Table  provides an overview of notable open datasets that have contributed to tree species classification, detailing their geographic coverage, size, and taxonomic diversity. For instance, the Seu Nico Forest dataset from Brazil provides geolocated samples for 228 species but is geographically constrained. Similarly, the Maraca Ecological Station dataset includes 110 species but is also region-specific. In Europe, the EUForest dataset offers broader coverage with data for 242 species. On a global scale, datasets such as Tallo provide significant taxonomic diversity, covering 5163 species across 187 families. However, these datasets lack integration with remote sensing or environmental variables, limiting their application in ecological modeling.

Advances in high-resolution imaging and lidar technologies have enabled datasets like PureForest and Individual Tree Point Clouds , which utilize aerial and point cloud data for species classification. While these datasets offer detailed structural information, they remain region-specific and lack the spectral and temporal depth of satellite-based datasets. The TreeSatAI dataset combines multi-sensor data, including aerial imagery and Sentinel-1/2, for tree species classification in Germany but covers only 20 species. Similarly, the Planted dataset focuses on only 40 planted species globally, limiting its broader applicability. Our work also complements broader biodiversity benchmarks like GeoLifeCLEF and GeoPlant . While these focus on regional, multi-lifeform plant prediction, GlobalGeoTree is the first dataset with a global scope specifically curated for over 21 000 forest tree species.

Collectively, while valuable, these existing datasets highlight the persistent need for a benchmark that synergizes global coverage, deep taxonomic information for forest tree species, and multimodal remote sensing data, a gap GlobalGeoTree aims to fill.

Vision-language models (VLMs) enable the integration of visual and textual information. Among these, Contrastive Language-Image Pretraining (CLIP) has demonstrated exceptional zero-shot transfer capabilities by jointly training image and text encoders through a contrastive learning objective, aligning image-text pairs within a shared embedding space. For tree species classification, CLIP's ability to learn from image-text pairings (such as satellite imagery and taxonomic labels) offers a path to capture complex visual and semantic relationships. Its proven zero-shot capabilities are particularly suited for addressing the challenges of identifying species within dynamic and evolving catalogs . Meanwhile, its few-shot capabilities tackle the issue of limited labeled data, a common obstacle in biodiversity research, offering an advantage over traditional supervised methods.

In remote sensing, VLMs have been applied to various tasks, including image classification, retrieval, and scene understanding, with domain-specific adaptations yielding significant improvements. For example, RemoteCLIP , the first VLM specifically tailored for remote sensing, leverages pretraining on large-scale remote sensing imagery paired with aligned text, achieving state-of-the-art performance in zero-shot classification, linear probing, and few-shot learning. Similarly, SkyCLIP , SkyCLIP-50 and CLIP-laion-RS extend the capabilities of CLIP through continual pretraining on semantically diverse remote sensing image-text pairs. These models demonstrate enhanced generalization and transferability, achieving substantial gains in tasks such as zero-shot scene classification, fine-grained classification, and cross-modal retrieval compared to the original CLIP model.

These advancements underscore the importance of domain-specific pretraining in adapting VLMs for remote sensing applications. Aligning models more closely with the unique characteristics of remote sensing tasks has demonstrated significant potential to advance progress in this field.

The GlobalGeoTree dataset provides unprecedented global and taxonomic coverage for tree species classification using remote sensing data, and the collection involved several key steps.

The filtering process results in the final GlobalGeoTree dataset, which comprises 6 263 345 high-quality samples distributed across 221 countries and regions. Approximately 70 % of the records are sourced from “Research Grade” iNaturalist observations, which require identification consensus from at least two independent contributors. Other samples are mainly from authoritative datasets available through GBIF, including National Forest Inventory (NFI) data from countries such as Sweden, France, and Colombia. The full list of source datasets and their provenance is traceable via the unique GBIF Derived Dataset DOI of GlobalGeoTree: 10.15468/dd.9qxqyy . A core feature of this dataset is its multi-granular taxonomic hierarchy, which is essential for developing models that can generalize across different levels of biological classification. This hierarchical structure is visualized in Fig. , illustrating the relationships branching from four broad functional types at the center (level 0), through 275 families (level 1) and 2734 genera (level 2), out to the 21 001 individual species at the periphery (level 3).

Each of the 6.3 million samples is structured as a single record that includes this full taxonomic lineage, its geolocation, and source metadata. This entire collection is made available as a downloadable CSV file (GlobalGeoTree.csv) for full transparency and accessibility. Table  presents several example records to illustrate the data structure. A comprehensive statistical breakdown and accessibility details of the dataset is provided in Appendix .

