SeasFire Cube: Wildfire Modeling Dataset

• SeasFire Cube Dataset is a specialized global resource encoding wildfire activity with environmental drivers and climate indices for continuous spatiotemporal forecasting.
• It organizes data on a one-degree grid into a three-level hierarchical graph that supports multi-scale message passing and adaptive feature fusion.
• The dataset underpins the HiGO framework, enabling neural ODE-based models to achieve state-of-the-art long-range wildfire prediction with class-imbalanced ordinal labeling.

The SeasFire Cube Dataset is a specialized data resource for global wildfire activity modeling, designed to support research on multi-scale, continuous-time dynamics of wildfires as governed by atmospheric, oceanic, and terrestrial influences. Developed as the experimental foundation for the Hierarchical Graph ODE (HiGO) framework, SeasFire Cube enables the training and evaluation of deep learning architectures capable of spatiotemporal wildfire prediction on a planetary scale, using a unified, multi-resolution graph-based representation of Earth’s surface and relevant environmental drivers (Xu et al., 4 Jan 2026).

1. Dataset Overview and Purpose

The SeasFire Cube Dataset encodes wildfire activity and associated driver variables on a global one-degree grid, structuring them for hierarchical graph-based modeling strategies. Each dataset entry comprises driver variables, climate indices, a burning area channel, and ground-truth labels for downstream wildfire prediction tasks. The dataset is optimized for scenarios requiring long-range spatiotemporal forecasting and is used for benchmarking predictive models such as the HiGO framework in continuous time settings. Its construction facilitates effective multi-level representation, aggregation, and learning of physical processes relevant to wildfire occurrence and propagation.

2. Data Structure and Granularity

At the core, the SeasFire Cube Dataset organizes information in a grid format at the base resolution of $1^\circ$ latitude–longitude cells covering Earth’s surface. The representation supports a multi-level (three-level) graph hierarchy:

• Level 1 (base): Nodes correspond to individual $1^\circ \times 1^\circ$ cells.
• Level 2: Nodes aggregate $2 \times 2$ blocks of Level 1.
• Level 3: Nodes aggregate $4 \times 4$ blocks of Level 1.

Each node is associated with a $D$-dimensional feature vector (with $D=256$), connecting to both intra-level neighbors (4-way connectivity) and unique inter-level parent nodes. Adjacent cells share scalar edge features, facilitating message passing and hierarchical feature fusion. This construction reflects the multi-scale and continuous characteristics of real-world wildfire dynamics.

3. Input Modalities and Feature Encoding

The dataset supplies three primary input modalities per example:

• Driver variables $$\mathcal{V}_{t_0} \in \mathbb{R}^{H\times W\times C_x}$$: Physical and meteorological predictors, spatially distributed.
• Climate indices $\mathcal{I}_{t_0} \in \mathbb{R}^{C_z}$: Global or regional climatic factors abstracted as vectors.
• Burning area channel $\mathcal{B}_{t_0}$: Direct observations or reconstructions of burning area at baseline time $t_0$.

Feature encoding procedures include:

• A learnable fusion of driver variables and climate indices via a gated mechanism: $\mathcal{V}_{t_0}$ is encoded to $\mathcal{Z} \in \mathbb{R}^{H\times W\times D}$, augmented with three channel-attention vectors $a_\omega$ derived via GeLU-activated MLPs on climate indices.
• Three small convolutional projections $\kappa_\omega$ produce $\mathcal{U}_\omega$, combined for the fused feature $$\mathcal{H} = \mathcal{Z} + \sum_{\omega=1}^3 a_\omega \odot \kappa_\omega(\mathcal{Z})$$.
• Cross-attention between $\mathcal{H}$ and the burning-area channel $\mathcal{B}_{t_0}$ produces refined features $\mathcal{X}_{t_0}$.
• Level-1 node features are initialized from $\mathcal{X}_{t_0}$; higher levels use learnable downsampling mechanisms.

4. Label Structure and Quantization

Ground truth in the SeasFire Cube Dataset is formulated as an ordinal quantization task for burning area (BA) prediction:

• For each spatial location $(i,j)$, the label $y_{ij}$ falls into one of $K$ discrete classes (number of classes $K$ is determined by the quantization scheme).
• Output prediction at future times $t_1 = t_0+\tau$ is structured as a softmax over the $K$ classes per location: $$\hat{\mathcal{P}} = \mathrm{softmax}(\hat{\mathcal{L}})$$, where $\hat{\mathcal{L}}$ is produced by a decoding MLP acting on the finest-level node features at $t_1$.
• Loss function is a weighted cross-entropy over spatial locations and quantization classes:

$$\mathcal{L} = -\frac{1}{HW} \sum_{i,j}\sum_{k=0}^{K-1} w_k \mathbf{1}(y_{ij}=k)\log \hat{p}_{ij,k}$$

where $w_k \propto 1/\mathrm{freq}(k)$ for class frequency reweighting.

5. Role in Hierarchical Graph ODE (HiGO) and Experimental Design

The SeasFire Cube Dataset underpins the HiGO model, which employs a multi-level graph hierarchy with adaptive filtering message passing and GNN-parameterized neural ODEs for continuous-time dynamics:

• Intra-level message passing utilizes attention-based aggregation with normalized coefficients computed from local node and edge embeddings.
• Inter-level propagation occurs via differentiable downsampling (fine→coarse fusion) and upsampling (coarse→fine correction) mechanisms, supporting effective cross-scale information flow.
• Stacked “HGODE” blocks at each hierarchy level yield a system of ODEs solved via an adaptive-step Dormand–Prince RK5 (dopri5) method.
• The model’s predictions on the SeasFire Cube Dataset demonstrate improved long-range forecasting ability relative to state-of-the-art baselines and show high observational consistency in predicted wildfire activity (Xu et al., 4 Jan 2026).

6. Implementation and Usage Considerations

The HiGO experimental setup on SeasFire Cube specifies:

• Hidden feature dimension $D=256$
• Layer normalization after upsampling and in the feature mixer
• GeLU activations in channel-attention modules; ReLU elsewhere
• Optimization via AdamW (learning rate $10^{-4}$, weight decay $10^{-5}$), cosine annealing over 100 epochs, and minibatch size 8
• Codebase constructed with PyTorch and PyTorch Geometric, employing TorchDiffEq for neural ODE integration.

Usage of SeasFire Cube thus facilitates reproducibility of multi-scale, continuous-time wildfire prediction frameworks, providing both raw modalities and ground-truth quantized outputs as required for state-of-the-art graph-based machine learning model development.

7. Significance and Research Directions

SeasFire Cube provides a benchmark enabling the creation and accurate assessment of global wildfire prediction models operating on continuous, multi-resolution spatiotemporal data. Its configuration for hierarchical graph learning, ODE-based dynamics modeling, and class-imbalanced ordinal prediction addresses core challenges in representing, forecasting, and understanding wildfire behavior across scales. A plausible implication is that further work leveraging SeasFire Cube could extend to Earth system modeling domains where multi-scale, temporally continuous, and data-sparse characteristics present analogous computational and scientific challenges (Xu et al., 4 Jan 2026).