--- license: cc0-1.0 pretty_name: WildGraph Wildlife Migration Trajectories task_categories: - other tags: - trajectory-generation - trajectory - wildlife - animal-movement - migration - gps - geospatial - movebank size_categories: - 10K Ali Al-Lawati, Elsayed Eshra, Prasenjit Mitra. > [WildGraph: Realistic Long-Horizon Trajectory Generation with Limited Sample Size](https://doi.org/10.1145/3678717.3691265). > SIGSPATIAL 2024. ([arXiv:2404.08068](https://arxiv.org/abs/2404.08068)) Each dataset is a small set of months-long GPS trajectories with one location per day and the same length for every trajectory. They were prepared for long-horizon trajectory generation from few samples: a model sees a few dozen real migrations and has to generate new, realistic ones. ## Contents Two configurations, each with a single `train` split and one row per daily location. | Configuration | Species | Clipping window | Days per trajectory | Trajectories | Rows | | --- | --- | --- | ---: | ---: | ---: | | `geese` | Greater white-fronted goose (*Anser albifrons*) | Mar 1 to Sep 1 | 185 | 60 | 11,100 | | `stork` | White stork (*Ciconia ciconia*) | Aug 15 to Nov 1 | 79 | 77 | 6,083 | The geese trajectories lie between 48.5° and 76.2° N and between 2.8° and 113.2° E, and the stork trajectories between 9.3° and 53.0° N and between 6.4° W and 37.7° E. ```python from datasets import load_dataset ds = load_dataset("aliwister/WildGraph", "geese", split="train") print(ds[0]) # {'idx': 9065, 'label': 0, 'day': 0, 'location.lat': 52.0339758, 'location.long': 6.1129115} # all trajectories as one array of shape (trajectories, days, 2), ordered (latitude, longitude) df = ds.to_pandas() trajs = df[["location.lat", "location.long"]].to_numpy().reshape(df["label"].nunique(), -1, 2) print(trajs.shape) # (60, 185, 2) ``` ## Fields | Field | Type | Description | | --- | --- | --- | | `idx` | int64 | Row identifier kept from the released CSV files. It is unique within a configuration and consecutive within a trajectory, but it does not order the trajectories. | | `label` | int64 | Trajectory identifier, 0 to 59 for `geese` and 0 to 76 for `stork`. One label is one animal in one season (see below). | | `day` | int64 | Position of the location within its trajectory, starting at 0. Day 0 corresponds to the first day of the clipping window. | | `location.lat` | float64 | Latitude in decimal degrees (WGS84). | | `location.long` | float64 | Longitude in decimal degrees (WGS84). | Rows are stored trajectory by trajectory and in day order. The files hold no timestamps, years, or animal identifiers. ## How the data was built Both datasets were obtained from the [Movebank Data Repository](https://www.movebank.org) and processed as described in Section 5.4 of the paper. 1. Observations were reduced to one per day for each animal with [MoveApps](https://www.moveapps.org). 2. Each track was clipped to the window in the table above, which covers one migration. 3. Animals tracked over several years were split into one trajectory per year, so a `label` is an animal in one season, and two labels can belong to the same animal. 4. Trajectories with significant missing readings were dropped, and the remaining isolated gaps were filled by simple interpolation. A small number of locations are therefore interpolated and not measured. ## Use The paper trains on the trajectories of four folds and compares generated trajectories with the held-out fold, using five-fold cross-validation over `label` without shuffling. ```python from sklearn.model_selection import KFold labels = range(df["label"].nunique()) for k, (train_labels, test_labels) in enumerate(KFold(n_splits=5).split(labels)): train, test = df[df["label"].isin(train_labels)], df[df["label"].isin(test_labels)] ``` Generated trajectories are scored by their distance to the real ones (Hausdorff distance, dynamic time warping, and final displacement error), by how many held-out trajectories they cover, and by how closely their spatial distribution matches the real one (Pearson correlation and chi-squared similarity). Code for WildGraph, the baselines, and the metrics is at [github.com/aliwister/wildgraph](https://github.com/aliwister/wildgraph). The code reads the columns `idx`, `label`, `location.lat`, and `location.long`, which keep their original names here, so `ds.to_pandas()` can replace the CSV files of that repository. ## Source data and license The trajectories are derived from two studies published in the Movebank Data Repository under CC0 1.0. This dataset is released under the same terms. Please cite the source data and the papers they accompany when you use it. - `geese`: Kölzsch A, Müskens GJDM, Moonen S, Kruckenberg H, Glazov P, Wikelski M (2019). Data from: Flyway connectivity and exchange primarily driven by moult migration in geese [North Sea population]. Movebank Data Repository. [doi:10.5441/001/1.ct72m82n](https://doi.org/10.5441/001/1.ct72m82n). Accompanies Kölzsch et al. (2019), Movement Ecology 7:3, [doi:10.1186/s40462-019-0148-6](https://doi.org/10.1186/s40462-019-0148-6). - `stork`: Carlson B, Rotics S, Nathan R, Wikelski M, Jetz W (2021). Data from: Individual environmental niches in mobile organisms. Movebank Data Repository. [doi:10.5441/001/1.rj21g1p1](https://doi.org/10.5441/001/1.rj21g1p1). Accompanies Carlson et al. (2021), Nature Communications 12:4572, [doi:10.1038/s41467-021-24826-x](https://doi.org/10.1038/s41467-021-24826-x). ## Citation ```bibtex @inproceedings{allawati2024wildgraph, title = {WildGraph: Realistic Long-Horizon Trajectory Generation with Limited Sample Size}, author = {Al-Lawati, Ali and Eshra, Elsayed and Mitra, Prasenjit}, booktitle = {Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems}, publisher = {ACM}, pages = {247--258}, year = {2024}, doi = {10.1145/3678717.3691265} } ```