Each sample in the GlobalGeoTree dataset is paired with a rich set of multimodal data to enable robust modeling. Table  provides a comprehensive overview of these features, including paired Earth Observation (EO) data and auxiliary environmental variables derived from remote sensing sources.

For effective model development and evaluation, the GlobalGeoTree dataset was partitioned into GlobalGeoTree-6M and curated evaluation subsets, primarily GlobalGeoTree-10kEval.

GlobalGeoTree-6M comprises the vast majority of the samples and is specifically designed for model pretraining. This large size allows models to learn robust and generalizable representations of tree species and their associated environmental contexts.

GlobalGeoTree-10kEval is a carefully curated dataset designed to benchmark model performance across taxonomic levels and species frequency categories. To address the characteristic long-tail distribution (detailed in Appendix ), we categorized species into three groups based on sample frequency: Frequent (>1500 samples), Common (100–1500 samples), and Rare (<100 samples), as shown in Fig. .

The primary GlobalGeoTree-10kEval dataset includes 30 species from each of these three categories, resulting in a total of 90 species. The sample proportions within this evaluation set are 12 % for Rare species, 33 % for Common species, and 55 % for Frequent species, culminating in around 10 000 samples. Figure  shows the geographical distribution of GlobalGeoTree-10kEval, which spans diverse regions across the globe. This global distribution ensures that the dataset captures a wide range of ecological and environmental contexts, making it representative of real-world scenarios. By focusing on a diverse set of species with varying levels of representation, GlobalGeoTree-10kEval serves as a robust evaluation benchmark for assessing the ability of models to tackle challenges posed by the long-tail distribution of species and shifts in geographical domains.

To further evaluate model robustness and scalability across broader taxonomic scopes, we constructed two additional evaluation subsets: GlobalGeoTree-10kEval-300 and GlobalGeoTree-10kEval-900, containing 100 and 300 species per category, respectively. Crucially, all samples within these evaluation sets, were subjected to the rigorous filtering criteria detailed in Sect.  and are strictly excluded from the GlobalGeoTree-6M pretraining set to ensure fair evaluation. Details of all evaluation subsets (Appendix ) and the corresponding evaluation results (Appendix ) are also provided. Given the complexity of global tree species classification, our primary analysis focuses on the 90-species GlobalGeoTree-10kEval, which serves as a practical starting point for systematic benchmarking.

We first quantified land-cover accuracy through a visual validation study. A random subset of 300 locations was sampled from GlobalGeoTree-10kEval. For each sample, we retrieved Very High Resolution (VHR) satellite imagery to verify whether the geolocated point fell within forest or tree cover. The inspection confirmed that 98.3 % of the samples were correctly located in forested areas, while a small minority were ambiguous (e.g., forest edges), validating the efficacy of our automated filtering pipeline.

To further enhance transparency and enable large-scale qualitative validation by the research community, we developed the GlobalGeoTree Explorer, a web-based application hosted on Google Earth Engine (Fig. ). This tool allows users to visualize the spatial distribution of all 6.3 million samples, overlay them on high-resolution satellite basemaps, and inspect detailed taxonomic attributes for any individual point. The application is publicly accessible at https://ee-yangm.projects.earthengine.app/view/globalgeotree-explorer (last access: 10 February 2026).

To validate taxonomic consistency at the functional type level, we performed a cross-comparison against the Copernicus Global Land Cover Layers (CGLS-LC100, Collection 3) . Since no global dataset currently provides high-resolution distribution maps at the species or genus level, the CGLS-LC100 product represents the most detailed global baseline available, offering discrete land cover layers at 100 m resolution that distinguish between broadleaf and needleleaf forest types.

We aligned the level0 labels of GlobalGeoTree with CGLS classes by aggregating them into two primary leaf types: Needleleaf and Broadleaf. To ensure global representativeness and mitigate regional data density biases, we employed a stratified sampling strategy, selecting 10 000 samples per continent.

Figure  presents the results of this cross-validation. The confusion matrix (left) reveals an overall agreement of 80.38 % at the leaf level. We observed high consistency for Broadleaf samples (81 %), while some Needleleaf samples (22 %) showed discrepancies with the coarser 100 m product, likely due to mixed pixels in transition zones. The regional breakdown (right) demonstrates exceptionally high agreement in South America (99.3 %) and Africa (98.7 %), where broadleaf forests dominate. Lower agreement rates in North America and Europe are expected given the higher prevalence of mixed forests and the resolution gap between our point-based data and the 100 m validation raster.

