Our Challenge
Humanity is at a turning point. While 55% of global GDP (approximately $58 trillion) is dependent upon nature, we continue to extract and destroy natural resources at an unsustainable rate. Since 1970, we’ve lost nearly three-quarters of all wildlife on Earth. Half of all species populations are in rapid decline, and one in three freshwater species now face extinction. Many people don’t understand that preserving biodiversity is essential to maintaining the complex ecosystems upon which we all rely – from climate stabilization and water provisioning to crop pollination and bioeconomy supply chains. This is why CEOs now consider biodiversity loss and ecosystem collapse to be amongst the top three risks facing the global economy over the coming decade.
Fortunately, the tides have turned. The world’s governments came together and boldly committed to protecting and conserving 30% of the planet by 2030, as part of a sweeping set of targets in the UN Kunming-Montreal Global Biodiversity Framework (GBF), which aims to build a nature-positive future for all. Concurrently, the private sector is starting to mobilize. A new report finds that shifting to a nature-positive economy could unlock over $10 trillion in business value. 733 firms have now committed to nature-related reporting under TNFD, including asset owners and managers with over $22.4 trillion under management and publicly-listed companies with a total market cap of $9.4 trillion. The central banks may soon consider tighter rules to address nature-related financial risks, and over 150 companies are now developing internal targets to adopt nature-positive business practices through SBTN.
Source: Nature Positive Initiative (NPI), 2023.
But we’ve got a significant problem. To take advantage of this unprecedented momentum, we need a robust nature ‘data ecosystem’ of well-funded scientific partners delivering a pipeline of high-quality data to decision makers in both the public and private sectors. Stakeholders urgently need a common “source of truth” to properly document the ‘State of Nature’ (SON) across an array of geographic domains. A cohesive integration of data providers has yet to be formed, and there is still no final consensus on how to measure SON. Progress has been made by the Nature Positive Initiative (NPI), which recently released a comprehensive set of draft metrics for review, organized by 9 indicators (detailed in Our Thesis). However, as we unpack the demands of each of the metrics, the reality sets in about the magnitude of work ahead that's needed to deliver high quality nature data at scale.
The view on the ground
As NPI explains, “there are an overwhelming number of ways to measure nature.” In the NPI consultancy phase, over 600 metrics were identified across thousands of peer-reviewed studies and white papers. Unlike the climate sector – which benefits both from a single common unit of measurement (1 tonne of CO2e) confined to one domain (the atmosphere) and a well-organized, well-funded knowledge system centered around the Intergovernmental Panel on Climate Change (IPCC) and the CMIP process – the nature sector relies upon a myriad of metrics and models across 5 domains (subterranean/soil, marine, freshwater, terrestrial, and troposphere). Recently there have been calls for a comparable BMIP process for biodiversity, but several core challenges regarding the collection and integration of data have to be overcome.
First, the biodiversity data landscape is highly segmented and complex, with a myriad of data sources – from academic research projects, NGO-led monitoring programs, and citizen science efforts (e.g. iNaturalist and eBird) to a plethora of start-ups selling novel nature and biodiversity data to support new corporate transparency policies or the nascent biocredit market (e.g. Biodiversity Net Gain in the UK). Second, the ad hoc nature of data collection means that areas with the most rare and threatened species have the least coverage given they are often found in remote (or dangerous) locations, making data highly biased to northern countries. And third, the heterogeneity of in situ sensors – everything from camera traps and bioacoustic sensors to chemical detection methods like eDNA or soil sampling – combined with the growing array of satellite sensors, makes data integration very challenging, requiring a high level of advanced scientific knowledge, compute time, and continual testing and validation of results.
To make matters worse, none of these critically important efforts are sufficiently funded. A 2016 study (the only of its kind) examined the costs associated with the “holy trinity” of biodiversity data providers – Key Biodiversity Areas, Protected Planet, and the IUCN Red List of Threatened Species – and found that total support for these three programs amounted to only ~$10M per annum (in 2025 inflation-adjusted dollars), with approximately 56% of these funds covered by private philanthropy (either through direct grants or via NGO programmatic expenses). The study found that all of these research efforts were greatly underfunded – and would require at least a doubling of annual expenditures on top of a large capital expenditure of $100M+ to meet critical baselines of data coverage.
Sources of funding for the three leading biodiversity data sets c. 2014, with totals in 2025 inflation-adjusted dollars, including the dollar-equivalent of volunteer time (roughly 8% of the total), adapted from Juff-Bignoli et al. 2016.
It’s important to point out that this trinity of biodiversity data sets is just the tip of the iceberg. They are the culmination of hundreds of thousands of field observations and reporting activities in every country, all of which require their own funding sources. These are aggregated in the Global Biodiversity Information Facility (GBIF), an international network and data facility “funded by the world's governments” to provide open access to data covering all types of life on Earth. This sounds impressive until you realize that the majority of observations submitted to GBIF since 2020 are from just two citizen science platforms – iNaturalist and eBird. These are effectively self-funded platforms, kept afloat by donations from the users themselves and a small group of charitable organizations. With its repository of over 200M observations, iNaturalist is now the leading force in biodiversity research, providing roughly half of all image records on GBIF, ranking second to eBird in audio records, and covering over 325,000 species.
