[Paper Review] The GeoLifeCLEF 2023 Dataset to evaluate plant species distribution models at high spatial resolution across Europe

Christophe Botella, Benjamin Deneu+7 more|arXiv (Cornell University)|Aug 7, 2023

Species Distribution and Climate Change6 citations

TL;DR

The GeoLifeCLEF 2023 dataset provides a benchmark for high-resolution plant species distribution modeling across Europe, integrating 5 million presence-only records from GBIF and 6,000 exhaustive presence-absence surveys (2017–2021) with 10 m resolution satellite imagery, climatic time series, and environmental rasters. It enables rigorous evaluation of deep learning-based SDMs using spatial block holdout, offering a critical resource to reduce biases and improve transferability in fine-scale biodiversity prediction.

ABSTRACT

The difficulty to measure or predict species community composition at fine spatio-temporal resolution and over large spatial scales severely hampers our ability to understand species assemblages and take appropriate conservation measures. Despite the progress in species distribution modeling (SDM) over the past decades, SDM have just begun to integrate high resolution remote sensing data and their predictions are still entailed by many biases due to heterogeneity of the available biodiversity observations, most often opportunistic presence only data. We designed a European scale dataset covering around ten thousand plant species to calibrate and evaluate SDM predictions of species composition in space and time at high spatial resolution (~ten meters), and their spatial transferability. For model training, we extracted and harmonized five million heterogeneous presence-only records from selected GBIF datasets and 6 thousand exhaustive presence-absence surveys both sampled during 2017-2021. We associated species observations to diverse environmental rasters classically used in SDMs, as well as to 10 m resolution RGB and Near-Infra-Red satellite images and 20 years-time series of climatic variables and satellite point values. The evaluation dataset is based on 22 thousand standardized presence-absence surveys separated from the training set with a spatial block hold out procedure. The GeoLifeCLEF 2023 dataset is open access and the first benchmark for researchers aiming to improve the prediction of plant species composition at a very fine spatial grain and at continental scale. It is a space to explore new ways of combining massive and diverse species observations and environmental information at various scales. Innovative AI-based approaches, in particular, should be among the most interesting methods to experiment with on the GeoLifeCLEF 2023 dataset.

Motivation & Objective

• To address the challenge of predicting plant species composition at high spatial resolution (~10 m) and continental scale across Europe.
• To reduce biases in species distribution models (SDMs) arising from heterogeneous, opportunistic presence-only data by integrating exhaustive presence-absence surveys.
• To create a standardized, open-access benchmark dataset for evaluating and improving deep learning-based SDMs at fine spatial grain.
• To support the development of AI-driven methods that combine massive, diverse biodiversity observations with multi-scale environmental predictors.

Proposed method

• The dataset integrates 5 million presence-only records from curated GBIF datasets and 6,000 exhaustive presence-absence surveys collected between 2017 and 2021.
• Environmental predictors include 10 m resolution RGB and near-infrared satellite imagery, 20-year time series of climatic variables, and standard bioclimatic and land cover rasters.
• A spatial block holdout procedure separates 22,000 standardized presence-absence surveys for evaluation, ensuring robust transferability assessment.
• Data harmonization and preprocessing are applied to align heterogeneous observation sources and environmental variables across space and time.
• The dataset is structured to support training and evaluation of deep learning models, particularly for multi-modal input fusion.
• The benchmark is designed to test model performance on species assemblages across diverse European biomes with fine-grained spatial resolution.

Experimental results

Research questions

• RQ1 How well can deep learning models predict plant species composition at 10 m spatial resolution using fused presence-only and presence-absence data?
• RQ2 To what extent do high-resolution remote sensing and climatic time series improve the accuracy and transferability of species distribution models?
• RQ3 How do biases in opportunistic presence-only data affect model performance, and how can they be mitigated using exhaustive presence-absence surveys?
• RQ4 What is the predictive performance of AI models on species assemblages across diverse European biogeographic regions using this benchmark?
• RQ5 How does the integration of multi-temporal and multi-spectral satellite data enhance fine-scale species distribution modeling?

Key findings

• The GeoLifeCLEF 2023 dataset contains approximately 10,000 plant species, 5 million presence-only records, and 6,000 exhaustive presence-absence surveys collected between 2017 and 2021.
• The dataset includes 10 m resolution satellite imagery and 20-year time series of climatic variables, enabling high-resolution spatio-temporal modeling.
• Evaluation is based on 22,000 presence-absence surveys separated via spatial block holdout, minimizing data leakage and enabling robust transferability testing.
• The dataset is open access and designed as the first benchmark for high-resolution plant species distribution modeling across Europe.
• It enables the objective comparison of deep learning and traditional SDM approaches on a standardized, large-scale, multi-modal dataset.
• The integration of diverse data types—especially high-resolution remote sensing and time-series climatic data—provides a foundation for reducing prediction biases and improving model generalization.