Beyond spatial validation, taxonomic reliability is ensured through our strict data provenance pipeline. We exclusively utilize observations which require community consensus for species identification, and 100 % of these samples have passed the GBIF taxonomic backbone check .

To establish a strong baseline on the GlobalGeoTree dataset, we developed GeoTreeCLIP, a vision-language model (VLM) specifically tailored for tree species classification. Our choice of the VLM paradigm was motivated by the inherent challenges of this task: preliminary experiments showed that traditional supervised models struggle with the vast, long-tailed, and hierarchical label space of GlobalGeoTree, especially at the fine-grained genus and species levels (see Sect. ).

In contrast, a VLM framework treats taxonomic labels not as discrete indices, but as structured text. This allows the model to leverage the rich semantic relationships between species, genera, and families. By jointly training on image-text pairs, GeoTreeCLIP learns powerful, transferable representations that align visual features with their corresponding taxonomic context. This approach is particularly effective for zero-shot and few-shot scenarios, which are critical for addressing the immense scale of global biodiversity and the practical impossibility of collecting exhaustive data for all taxa .

As shown in Fig. , the GeoTreeCLIP model architecture consists of the following components:

Visual Encoder. A ViT-B/16 backbone adapted to process the multi-spectral, time-series Sentinel-2 data. To handle the 10-channel, 5×5 pixel input patches, we modify the initial patch embedding layer of the ViT. Each 5×5 patch is first upsampled to 16×16 pixels using bicubic interpolation. The modified embedding layer then processes these 10 channels and projects the entire 16×16 patch into a single visual token. This approach allows the model to capture the holistic spectral information of the patch at each timestep. These 12 monthly tokens are then fed into a temporal attention module to learn and fuse key phenological patterns across the year. This mechanism allows the model to dynamically weigh the importance of different months, implicitly learning which parts of the phenological cycle are most informative for identifying a particular species.

We evaluated GeoTreeCLIP against two advanced pretrained vision-language models: CLIP and RemoteCLIP , a specialized VLM for remote sensing applications. To ensure a fair comparison with models not explicitly designed for time-series data, we employed an ensemble-like approach for CLIP and RemoteCLIP. Specifically, features or probabilities were computed independently for each of the 12 monthly images, and the final results were obtained by averaging these values. Additionally, the results of other baseline models, including SkyCLIP-50 and CLIP-laion-RS , are provided in Appendix .

All models were evaluated on the GlobalGeoTree-10kEval benchmark using zero-shot and few-shot learning settings. Performance was measured using top-1 and top-5 prediction accuracy, with separate evaluations for each taxonomic level (family, genus, and species). To ensure robustness, we repeated each experiment 5 times using different random seeds and reported the mean accuracy and variance.

To further contextualize the performance of our contrastive learning-based GeoTreeCLIP model, we conducted an additional experiment using a traditional supervised learning paradigm. This allows for a more direct comparison of learning objectives (contrastive vs. supervised) while keeping core architectural components and training settings as consistent as possible.

To qualitatively assess and compare the learned feature representations from different modeling paradigms, we performed t-SNE visualizations . We used zero-shot image embeddings (or the final fused visual-auxiliary features for SupervisedGeoTree before the classification heads) extracted from a subset of the GlobalGeoTree-10kEval-300 dataset. This analysis includes our proposed GeoTreeCLIP, the CLIP pretrained by OpenAI, and the SupervisedGeoTree model. Given the large number of classes and the hierarchical nature of the labels, we adopted a selective visualization strategy: first examining embeddings at the family level for five randomly selected families (Anacardiaceae, Berberidaceae, Cactaceae, Hernandiaceae, and Rosaceae); then focusing on five genera within the Rosaceae family (Prunus, Pyrus, Rhodotypos, Rubus, and Spiraea); and finally, visualizing five species within the Prunus genus (Prunus avium, Prunus caroliniana, Prunus ilicifolia, Prunus laurocerasus, and Prunus obtusata).

The comparative t-SNE visualizations are presented in Fig. . Across all three taxonomic levels (Family, Genus, and Species, shown as columns), GeoTreeCLIP (top row) consistently demonstrates the most effective separation and formation of distinct clusters. At the family level (left column), GeoTreeCLIP clearly distinguishes between the selected families. The CLIP (middle row) exhibits considerable overlap, particularly for the diverse Rosaceae family. SupervisedGeoTree (bottom row) shows some separation but less defined clusters compared to GeoTreeCLIP, with Rosaceae still forming a very broad distribution.

