Advanced Deep Learning Approaches for Forecasting High-Resolution Fire Weather Index (FWI) over CONUS: Integration of GNN-LSTM, GNN-TCNN, and GNN-DeepAR

Abstract

Wildfires in the United States have increased in frequency and severity over recent decades, driven by climate change, altered weather patterns, and accumulated flammable materials. Accurately forecasting the Fire Weather Index (FWI) is crucial for mitigating wildfire risks and protecting ecosystems, human health, and infrastructure. This study analyzed FWI trends across the Continental United States (CONUS) from 2014 to 2023, using meteorological data from the gridMET dataset. Key variables, including temperature, relative humidity, wind speed, and precipitation, were utilized to calculate the FWI at a fine spatial resolution of 4 km, ensuring the precise identification of wildfire-prone areas. Based on this, our study developed a hybrid modeling framework to forecast FWI over a 14-day horizon, integrating Graph Neural Networks (GNNs) with Temporal Convolutional Neural Networks (TCNNs), Long Short-Term Memory (LSTM), and Deep Autoregressive Networks (DeepAR). The models were evaluated using the Index of Agreement (IOA) and root mean squared error (RMSE). The results revealed that the Southwest and West regions of CONUS consistently exhibited the highest mean FWI values, with the summer months demonstrating the greatest variability across all climatic zones. In terms of model performance on forecasting, Day 1 results highlighted the superior performance of the GNN-TCNN model, achieving an IOA of 0.95 and an RMSE of 1.21, compared to the GNN-LSTM (IOA: 0.93, RMSE: 1.25) and GNN-DeepAR (IOA: 0.92, RMSE: 1.30). On average, across all 14 days, the GNN-TCNN outperformed others with a mean IOA of 0.885 and an RMSE of 1.325, followed by the GNN-LSTM (IOA: 0.852, RMSE: 1.590) and GNN-DeepAR (IOA: 0.8225, RMSE: 1.755). The GNN-TCNN demonstrated robust accuracy across short-term (days 1–7) and long-term (days 8–14) forecasts. This study advances wildfire risk assessment by combining descriptive analysis with hybrid modeling, offering a scalable and robust framework for FWI forecasting and proactive wildfire management amidst a changing climate.

1. Introduction

Wildfires in the United States have become increasingly severe and widespread over the past four decades, with the area burned by wildfires quadrupling during this period. This alarming rise is attributed to the accumulation of flammable materials from decades of fire suppression and recent drier conditions exacerbated by climate change [1]. A study reveals rising wildfire potential in the U.S., with annual mean Keetch–Byram Drought Index (KBDI) increasing by 5.2 and 2.9 per year in the Southwestern and Northwestern CONUS since 1982, and extreme fire weather days up by 16 and 25 days. By the century’s end, high-risk days could increase by 90–189 annually, with Southern CONUS high-risk areas expanding 60-fold, driven primarily by warming temperatures [2]. These fires are no longer confined to the Western region but have spread nationwide, affecting the nation’s economy. According to Nielsen-Pincus et al. [3], the average annual expenditure on wildfire suppression in the U.S. has exceeded $1.5 billion, primarily in the Western states. Consequently, wildfires have become a burning issue in the US. In addition, wildfire is also associated with profound and multifaceted health issues. The smoke released from wildfire comprises a complex combination of air pollutants, such as carbon monoxide (CO), nitrogen oxides (NOx), volatile organic compounds (VOCs), and particulate matter, which have the ability to deeply infiltrate the lungs and enter the bloodstream of humans [4,5]. Thus, it presents substantial risks to human well-being, such as increased death rates, respiratory and cardiovascular problems, negative effects on childbirth, and mental health difficulties [6,7,8,9,10,11]. In particular, children are more vulnerable, facing significant effects on their physical and mental well-being [6]. Despite these negative impacts, wildfires also play a crucial role in Earth’s system. In addition to influencing the climate through radiative forcing, wildfires play a role in the regeneration of Amazon forests following deforestation [12], contribute to the thawing of permafrost [13], and trigger notable events of ocean fertilization [14]. Consequently, the timely and reliable forecasting of potential wildfire events and the assessment of associated risks is essential for ensuring public health and safety and protecting the environment, especially in the context of climate change.

