Accurate Species Distribution Modelling (SDM) is essential for biodiversity conservation, however the limited and spatially biased nature of Presence-Absence (PA) data poses a challenge. In contrast, Presence-Only (PO) datasets are abundant but lack explicit absence records. This paper examines a two step deep learning approach to combining both PO and PA data to generate an SDM. In the first step, the model was trained on a larger PO dataset, and in the second the model was then tuned on a smaller PA dataset. Results indicate that pre-training with PO data improved the performance by 7% when subsequently fine-tuned with PA data, as measured by the samples-averaged F1-score. This approach demonstrates the potential of combining diverse data types to create more reliable species distribution models for plant biodiversity conservation.
Understanding the complex patterns of plant species distribution can help in managing and protecting
species, that are rare [
Deep learning techniques, especially Convolutional Neural Networks (CNNs), have become popular for this purpose. CNNs are particularly good at identifying complex patterns in large, high-dimensional environmental data.
These models typically rely on two kinds of data. The first is Presence-Only (PO) data, which is
widely available from sources such as citizen science projects and museum collections. However, this
data only tells us where a species has been found, not where it’s absent, and it can be geographically
skewed [
Through the combination of both PA and PO data more powerful models may be developed, namely
Integrated Distribution Models. Development of these models is, however, both technically and
computationally complex, and often requires notable amounts of time and resources [
CNNs are a natural fit for SDMs because they excel at analyzing high-resolution data from remote
sensing. Architectures such as ResNet have been used efectively in modeling species distribution, and
served as key benchmarks in previous GeoLifeCLEF labs [
In previous GeoLifeCLEF labs multi-modal models, which include more than one predictor per
model, have been shown to perform well on similar tasks [
To take advantage of both Presence-Only (PO) and Presence-Absence (PA) data, a two-step deep learning
approach was used, with the goal of improving on previous models that exclusively used PA data [
This approach was tested with the GeoLifeCLEF 2025 dataset [
Ultimately, this research investigates whether pre-training a model on PO data can improve predictions when it is later trained on PA data. This paper demonstrates that this combined method will produce more accurate results than using PA data by itself.
The data for this study [
The provided data contained two main types of species information: approximately 5 million PresenceOnly (PO) sightings and 90,000 Presence-Absence (PA) survey records, all linked to specific GPS coordinates and survey ID values.
Each record is paired with a comprehensive set of environmental variables, including: • Satellite imagery (from Sentinel-2 and Landsat) • Climate data (CHELSA time series and bioclimatic variables) • Land cover maps • Human footprint indexes • Soil data
The dataset is organized to frame the task as a multi-label classification problem, where the goal is to predict the presence of various species. This research specifically drew upon three main data sources within this collection. These three data sources were selected as they were provided as PyTorch tensors in a structure that could easily be input into a CNN such as ResNet18 without additional manipulation data: • Sentinel-2 (sen): Pre-processed raster files scaled to the European continent, representing Sentinel-2 Level-2A observations. Each TIFF file corresponds to a unique observation location (surveyId). • Landsat (lan): Over 20 years of Landsat satellite imagery extracted from Ecodatacube, aggregated into CSV files and data cubes representing mean spectral band values for three months preceding each observation date. Cubes are structured as (n_bands, n_quarters, n_years) where n_bands = 6, n_quarters = 4, and n_years = 21. • Bioclimatic Cubes (bio): Four monthly CHELSA climatic rasters (precipitation, maximum temperature, minimum temperature, and mean temperature) with a resolution of 30 arc seconds, spanning January 2000 to June 2019. Cubes are structured as (n_year, n_month, n_bio) where n_year = 19, n_month = 12, and n_bio = 4.
Statistical analyses were performed to determine whether the PO and PA training and test datasets exhibit similar geographic spreads. This was done to ensure the test data’s spatial distribution is an accurate representation of the training data.
The test compared the spatial density of GPS coordinates between the Presence-Absence training data,
the Presence-Only training data, and the Presence-Absence test data. The Jensen-Shannon Divergence
(JSD) [
The following hypotheses were tested:
Null Hypothesis: The training and test datasets originate from the same geographic area, and observed spatial diferences are attributable to random chance. Alternative Hypothesis: The datasets come from diferent geographic areas, meaning there is a significant spatial mismatch between them.
To do this, a Gaussian Kernel Density Estimate (KDE) was created for the coordinates of each dataset to visualize its spatial distribution. The JSD score was then calculated to quantify the diference between these distributions. The resulting KDE heatmaps for the PA and PO datasets are shown in Figure 2.
Second, to evaluate the statistical significance of the observed JSD, a permutation test was then conducted under the null hypothesis that the training and test samples originate from the same underlying spatial distribution. The combined dataset was randomly permuted, and samples were reassigned into surrogate training and test groups matching the original group sizes. For each permutation, KDEs and the corresponding JSD were recomputed, generating a null distribution of divergence values.
Value counts for each individual species were calculated for the data in order to remove low-occurrence outlier species. These counts were performed on the PA training data, as this provides a comprehensive overview of each survey site, mitigating the possibility of certain harder to identify species being overlooked, which might occur in the PO data.
The top 1,000 species were selected for use. These represent 94.12% of the PA training data and include species with occurrences of 130 or greater within this dataset.
