This repository is related to the GeoLifeCLEF challenge.
The details of the challenge, the data, and all other useful information are present on the challenge page: https://www.kaggle.com/competitions/geolifeclef-2024
In this repository you will find a ready-to-use code to run one of the baseline models for the GeoLifeCLEF 2024 challenge: the MultimodelEnsemble baseline.
This example can be run using the following command:
python glc24_cnn_multimodal_ensemble.pyand uses a custom dataset from malpolon.data.datasets.geolifeclef2024_pre_extracted as well as a custom model from malpolon.models.geolifeclef2024_multimodal_ensemble.
The data used in this example is the pre-extracted data from the GeoLifeCLEF 2024 challenge. The data is available on the [https://www.kaggle.com/competitions/geolifeclef-2024/data](challenge "data" page).
It consists of points extractions from Sentinel-2A patches, Landsat time series, and Bioclim time series for each species occurrence of Presence/Absence (PA) surveys.
The model used in this example is the Multimodal Ensemble model. This model is a simple ensemble of 3 models:
- ResNet18 for Landsat time series extractions
- ReNet18 for Bioclim time series extractions
- Swin V2 Transformer for Sentinel-2A patches extractions
These 3 models are then merged with a simple late fusion strategy (concatenation of features + Multi-Layer Perceptron (MLP)). A single loss is used for the training of the 3 models.