Forecasting the risk of wildfires is challenging to undertake due to the complex nature of the climate system and its interactions with vegetation and socio-economic factors [15]. Nevertheless, notable progress in numerical weather forecasting in the field of atmospheric science has also enhanced fire forecasting, prompting certain government agencies to provide fire weather forecasting services [16]. Among them, the US Forest Service (USFD), the National Interagency Fire Center (NIFC), and the US Geological Survey (USGS) offer public forecasts of wildfire hazards in the US. For instance, NIFC provides the 7-Day Significant Fire Potential Outlook. Similarly, the European Forest Fire Information System (EFFIS) and Global Wildfire Information System (GWIS) use the FWI system to generate fire danger forecasts. Additionally, satellites like Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) and Goddard Earth Observing System Model, Version 5 (GEOS-5) generate daily datasets of the FWI on a global scale, where the spatial resolutions are 0.5° × 2/3° and 0.25° × 0.25°, respectively [17]. While these datasets are valuable for large-scale assessments, the coarse resolution limits their utility for more localized wildfire risk assessments and predictions, which, therefore, necessitates the utilization of a finer resolution of the FWI.

However, to achieve finer spatial resolution, numerical weather forecasting models require substantial computational resources and the integration of regional models, which somewhat limit the advancements [18,19]. In recent years, to overcome the complexity of the numerical models, artificial intelligence (AI) methods, i.e., machine learning (ML) and deep learning (DL) models, have achieved significant breakthroughs, including applications in climate domains such as wildfire prediction [20,21]. However, ML and DL models alone may not fully capture the dynamic processes governing wildfire occurrence, as they primarily rely on historical data and may lack the capacity to model future changes in climate and vegetation. This concern has led to hybrid models combining the temporal and spatial characteristics of the data [22]. For example, a study [23] developed a hybrid model, CFS-SR, that integrates the Climate Forecast System version 2 with DL techniques, significantly improving fire weather forecasts at fine resolutions (4 km) for California. While these models offer enhanced prediction, they also require complex processing and computational resources. Yet, the computational time for these hybrid models remains far shorter than that of numerical models that simulate wildfire forecasts on a fine grid [24].

The previous studies [25,26,27] demonstrate significant progress in applying ML, i.e., Random Forest (RF), Gradient-Boosted Trees (GBTs), Extreme Gradient Boosting (XGBoost), multilayer perceptron (MLP), Long Short-Term Memory (LSTM) and logistic regression for wildfire classification and prediction. However, each approach has inherent limitations that restrict its capacity to address the complex spatial–temporal dynamics associated with wildfire behavior. Models like RF and GBT, although effective in handling tabular data and nonlinear relationships, assume feature independence and lack mechanisms to explicitly model temporal and spatial dependencies [28,29]. This limitation hinders their ability to capture dynamic interactions between environmental variables such as temperature, wind speed, and vegetation moisture [30]. Similarly, XGBoost improves computational efficiency and reduces overfitting but, like RF and GBT, cannot inherently account for the interconnectedness of spatial and temporal features [31]. Neural networks, such as MLP and LSTM, offer more adaptability in nonlinear and sequential data modeling. However, MLPs fail to capture sequential dependencies, while LSTMs, despite their strength in temporal modeling, lack the explicit spatial reasoning necessary for understanding wildfire spread [32,33]. Logistic regression, though computationally efficient and interpretable, is unsuitable for capturing the nonlinear and multifaceted relationships that govern wildfire dynamics [34].

These limitations underscore the need for a more robust approach that integrates spatial and temporal dimensions effectively. GNNs provide a powerful alternative by explicitly representing spatial relationships through graph structures, enabling the modeling of complex interactions between geographically connected regions [35]. This capability is critical for wildfire forecasting, as fire behavior is strongly influenced by spatial factors like vegetation connectivity, topography, and climatic gradients. Several studies have explored the use of GNNs for wildfire forecasting due to their ability to model spatial dependencies and relationships among environmental variables such as vegetation, topography, and weather conditions [36,37,38]. These studies demonstrate the effectiveness of GNNs in capturing the spatial spread of wildfires and understanding how local features influence fire behavior across connected regions. However, using GNNs alone presents certain limitations. While GNNs excel at representing spatial correlations, they often lack the capability to fully capture temporal dynamics, such as the progression of fire risk over time due to changing meteorological conditions [39]. This limitation can lead to reduced accuracy in modeling wildfire evolution, particularly in scenarios where temporal patterns, i.e., prolonged drought or sudden changes in wind speed, play a critical role.

