Overview of GeoLifeCLEF 2022: Predicting Species Presence from Multi-modal Remote Sensing, Bioclimatic and Pedologic Data

Titouan Lorieul, Elijah Cole, Benjamin Deneu, Maximilien Servajean, Pierre Bonnet, Alexis JolyCLEF (Working Notes) 2022 · OpenReview

Introduction

The GeoLifeCLEF 2022 challenge represents a significant advancement in Species Distribution Modeling (SDM), bridging the gap between traditional ecology and modern deep learning. By pairing 1.6 million citizen-science observations with high-resolution remote sensing imagery, the competition explored how multi-modal data can improve our ability to predict the geographic distribution of plants and animals. The findings suggest that while traditional climatic data remains a staple, deep learning models leveraging satellite imagery provide critical local context that significantly enhances predictive accuracy.

The Challenge of Species Distribution Modeling

Predicting where species live is fundamental to conservation biology. Historically, researchers have relied on Species Distribution Models (SDMs) built from coarse, low-resolution environmental data, such as average annual temperature or soil pH. However, these kilometer-scale variables often fail to capture the local micro-topography or specific land cover—such as a small patch of forest or a rocky outcrop—that determines a species' actual presence.

GeoLifeCLEF 2022 addressed two primary technical hurdles in this field:

  1. Data Resolution: Transitioning from kilometer-scale climate averages to meter-scale remote sensing imagery to capture fine-grained environmental features.
  2. Presence-Only Data: Most citizen science data records where a species was seen, but rarely where it was searched for and not found. This lack of confirmed absences makes standard classification metrics unreliable and necessitates specialized evaluation strategies like the Top-K error rate.

Multi-modal Fusion Strategies

The core technical challenge for the 52 competing teams was multi-modal fusion: effectively combining 2D spatial satellite images with 1D environmental vectors (such as elevation, rainfall, or soil properties). Participants generally gravitated toward three architectural patterns:

  • Early Input Aggregation: Stacking all available spatial data (RGB, Near-Infrared, Altitude) into a single multi-channel "image" to be processed by a standard Convolutional Neural Network (CNN) like ResNet.
  • Independent Feature Extractors: Utilizing separate networks for different modalities—for instance, a CNN for imagery and a Multi-Layer Perceptron (MLP) for tabular climate data—and concatenating their features before a final classification layer.
  • Late Predictions Averaging: Training independent models for each data type and averaging their final probability outputs in an ensemble approach.

The winning solution, developed by the Sensio Team, utilized a sophisticated ensemble of two bi-modal networks (one based on ResNet-34 and another on MobileNetV3) integrated with a Random Forest model for tabular data.

The GeoLifeCLEF 2022 Dataset

The competition utilized the GeoLifeCLEF 2022 dataset, currently the largest publicly available resource pairing remote sensing imagery with species observations. It covers 1.6 million observations across 17,037 species in the United States and France.

A sample multi-modal observation from the GeoLifeCLEF 2022 dataset, showcasing RGB, Near-IR, Altitude, and Land cover layers for a specific species occurrence.

As shown in the figure above, each observation is associated with several high-resolution covariates, including 256x256 pixel patches (1 meter/pixel resolution) of RGB imagery, Near-Infrared (NIR) imagery, land cover maps, and altitude data. These are supplemented by low-resolution covariates consisting of 19 bio-climatic variables and 8 pedologic (soil) variables.

Key Insights and Results

The competition revealed several critical insights into modern SDM:

  • The Power of Imagery: High-resolution imagery provides complementary information that traditional climate variables miss. CNNs trained on image patches consistently outperformed models relying solely on tabular environmental data.
  • The "Presence-Only" Constraint: To handle the lack of confirmed absences, the organizers used a Top-30 error rate. This treats the problem as a set-valued classification task, asking if the observed species is among the top 30 most likely species predicted for a given location.
  • Pre-training Benefits: Models pre-trained on ImageNet consistently outperformed those trained from scratch, despite the visual differences between satellite imagery and standard photographic datasets.

The performance gap between the top models and baselines was stark. The winning Sensio Team achieved a Top-30 error rate of 68.5%, significantly better than the standard CNN baseline on RGB patches (73.7%) and the Random Forest baseline using traditional environmental vectors (76.2%).

Limitations and Future Work

Despite these successes, the authors identified areas for future development. Citizen science data is inherently biased toward areas with high human activity, a factor the current evaluation metric does not fully address. Additionally, the Top-K error rate can be sensitive to regional biodiversity; a Top-30 list may be too restrictive in high-diversity areas while being too lenient in low-diversity ones. Future iterations of the challenge aim to incorporate "presence-absence" data from structured surveys to provide a more robust ground truth for evaluation.

Summary updated

BibTeX
@inproceedings{lorieul2022overview,
  author = {Titouan Lorieul and Elijah Cole and Benjamin Deneu and Maximilien Servajean and Pierre Bonnet and Alexis Joly},
  title = {{Overview of GeoLifeCLEF 2022: Predicting Species Presence from Multi-modal Remote Sensing, Bioclimatic and Pedologic Data}},
  booktitle = {CLEF (Working Notes)},
  year = {2022},
  url = {https://openreview.net/forum?id=LAn-wR8-zys},
}
Figure