with vocal activity rates normalized to a 0–1 scale to serve as a continuous response variable.
Remote Sensing Predictors: The authors utilized 129 geospatial variables from the BioCube framework, grouped into six categories:
- Climate (46 variables): Temperature, precipitation, aridity, snow season length, etc.
- Disturbance (26 variables): Fire history, human modification, night lights, etc.
- Terrain (18 variables): Elevation, slope, ruggedness, soil properties.
- Phenology (21 variables): NDVI, gross primary productivity, growing season timing.
- Forest Structure (4 variables): Derived entirely from NASA's GEDI LiDAR (canopy height, mid-canopy height, biomass density, structural richness).
- Foliar Traits (14 variables): Derived from airborne hyperspectral imagery (Western Diversity Time Series), including leaf mass per area (LMA) and concentrations of nitrogen, carbon, lignin, sugars, etc.
Modeling Approach: The authors employed Extreme Gradient Boosting (XGBoost) regression to predict normalized vocal activity rates. To assess the utility of new data streams, they ran 100 replicates of models for each species using seven variable subsets: one full model (all 129 variables) and six models where one major category was systematically excluded. Model performance was evaluated using test $R^2$ (Nash-Sutcliffe efficiency) and Root Mean Squared Error (RMSE). Variable importance was quantified using Shapley Additive Explanations (SHAP) to determine which specific variables drove predictions.

Key Results

Incremental Improvement: No single variable or category improved models for all 94 species. However, the inclusion of emerging datasets significantly improved model accuracy for a substantial subset of species.
- Foliar traits significantly improved models for 39 species (median $\Delta R^2 = 0.028$ ).
- Forest structure improved models for 41–47 species (median $\Delta R^2 = 0.019$ ).
- Climate improved models for 33–35 species (median $\Delta R^2 = 0.023$ ).
Variable Importance: Climate variables contributed the most predictive power on average (~32% of model accuracy), largely due to the high number of climate variables available (46). However, forest structure and foliar traits provided outsized importance for specific species. For instance, canopy height and mid-canopy height were the single most important predictors for 13 species. Foliar starch content accounted for >30% of SHAP importance for the Golden-crowned Kinglet and Yellow-rumped Warbler.
Species-Specificity: Variable importance was highly species-specific, reflecting individual niche requirements. Generalist species showed weaker correlations with habitat variables.
Biological Interpretability: The inclusion of vegetation data allowed for more nuanced and interpretable models. For example, the Red-breasted Nuthatch model identified a preference for tall canopies (>20–40m), high leaf mass per area (LMA), and low sulfur concentrations—traits specific to mature coniferous forests. Similarly, species preferring open habitats (e.g., Western Bluebird) were strongly associated with shorter canopies and lower LMA.

Significance and Claims
The paper claims that emerging remote sensing technologies, specifically spaceborne LiDAR (GEDI) and imaging spectroscopy, provide scientists with an unprecedented ability to remotely and quantitatively measure vegetation at global scales. The study demonstrates that these datasets are not merely redundant with climate or terrain data; rather, they capture unique ecological dimensions related to vegetation structure and physiological quality that are critical for modeling fine-scale species distributions.

The authors argue that while climate variables define broad physiological niches, variables describing local vegetation conditions (structure, composition, fire history) are necessary to resolve habitat heterogeneity that broad climatic gradients cannot distinguish. The integration of these data streams improves both the accuracy and the biological interpretability of species distribution models. The paper posits that these capabilities will greatly enhance the ability to understand ecosystems at global scales, supporting the monitoring of Essential Biodiversity Variables (EBVs) and the goals of the Kunming-Montreal global biodiversity framework. The authors conclude that continued investment in upcoming satellite missions (e.g., NASA's EAGLE-VSWIR, ESA's CHIME) is empirically justified, as these will provide the global, high-temporal-resolution data necessary to fully realize the potential of functional vegetation mapping for biodiversity conservation.

Technical Summary: The Predictive Power of New Remote Sensing Data for Biological Distributions

More like this