Contributions to GBIF from iNaturalist versus other data publishers. The largest quantity of data contributions to GBIF over the past decade is from eBird, though these observations are limited to a relatively small number of bird species (< 10,000). A group of publishers contributing camera trap observations (e.g. Smithsonian Institution) increases coverage to approximately 60,000 species. Beyond these species, iNaturalist expands coverage to over 325,000 species documenting multiple vertebrate, invertebrate, and plant taxa. Adapted from iNaturalist, 2024.
The 3,000,000 foot view
What about sensors in orbit? With over 1100 functioning Earth Observation satellites (1129 have been launched since 2019 and typically have a 7-year lifespan), the question naturally arises, don’t we have enough remote sensing data to determine ‘State of Nature’ globally? The resounding answer to these questions is “No”. Despite a plethora of increasingly advanced instruments encircling our globe (producing over 2 petabytes of data per day), we don’t yet have data of sufficient quality or data integration methods to confidently attribute land cover characteristics or ecosystem dynamics at precise scales. Our analysis of 14 leading global land cover models found that none are able, for example, to effectively distinguish between native grassland and cropland, natural forests and tree plantations, or freshwater and saline water – precisely the characteristics we need to understand in order to build a comprehensive information system defining SON.
Artemis 1 Orion capsule Nov. 16, 2022, composite with ‘Earth from Afar’ (Image credits: NASA).
There are certainly some exciting breakthroughs in remote sensing technology on the horizon. We supported early work on laser-guided imaging spectroscopy (LGIS) led by Greg Asner at Arizona State University, which measures how 21 plant foliage chemicals interact with sunlight at very high resolution (~30 cm). This process enables the remote detection of specific tree species, creating maps that reveal ‘beta diversity’ in a given geographic area – the ratio of unique species found in one area versus those found in adjoining areas (aka “species turnover”). High beta diversity becomes a direct measurement of ecosystem complexity, one aspect of biodiversity, which gives us a more refined understanding of the importance of a region beyond its ‘alpha diversity’ – the total number of species found in the region (aka “species richness”).
While LGIS required airplane-mounted instrumentation, Asner and his team at Planet Labs recently launched the Tanager mission, which brings advanced hyperspectral imaging into orbit. This mission, however, is primarily tasked with observing methane emissions and doesn’t yet provide the direct observation of beta diversity that we have dreamed of. The European Space Agency (ESA) and the European Union are now set to launch two satellites as part of the CHIME mission, which will carry advanced hyperspectral instruments detecting “chemistry from space” at medium resolution (~30 m), specifically designed for soil, land cover, and biodiversity observation. This could be a game-changer, but it’s a ways off. The first satellite is not expected to launch until late 2029.
Map of beta diversity in a Peruvian forest, showing the quantity and distribution of unique tree species assemblages, or beta diversity. Courtesy of Greg Asner, ASU.
Given the uneven and sporadic availability of ground observations on biodiversity, and the lack of accurate maps of land cover and ecosystem condition due to limited satellite imagery capabilities, those working in the nature data space have had to rely on synthetic models and composite metrics to glean information about the condition of ecosystems and biodiversity habitats globally. Area of Habitat (AoH), derived from IUCN range maps (assembled from GBIF data), is used to produce Species Richness and Rarity-weighted Richness maps (per IUCN), which in turn feed synoptic models like the Biodiversity Intactness Index (BII), the Ecosystem Integrity Index (EII), and more advanced products like Map of Life (MOL). These then become inputs for global prioritization models, such as NatureServe’s MoBI project, and the ‘Areas of Global Importance’ optimization led by IIASA (Jung et al. 2021).
A new paper exploring incentives for biodiversity data production, “Models and metrics aren’t enough..” lays out our current challenge plainly:
Biodiversity loss is accelerating, and our ability to meet public, private, and international commitments to halt and reverse this decline depends on access to high-quality information about biodiversity. Policy effectiveness is constrained by data quality, and advances in modeling cannot compensate for gaps, biases, and misrepresentation in underlying observations. Our biodiversity information landscape is currently fractured, geographically biased, and systematically incomplete… and the scale of biodiversity monitoring required to meet global commitments substantially exceeds current capacity from public funding and nonprofit infrastructure (Hulkund et al. in review).
So… can’t AI just fix it?
While Large Language Models (LLMs) and Agentic AI tools have recently captured the public’s imagination, Earth Observation scientists have been working on advanced machine learning techniques, like Convolution Neural Networks (CNNs), for well over two decades. This community knows far too well that you cannot model your way out of bad data. Even if we look at LLMs, which have received hundreds of billions of dollars in capital investment over the past few years, we see persistent issues like the Strawberry Problem and the inability for LLMs to form working memory. LLM hallucinations, either by commission or omission, are a growing challenge. And for those worried about AI taking over all white collar jobs, they might find comfort in a new paper just released by Microsoft Research that shows even well-trained LLMs, when delegated a simple a task, consistently introduce compounding errors, resulting in an average corruption rate of 25% for long workflows. If LLMs, arguably the best funded of AI technologies, continue to struggle with basic linguistic tasks, one can imagine the exponentially greater challenges presented by processing large amounts of quantitative data.