A CNN based on the ResNet18 architecture was trained to predict species presence from satellite imagery. The model was designed to classify species presence or absence at specified geographic locations, utilizing image tiles centered on survey coordinates as input. To optimize spatial generalization and reduce overfitting, a data strategy that combined diferent types of occurrence data while accounting for their respective strengths and limitations was implemented. 3.1. Incorporating Presence-Only Data to Supplement Sparse Presence-Absence
Observations Presence-Absence data were used as the primary target for model training and evaluation, as they provided both positive and negative labels necessary for supervised learning. However, the available PA dataset was relatively limited in size and spatial coverage. Many regions were underrepresented, including regions covered in the test data. To address this, the PA were supplemented with a larger and more spatially uniform Presence-Only (PO) dataset, which, although lacking explicit absence information and label balance, ofered broader geographic coverage.
The PO data were used to augment the spatial diversity of the training inputs, exposing the model to a wider range of environmental and landscape contexts associated with species presence. While PO data could not be used for direct training on absence, they contributed to pretraining and representation learning stages, improving the model’s ability to extract ecologically meaningful spatial features. This combined data approach allowed the model to benefit from both the accuracy of PA labels and the spatial representativeness of the PO observations.
This study employs a multi-label classification approach to predict species presence based on the
environmental predictors described above. ResNet18 [16], a commonly used convolutional neural
network architecture, was selected as the base model for all experiments. This model was chosen for its
eficiency and the ability to train it on consumer hardware. The code developed for this experiment
was adapted from the baselines provided as part of the Kaggle competition [
All models were trained using an RTX 3090 GPU with 24GB VRAM. This use of consumer grade hardware influenced the choice of model architecture for smaller networks.
For this research, it was necessary to make some adjustments to the ResNet18 model. Model for processing the Sentinel-2 data: • Used pre-trained weights IMAGENET1K_V1 • The first convolutional layer was modified from 3 to 4 channels to accommodate the Infrared spectrum imagery.
Model for processing the Bioclimatic Cubes data:
• Used randomized initial weights
• Input channels were set to [
Model for processing the Landsat data:
• Used randomized initial weights
• Input channels were set to [
The performance was assessed using the samples-averaged F1-score. This metric quantifies the degree of overlap between the species predicted to be present at each location and the actual species observed (ground truth). Submissions were then created consisting of a list of predicted species IDs for each test survey. These were compared against the known set of species present at that location and time, which was stored on the Kaggle platform. The F1-score was calculated to provide an evaluation of the model’s predictive accuracy.
Where: 1 =
1 ∑︁
=1 + ( + )/2 = Number of predicted species truly present = Number of predicted species that are absent = Number of present species not predicted (1)
The hyperparameters used in training the models are shown in Table 2. These were adopted from
previous research [
As shown in Figure 3 all three models pre-trained on PO data outperformed their equivalent models trained on just PA training data with random weight initialization. Training solely on PO data (the first 10 epochs of the 10 PO + 10 PA results in Figure 3) proved insuficient to produce a useful model, as the PO-trained Bioclimatic Cubes, Sentinel-2 models did not outperform a simple top-25 most frequent species prediction and the Landsat outperformed this baseline (0.08372) with a samples-averaged F1score of 0.09374. However, when subsequently fine-tuned with the PA training data, the three models demonstrated an approximate 7% increase in performance, as measured by the F1-score.
Looking at the individual models, the Landsat model has the highest F1-score in both PO+PA and PA-only training. Further, the Landsat PA-only model outperforms both the Bioclimatic and Sentinel PO+PA models. Between the two versions of the Landsat model, an improvement is also seen in the PO+PA model, which outperforms the PA-only model, resulting in an F1-score of 0.1784.
Pre-training our three baseline ResNet18 models with PO data resulted in improved performance for all three models. This improvement was observed despite the demonstrated diferences between the PO and PA datasets. Notably, substantial geographical areas existed where the PA training and test data exhibited no overlap. The diferent data coverage for Switzerland, a country with very distinctive terrain, is shown in Figure 4. It is largely covered in the PA test data, but is only sparsely covered in the PA training data. It is, however, well covered in the PO data.
Due to the disparity in data size, training the model on PO data took longer per epoch than training on PA data. This cost in time would be a factor in choosing to use PO data, however, due to the eficiency of ResNet18, this did not push the hardware requirements beyond what can be delivered by modern consumer hardware. Furthermore, as seen in Figure 3 the models did not exhibit improvement in performance over extended training and the number of epochs could be reduced without sacrificing performance.
The PO data could, in theory, allow these trained models to generalize better to areas not adequately covered by PA training data, but this would require further analysis of the data, and access to the ground truth labels for the test set.
Due to resource limitations, limited hyperparameter tuning was performed. Further tuning is likely to result in improved performance, and when combined with more complex pre-training/fine-tuning methods, could result in a more accurate model.
Future research would be assisted by the addition of more data related to the plant species, such as family or genus information. This would allow for a deeper analysis of plant groups where the models underperform.
This research has shown that Presence-Only data can still improve predictive multi-label classification models, when supported by accompanying Presence-Absence data. Improved predictions could allow for better planning in protecting biodiversity.
During the preparation of this work, the authors used Gemini 2.5 in order to: Grammar and Spelling Check; and Improve writing style. Further, the initial reviews used ChatGPT-4o for: Peer review simulation, the results of which were taken into account in later drafts. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [16] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, 2015. URL: https: //arxiv.org/abs/1512.03385. arXiv:1512.03385.