To address these gaps, hybrid models that integrate GNNs with temporal models can be proposed. When combined with temporal architectures such as LSTM, TCNN, or DeepAR, GNN-based models can seamlessly capture both dynamic spatial patterns and long-term temporal dependencies [40,41,42]. Unlike traditional ML models, GNN-based hybrid frameworks excel in handling sparse or irregular datasets, such as station-based measurements, and offer enhanced scalability and computational efficiency compared to numerical models [35]. By integrating GNN-TCNN, GNN-LSTM, and GNN-DeepAR architectures, this study addresses the shortcomings of existing methods, providing a scientifically robust and computationally efficient solution for high-resolution FWI forecasting in CONUS. These models not only bridge the gap between spatial and temporal complexity but also deliver improved accuracy and a deeper understanding of wildfire dynamics, making them a promising alternative to conventional approaches.

2. Materials and Methods

This study aims to forecast the FWI in CONUS over a 14-day horizon by integrating spatial and temporal dependencies using hybrid models: GNN-TCNN, GNN-LSTM, and GNN-DeepAR. This region was selected due to its diverse climatic zones and significant wildfire activity, making it a suitable study area for forecasting the Fire Weather Index (FWI). The CONUS encompasses nine distinct climate regions as classified by the National Centers for Environmental Information (NCEI): the Northeast, Northern Rockies and Plains, Upper Midwest, Ohio Valley, Southeast, South, Southwest, Northwest, and West. Each of these regions exhibits unique meteorological and environmental conditions that influence wildfire behavior, ranging from arid deserts in the Southwest to humid subtropical climates in the Southeast (Figure 1). The diversity of fire-prone ecosystems and the increasing frequency of extreme fire weather events across the CONUS underscore the importance of developing robust wildfire forecasting models tailored to the region’s heterogeneity.

This study selected a 14-day period for FWI forecasting over CONUS to assess both the short-term (1–7 days) and medium-term (8–14 days) predictive capabilities of the models, capturing varying temporal dependencies. Additionally, longer horizons, i.e., more than 14 days, were avoided as they exhibit diminishing accuracy in fire weather forecasting due to increasing meteorological uncertainties, as supported by other studies. The methodology is structured into three key stages: data collection and preprocessing, model development, and model evaluation. The overall methodology is illustrated in Figure 2.

2.1. Data Acquisition

The meteorological data of CONUS used in this study were collected from the gridMET dataset, a high-resolution gridded surface meteorological dataset (https://www.climatologylab.org/gridmet.html; accessed on 15 January 2024). The gridMET dataset provides data at a spatial resolution of 4 km × 4 km and data of daily resolution from 1979 to present, ensuring an accurate representation of local weather conditions across the study region. For this study, the data span a period of ten years, from 2014 to 2023, and include daily records of key meteorological variables required for the calculation of fire weather indices. These variables include daily mean temperature (°C), precipitation (mm), relative humidity (%), and wind speed (km/h), which are crucial for calculating the Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and Drought Code (DC).

2.2. FWI Calculation

The FWI was calculated using the Canadian Forest Fire Weather Index System, according to the studies [43,44]. It integrated daily meteorological variables into a series of subcomponents that represent different aspects of fire behavior and fuel moisture conditions. The FWI required four key meteorological inputs: daily mean air temperature (T), daily relative humidity (RH), daily wind speed (W), and daily total precipitation (P). The first step in calculating the FWI involved the computation of the FFMC for the study period, which estimated the moisture content in fine surface fuels. The calculation FFMC begins with the moisture content of fine fuels from the previous day (M₀), which serves as the baseline. If precipitation occurs, only rainfall exceeding 0.5 mm is considered “effective”, and the moisture content is adjusted using an empirical rainfall adjustment formula. This adjustment is given by

$$M_r=M_0+1000(1-\exp \left(-0.23P\right))$$

(1)

where M_r is the adjusted moisture content, and P represents the effective precipitation in millimeters. This step accounts for the wetting of fine fuels due to precipitation. In the absence of rainfall or after adjusting for it, the moisture content is further modified to reflect the drying effects of weather conditions. The drying rate is governed by the following equation:

$$M=M_r+k_{a}D$$

(2)

where M is the final moisture content, k_a is the drying rate constant, and D is the day-length factor, which varies with latitude and season. The drying rate constant (k_a) is empirically derived as

$$k_a=0.424\times {\left(1-{\displaystyle \frac{RH}{100}}\right)}^{0.5}+0.0694W^{0.5}$$

(3)

Finally, the updated moisture content is converted into the FFMC scale using

$$FFMC=59.5\times {\displaystyle \frac{250-M}{147.2+M}}$$

(4)