For quantitative data processing, one major business category has received considerable AI investments – financial data services. A recent study examined the performance of AI models in ingesting and interpreting quantitative financial data, and while most did well locating and classifying financial facts in documents, they struggled to understand the semantic framework behind the data, preventing them from accurately tagging data for financial analysis. When attempting to link extracted numbers to an industry standard taxonomy, even the best models achieved only 17% accuracy. A consortium of AI researchers came to the conclusion that data inputs must be highly structured if they are to work for AI-driven routines and have just introduced an LLM-ready system for preparing data called FinTagging, which supports higher quality semantic reasoning to meet the demands of financial analysis.
This discovery is extremely relevant to Ecosystem Service (ES) modeling, an approach that utilizes data to quantify the value of specific geographic locations based on dollar-denominated contributions of those ecosystems to society – from water provisioning and climate stabilization to ecotourism and pollination. There are numerous approaches to ES analysis, and in the past the field has been greatly hindered by a lack of interoperability. Many have hoped that AI could help, but a group of scientists led by USGS has proposed that a shared semantic convention must first be developed to produce semantically enriched (tagged) and "machine-actionable" data and models. We don’t yet even have globally agreed definitions of basic concepts like “urban” or “forest”. A recent study on the extent of urban land found a four-fold difference between models, and a new global comparison of 8 leading global forest data sets found only a 26% congruence of mapped forest area. It’s clear that we still need better data organized in accordance with a commonly held taxonomy before we’re ready for AI primetime.
Spatial agreement of forest cover classifications between eight land cover datasets (Castle et al. 2026). Spatial agreement is defined as the number of datasets (out of 8 total) which designate a specific pixel as 'forest'. Full agreement between all eight datasets corresponds to a value of eight (dark green), and no agreement between the datasets corresponds to a value of 1 (dark purple). No color (gray) indicates that none of the datasets classified the area as forest. All data are from the year 2019 except for ESA-WC and GLAD-LCLUC, which are from 2020. Via Mongabay.
If AI is currently struggling with linguistic and quantitative data, the challenges facing AI researchers when it comes to reliably extracting information from geospatial data (GeoAI) is far greater. Provided either as rasters (pixels) or vectors (points) encoded with a specific location, these data can define two-dimensional boundaries, three-dimensional elevations, and change over time, mapping a wide array of characteristics. Even in scientific disciplines with a long history of utilizing advanced machine learning techniques and Earth Observation data – for example, aboveground biomass (AGB) in forest ecosystems or dynamic global vegetation modeling (DGVM) – results are still highly variable, with modeled outputs in total carbon sequestration over a 100-year period differing by 85% or more.
Yet it’s these very data that are now being used to train an ever-expanding number of geospatial foundation models (IBM’s TerraMind, Google’s AlphaEarth, and TESSERA from the University of Cambridge, just to name a few), posing significant problems. A recent analysis by the European Space Agency (ESA), which compared 12 popular global foundation models (GFMs) on tasks like land cover classification, change detection, environmental monitoring and multi-sensor and multi-temporal analysis, found only modest gains in some tasks even with enormously more compute capacity:
Today’s GFMs aren’t failing because they’re “too small.” They’re failing because they’re under‑trained relative to capacity and misaligned with pixel-wise segmentation tasks. Before burning another million dollars in AWS credits and scaling models and datasets up further, we need sharper and more relevant pre‑training objectives and more information‑dense datasets… – Christopher Ren, 2025
In other words, if we want big AI models to succeed, we have to feed them a diet of high-quality training data. And there is still a lot of fundamental science that needs to be done to provide that supply of data. To build accurate forest biomass estimates, for example, we need investment to better understand forest dynamics across diverse geographies, incorporating new scientific findings, for example:
- Biomass carbon is underestimated in temperate forests due to bias towards smaller trees in allometric models. In the Pacific Northwest, large trees only make up 3% of the forest canopy, yet they account for 42% of total carbon storage.
- Understanding wood density (WD) of various tree and shrub species is important to better carbon storage in successional forests, with some families of species containing nearly 3 times more wood density than others.
- We’re just beginning to understand the role of edge effects in greatly reducing the quantity and resilience of AGB due to forest fragmentation, potentially reducing total biomass storage by 9% globally.
- While most of the focus is on aboveground biomass, root biomass is an important component of forest carbon sequestration, adding an average of 25% to total storage estimates (lower than previous estimates).
- Mapping efforts like SPUN and others, are showing the specialized role of mycorrhizal fungi and soil microbiology in carbon sequestration, particularly in old-growth forests and savanna landscapes.
While rapid advances are being made in the GeoAI space, and agentic and quantitative AI tools show promise to help fill global shortfalls in biodiversity knowledge, significant investments are needed to structure machine-readable data before we can unlock the full power of the AI revolution. And this takes us to Our Thesis behind the Nature Data Lab.