Adaptive Filtering Message Passing

• Adaptive Filtering Message Passing is a framework that dynamically adjusts node communications based on input features and edge weights for context-aware information flow.
• It employs fusion of heterogeneous modalities, such as environmental drivers and observed data, to construct high-fidelity multi-scale graph representations.
• Practical applications include wildfire forecasting and Earth system modeling, where adaptive filtering improves predictive accuracy over static methods.

The SeasFire Cube dataset is a foundational resource for global-scale wildfire activity modeling, serving as the principal empirical basis for the development and evaluation of advanced deep learning frameworks such as the Hierarchical Graph ODE (HiGO). This dataset encapsulates the complex, multi-scale spatiotemporal characteristics of wildfire behavior across the Earth's surface, with representations tailored for graph-based and continuous-time neural modeling (Xu et al., 4 Jan 2026). The unique integration of driver variables, climate indices, and observed burning-area distributions in a structured, gridded format enables high-fidelity study of fire dynamics across atmospheric, oceanic, and terrestrial interactions.

1. Dataset Composition and Structure

The SeasFire Cube dataset organizes global wildfire-related observations across three spatial resolutions, referred to as levels ($l = 1, ..., L$, with $L = 3$), corresponding to increasingly coarse gridded representations. At each level $l$, the domain is discretized into a grid of size $H_l \times W_l$:

• Level 1 Nodes ($V^{(1)}$): Each node represents a $1^\circ$ latitude-longitude cell, initialized with features embedding both environmental driver variables and burning-area information.
• Higher Levels ($l > 1$): Nodes correspond to blocks of $2^{l-1} \times 2^{l-1}$ grid cells, with features computed via learnable downsampling.
• Edges: Intra-level edges connect 4-way neighbors, supporting spatial information flow within each level. Inter-level edges connect each child node to its unique parent at the next coarser level.

Every node feature at time $t$ is a real-valued vector (base level: $\mathbf{x}_i^{(1)}(t)\in\mathbb{R}^D$, with $D=256$), incorporating fused driver and burning-area embeddings. Scalar edge features $\mathbf{e}_{ij}^{(l)} = 1$ support message passing and attention-based aggregation.

2. Data Modalities and Feature Engineering

The dataset fuses heterogeneous modalities to create node feature embeddings:

• Driver Variables ($\mathcal{V}_{t_0}$): $H \times W \times C_x$ arrays encoding atmospheric, oceanic, and terrestrial covariates.
• Climate Indices ($\mathcal{I}_{t_0}$): Global descriptors ($\mathbb{R}^{C_z}$), such as ENSO and PDO indices.
• Burning-Area Channels ($\mathcal{B}_{t_0}$): Gridded observed burning fractions, provided as an image-like tensor.
• Feature Fusion: Channel-attention MLPs produce climate-informed weights ($a_\omega$), applied to convolutional transforms of driver variables ($\kappa_\omega(\mathcal{Z})$), yielding a composite embedding ($\mathcal{H}$). Subsequent cross-attention with burning-area information generates a spatiotemporal feature field ($\mathcal{X}_{t_0}$) for level-1 initialization.

This design facilitates the model's ability to contextualize local wildfire behavior within global climate patterns and real-time fire observations.

3. Spatiotemporal Graph Representation

Each temporal snapshot in the SeasFire Cube dataset is encoded as a tri-level graph hierarchy:

Level ($l$)	Node Interpretation	Feature Construction
1	$1^\circ$ grid cell	Fused driver + BA embedding
2	$2\times2$ block of level-1	Learnable downsampling
3	$4\times4$ block of level-1	Learnable downsampling

Edge sets comprise intra-level 4-way adjacency and inter-level parent-child relations. Downsampling and upsampling operations are realized via learnable $\softmax$-normalized MLPs. This structure enables multi-scale feature propagation and supports continuous-time dynamic modeling with Graph Neural ODEs.

4. Usage in Hierarchical Graph ODE Frameworks

The SeasFire Cube dataset is central to the HiGO framework, which leverages it as follows (Xu et al., 4 Jan 2026):

1. Initialization: Climate-informed fusion yields node features at $t_0$, mapped to graph nodes.
2. Graph ODE Integration: GNN-parameterized Neural ODEs propagate feature states forward in time, modeling continuous wildfire dynamics.
3. Decoding: Output node states at the finest level ($\mathbf{X}^{(1)}(t_1)$) are decoded into class probabilities for ordinal burned-area quantization via MLP and softmax.

Training utilizes weighted cross-entropy loss, balancing class frequencies and optimizing predictive skill over long horizons.

5. Key Implementation and Evaluation Practices

SeasFire Cube supports high-throughput, multi-resolution deep learning experiments with the following configuration parameters:

• Hidden dimension: $D=256$
• Levels: $L=3$
• Normalization: LayerNorm post-upsampling and in fusion layers
• Activation: GeLU in channel-attention, ReLU elsewhere
• ODE Solver: Dormand–Prince RK5 (dopri5 in TorchDiffEq, default tolerances)
• Optimization: AdamW ($\text{lr}=10^{-4}$, weight decay $10^{-5}$), cosine annealing schedule (100 epochs), minibatch size 8
• Software stack: PyTorch and PyTorch Geometric

A practitioner is thus provided a complete specification for multi-scale graph construction, adaptive message passing, ODE-based feature flow, and training regime.

6. Significance and Applications

The SeasFire Cube dataset enables rigorous study of wildfire dynamics at a scale and level of integration not previously accessible. Its multi-resolution graph encoding, combined with harmonized driver, climate, and fire observations, underpins benchmarks for continuous-time global wildfire forecasting. Experiments have demonstrated substantial performance improvements in forecasting skill and observational consistency when used with the HiGO framework (Xu et al., 4 Jan 2026). These advances suggest enhanced potential for operational wildfire early warning and climate impact assessment.

7. Connections to Broader Research Directions

The architectural approach facilitated by SeasFire Cube exemplifies cutting-edge trends in machine learning for Earth system modeling: multi-scale graph representations, attention mechanisms for feature fusion, and Neural ODEs for continuous-time prediction. The dataset's structure and integration with HiGO offer a template for similar efforts in related domains, such as global weather, hydrology, and ecosystem modeling. A plausible implication is that future extensions might incorporate additional process layers (e.g., vegetation dynamics), further enriching multi-modal, physically consistent forecasting frameworks.