README.md

---
configs:
- config_name: Kenya
  data_files:
  - split: train
    path: Kenya/train_filtered.csv
  - split: val
    path: Kenya/valid_filtered.csv
  - split: test
    path: Kenya/test_filtered.csv
  default: true
- config_name: South_Africa
  data_files:
  - split: train
    path: South_Africa/train_filtered.csv
  - split: val
    path: South_Africa/valid_filtered.csv
  - split: test
    path: South_Africa/test_filtered.csv
- config_name: USA_Summer
  data_files:
  - split: train
    path: USA_Summer/train_filtered.csv
  - split: val
    path: USA_Summer/valid_filtered.csv
  - split: test
    path: USA_Summer/test_filtered.csv
- config_name: USA_Winter
  data_files:
  - split: train
    path: USA_Winter/train_filtered.csv
  - split: val
    path: USA_Winter/valid_filtered.csv
  - split: test
    path: USA_Winter/test_filtered.csv
license: cc-by-nc-4.0
---

license: cc-by-nc-4.0
---

# BATIS: Benchmarking Bayesian Approaches for Improving Species Distribution Models

This repository contains the dataset used in experiments shown in BATIS: Benchmarking Bayesian Approaches for Improving Species Distribution Models (preprint). To download the dataset, you can use the `load_dataset` function from HuggingFace. For example :

```python
from datasets import load_dataset

# Training Split for Kenya
training_kenya = load_dataset("cathv/batis_benchmark_2025", name="Kenya", split="train")

# Validation Split for South Africa
validation_south_africa = load_dataset("cathv/batis_benchmark_2025", name="South_Africa", split="val")

# Test Split for USA-Summer
test_usa_summer = load_dataset("cathv/batis_benchmark_2025", name="USA_Summer", split="test")
```

The code to reproduce the experiments presented in the paper is available in our [GitHub repository](https://github.com/cath34/batis_benchmark_neurips_2025). Please note that the checklists data is **NOT AVAILABLE** in this current repository, in order to comply with the [eBird Data Access Terms of Use
](https://www.birds.cornell.edu/home/ebird-data-access-terms-of-use/). If you are interested in downloading the checklists data, please formulate a [request for access](https://support.ebird.org/en/support/solutions/articles/48000838205-download-ebird-data) to the eBird team. 

## ⚠️ !!! ERRATUM IN THE MAIN PAPER !!! ⚠️

We would like to apologize to the reviewers for a typo in Table 1 of the main paper. The table incorrectly suggests that hundreds of thousands of species can be observed in the United States during summer, and nearly 50,000 in winter. While many birders would surely dream of such an extraordinary high avian biodiversity, these numbers are clearly far from the reality. The values intended for the `number_of_hotspots` column were unfortunately placed in the `number_of_species` column. The first table of the Appendix reports the appropriate numbers, but we also include it here to avoid any confusion : 

| **Region**           | **Date Range**              | **Number of Checklists** | **Number of Hotspots** | **Number of Species** | **Species List** |
|----------------------|-----------------------------|---------------------------|-------------------------|------------------------|------------------|
| Kenya (KE)           | 2010-01-01 to 2023-12-31    | 44,852                    | 8,551                   | 1,054                  | Avibase          |
| South Africa (ZA)    | 2018-01-01 to 2024-06-17    | 498,867                   | 6,643                   | 755                    | BirdLife         |
| USA-Winter (US-W)    | 2022-12-01 to 2023-01-31    | 3,673,742                 | 45,882                  | 670                    | ABA 1-2          |
| USA-Summer (US-S)    | 2022-06-01 to 2022-07-31    | 3,920,846                 | 98,443                  | 670                    | ABA 1-2          |

## Dataset Configurations and Splits

The dataset contains the following four configurations :

- **Kenya :** Containing the data used to train our models for predicting bird species distribution in Kenya.
- **South Africa :** Containing the data used to train our models for predicting bird species distribution in South Africa.
- **USA-Winter :** Containing the data used to train our models for predicting bird species distribution in the United States of America during the winter season.
- **USA-Summer :** Containing the data used to train our models for predicting bird species distribution in the United States of America during the summer season.

Each subset can be further divided into `train`, `test` and `split`. These splits are the same as the one we used in our paper, and were generated by following the pre-processing pipeline described in our paper, which can be easily reproduced by re-using our [code](https://github.com/cath34/batis_benchmark_neurips_2025/blob/main/Splits/combined_script.py). 

## Dataset Structure

```
/batis_benchmark_2025/
    Kenya/
        images.tar.gz
        environmental.tar.gz
        targets.tar.gz
        train_filtered.csv
        test_filtered.csv
        valid_filtered.csv
    South_Africa/
        images.tar.gz
        environmental.tar.gz
        targets.tar.gz
        train_filtered.csv
        test_filtered.csv
        valid_filtered.csv
    USA_Winter/
        images/
          images_{aa}
          ...
          images_{ad}
          
        environmental.tar.gz
        targets.tar.gz
        train_filtered.csv
        test_filtered.csv
        valid_filtered.csv
    USA_Summer/
        images/
            images_{aa}
            ...
            images_{af}
        images.tar.gz
        environmental.tar.gz
        targets.tar.gz
        train_filtered.csv
        test_filtered.csv
        valid_filtered.csv
    Species_ID/
        species_list_kenya.csv
        species_list_south_africa.csv
        species_list_usa.csv

```

The files `train_filtered.csv`, `test_filtered.csv` and `valid_filtered.csv` are containing the informations one can see from the Dataset Viewer. The archives `targets`, `images`, `environmental` are respectively containing the target vectors (i.e., the estimated ground truth encounter rate probability). The `Species_ID/` folder contains the species list files for each subset. 


## Data Fields

- `hotspot_id` : The unique ID associated with a given hotspot. The `hotspot_id`value can be used to upload date coming from either `targets`, `environmental` or variance, as they are all formulated as

```
/batis_benchmark_2025/
    images/
        {hotspot_id_1}.tar.gz
        ...
        {hotspot_id_n}.tar.gz
    environmental/
        {hotspot_id_1}.tar.gz
        ...
        {hotspot_id_1}.tar.gz
    targets/
        {hotspot_id_1}.tar.gz
        ...
        {hotspot_id_1}.tar.gz
```
- `lon` : Longitude coordinate of the hotspot
- `latitude` : Latitude coordinate of the hotspot
- `num_complete_checklists` : Number of complete checklists collected in that hotspot
- `bio_1` to `bio_19`: Environmental covariates values associated with that hotspot, extracted from the WorldClim model. For more details on each of these variables, please refer to Table 5 of the appendix.
- `split` : The split associated with that hotspot (either `train`, `valid` or `test`)

## Authors 

- **Curated by:** Catherine Villeneuve, Mélisande Teng, Benjamin Akera, David Rolnick
- **Language:** English
- **Repository:** https://huggingface.co/datasets/cathv/batis_benchmark_2025
- **Paper:** IN REVIEW

## Licenses 

The **BATIS Benchmark** is released under a [Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License](https://creativecommons.org/licenses/by-nc/4.0/).

The use of our dataset should also comply with the following:

- [eBird Terms of Use](https://www.birds.cornell.edu/home/terms-of-use/)
- [eBird API Terms of Use](https://www.birds.cornell.edu/home/ebird-api-terms-of-use/)
- [eBird Data Access Terms of Use](https://www.birds.cornell.edu/home/ebird-data-access-terms-of-use/)