Основні поняття
WildGraph generates realistic long-horizon trajectories for wildlife movement by constructing a hierarchical prototype network of regions and leveraging a variational recurrent network to probabilistically generate paths over the regions.
Анотація
The paper proposes a framework called WildGraph to generate realistic long-horizon trajectories for wildlife movement. The key components of WildGraph are:
-
Hierarchical Network Generator (HNG):
- Discretizes the geographic area into a prototype network of regions using the H3 spatial indexing system.
- Recursively refines the regions to capture local movement characteristics while preserving global movement patterns.
- Constructs a graph representing transitions between the regions based on the observed trajectories.
-
Graph Embedding Layer:
- Utilizes node2vec to learn low-dimensional embeddings for the nodes (regions) in the graph, capturing both local and global structural information.
-
Variational Recurrent Network (VRN):
- Processes the embedded trajectory sequences to learn a transition kernel that models the sequential decision-making under uncertainty.
- Generates new trajectories by iteratively updating a hidden state and sampling the next region based on the learned transition probabilities.
-
Latent Dictionary:
- Stores the latent space representations (from the VRN encoder) for each region, enabling stochastic sampling during the generation process.
-
Occupancy Sampler:
- Converts the selected regions into geographic points by sampling from a heatmap based on the occupancy of the real trajectories within each region.
The proposed WildGraph framework is evaluated on two wildlife migration datasets and compared against baselines such as VAE, GAN, and Transformer-based approaches. WildGraph demonstrates superior performance in terms of trajectory similarity and distribution alignment with the real trajectories, while also achieving high coverage of the test set.
The key advantages of WildGraph are its ability to:
- Generate realistic long-horizon trajectories using a small set of real samples.
- Improve the generalization of the generated trajectories by leveraging the hierarchical prototype network.
- Capture both global movement patterns and local movement characteristics through the combination of the HNG and VRN.
Статистика
The average Hausdorff distance between the generated and real trajectories is 9.38 for the geese dataset and 7.29 for the stork dataset.
The average Dynamic Time Warping (DTW) distance is 67.87 for the geese dataset and 40.16 for the stork dataset.
The average Final Displacement Error (FDE) is 2.61 for the geese dataset and 3.74 for the stork dataset.
Цитати
"WildGraph successfully generates realistic months long trajectories using a sample size as small as 60."
"Experiments performed on two wildlife migration datasets demonstrate that our proposed method improves the generalization of the generated trajectories in comparison to existing work, while achieving superior or comparable performance in several benchmark metrics."