This pattern of superior clustering by GeoTreeCLIP continues at the genus level within Rosaceae (middle column). GeoTreeCLIP forms relatively distinct groups for genera like Prunus, Pyrus, and Rhodotypos. Both CLIP and SupervisedGeoTree struggle more, with CLIP showing a highly condensed and overlapping structure, while SupervisedGeoTree offers some separation but with less clarity than GeoTreeCLIP. The most striking difference is observed at the species level within the Prunus genus (right column). GeoTreeCLIP achieves remarkable separation, forming visually distinct clusters for each of the five Prunus species. In contrast, both CLIP and SupervisedGeoTree largely fail to differentiate these closely related species, with their embeddings heavily intermingled.

These visualizations qualitatively affirm that GeoTreeCLIP, through its contrastive vision-language learning approach tailored with hierarchical taxonomic information and domain-specific data, learns more semantically meaningful and discriminative representations across all taxonomic ranks. This aligns with its superior quantitative performance in classification tasks compared to both general-domain VLMs and a traditional supervised approach.

To quantify the contribution of different input modalities, we conducted an extensive ablation study. We trained several variants of GeoTreeCLIP (for 10 epochs on GlobalGeoTree-6M) and evaluated their zero-shot performance on the GlobalGeoTree-10kEval benchmark. The results are summarized in Table .

The results lead to several key insights. First, the full input consistently outperforms all ablated versions, highlighting the synergistic effect of integrating spectral-temporal imagery with environmental context. Second, comparing the “Image Only (10-band)” and “RGB Only” variants reveals the critical importance of non-visible bands (e.g., NIR, SWIR) for vegetation analysis, as they provide a substantial performance boost (e.g., 15.83 % vs. 6.38 % at the family level). Third, adding geolocation to RGB imagery (“RGB + Geolocation”) improves performance over “RGB Only”, confirming that spatial priors help capture species' geographical ranges. The reason “Image Only (10-band)” outperforms “RGB + Geolocation” is that the rich information from the 7 additional spectral bands is more discriminative than the signal provided by geolocation alone.

Our data collection process was designed to adhere to ethical standards for data reuse. All geolocated occurrence records sourced from GBIF were filtered to include only those with permissive licenses (e.g., CC0, CC-BY) explicitly chosen by the original data contributors. We also respected the privacy safeguards of the source platforms; for instance, the coordinates of sensitive or threatened species are often automatically obscured by platforms like iNaturalist, and our dataset preserves these privacy-preserving modifications.

The Sentinel-2 L2A data were accessed via Google Earth Engine. The auxiliary bioclimatic, geographic, and soil data were sourced from WorldClim , USGS SRTM , and SoilGrids , respectively, all of which are publicly available.

While GlobalGeoTree represents one of the largest and most comprehensive datasets of its kind, users should be aware of its inherent limitations. As with large part of datasets derived from citizen-science observations, it contains geographic and taxonomic biases. Data coverage is higher in regions with active observer communities (e.g., North America, Europe) and is skewed towards more common or easily identifiable species, resulting in a long-tail distribution. Furthermore, since we assimilated available official inventory data (e.g., NFIs) into the training set to maximize coverage, independent external validation remains challenging, and shared observation biases may influence evaluation metrics. The ambiguous boundary between trees and shrubs in botanical classification further complicates the dataset, as some samples may represent shrubs rather than trees. Tree species taxonomy is also subject to frequent revisions driven by new genetic evidence, which may misalign dataset labels with updated classifications over time. Although CLIP-based models can identify unseen species to some extent, such taxonomic shifts may still affect model interpretability and evaluation consistency. Additionally, the dataset relies on Sentinel-2 data from 2020, restricting its ability to capture long-term vegetation dynamics or recent disturbances. Future versions incorporating multi-year observations could better account for phenological changes and climate-driven impacts.

GlobalGeoTree holds significant potential for advancing forest monitoring, biodiversity conservation, and climate change mitigation. Beyond benchmarking, the dataset and GeoTreeCLIP model support practical operational workflows. For large-scale mapping, users can employ a sliding-window inference approach on Sentinel-2 tiles to generate pixel-wise species maps. To handle prediction uncertainty in diverse ecosystems, we recommend a hierarchical inference strategy, where the model defaults to genus or family-level classifications if species-level confidence falls below a threshold. Furthermore, to address regional domain shifts, the pretrained model serves as a robust foundation that can be efficiently fine-tuned with small sets of local field data. This “local adaptation” workflow allows practitioners to leverage global knowledge while tailoring predictions to specific forest management needs.