This transformation ensures that the FFMC is expressed on a scale ranging from 0 to 101, where higher values signify drier and more flammable conditions. By incorporating both wetting and drying processes, the FFMC provides a dynamic representation of the fine fuel moisture status, which is critical for assessing wildfire potential under varying weather conditions. The next component, the DMC, represents the moisture content in loosely compacted organic soil layers. It was calculated daily based on P and temperature-driven drying (K). The calculation incorporates two primary cases, depending on whether the daily precipitation exceeds the threshold of 1.5 mm. The calculation incorporates two primary cases, depending on whether the daily precipitation exceeds the threshold of 1.5 mm. The general form of the equation is

$${DMC}_t=\left\{\begin{array}{c}{DMC}_{t-1}+100k, if P\le 1.5\\ {DMC}_r+100k, if P>1.5\end{array}\right.$$

(5)

where DMC_t is the Duff Moisture Code for the current day, DMC_t−1 is the Duff Moisture Code from the previous day, DMC_r is the rain-adjusted Duff Moisture Code, and K is the drying factor. Following the DMC, the DC was calculated including the adjustments based on P and temperature-driven evaporation (V), ensuring a dynamic and accurate estimation of soil dryness.

$${DC}_t=\left\{\begin{array}{c}{DC}_{t-1}+0.5V, if P\le 2.8\\ {DC}_r+0.5V, if P>2.8\end{array}\right.$$

(6)

where DC_t is the current day’s Drought Code, DC_t−1 is the previous day’s Drought Code, and DC_r is the rain-adjusted Drought Code. The evaporation term, i.e., V, reflects the drying effect of T and was calculated as

$$V=0.36\times \left(T+2.8\right)$$

(7)

If P > 2.8 mm, it was considered effective rainfall, and the DC_r was computed as

$${DC}_r={DC}_{t-1}\times e^{(-0.02\times \left(P-2.8\right))}$$

(8)

The final DC was updated by adding the evaporation term 0.05 V to either the rain-adjusted or the previous day’s Drought Code, capturing both drying and wetting processes to dynamically reflect long-term soil moisture conditions. Once the fuel moisture codes were calculated, the Initial Spread Index (ISI) was derived by combining the FFMC with wind speed to estimate the potential rate of fire spread. The ISI was computed as

$$ISI=0.208\times f\left(W\right)\times f\left(FFMC\right)$$

(9)

where f(W) represents wind function and f(FFMC) represents fuel moisture function. Following ISI, the Buildup Index (BUI) was calculated with the DMC and DC to represent the total amount of available fuel. It was calculated based on two scenarios, which are as follows:

$$BUI=\left\{\begin{array}{c}0.8\times {\displaystyle \frac{DMC\times DC}{DMC+0.4DC}};if DMC\le 0.4DC\\ DMC-\left(1-{\displaystyle \frac{0.8DC}{DMC+0.4DC}}\right)\times \left[0.92+{\left(0.0114\times DMC\right)}^{1.7}\right];if DMC>0.4DC\end{array}\right.$$

(10)

Finally, the FWI was calculated by combining the ISI and the function of BUI (f(D)) to represent the overall intensity of a potential wildfire. The FWI is given by

$$FWI=0.1\times ISI\times f\left(D\right)$$

(11)

The validation of FWI calculations was performed using fire radiative power (FRP) data from MODIS, implementing randomized sampling to ensure robustness and minimize bias. FWI values were computed using high-resolution meteorological data and spatially matched with MODIS FRP observations. The analysis revealed a strong positive correlation between FWI and FRP, with a correlation coefficient of 0.75, indicating that higher FWI values correspond to higher fire radiative power, which is identical to the existing literature [45]. As MERRA-2 and GEOS-5 datasets do not share the same resolution as the FWI calculations, they were excluded from this comparison.

2.3. Model Development

2.3.1. GNN-TCNN, GNN-LSTM, GNN-DeepAR, and Random Forest

This study employed a hybrid GNN-TCNN, GNN-LSTM, and GNN-DeepAR model to process meteorological variables and forecast the FWI for a 14-day horizon. The input data consist of meteorological variables, including T, P, RH, and W, obtained from the gridMET dataset for 2014–2023. The target variable is the FWI. The data are structured into grid cells (n) with a spatial resolution of 4 km × 4 km, and each cell contains 14-day sequences of input features (X):

$$X_i^{\left(t\right)}=\left[T^{\left(t\right)},P^{\left(t\right)},{RH}^{\left(t\right)},W^{\left(t\right)}\right], fort=t_1,t_2,t_3,\dots ,t_{14}$$

(12)

where i is the grid cell index, and t represents the time step. Each input sequence X_i is paired with the corresponding 14-day FWI forecasts (Y_i) as the target:

$$Y_i=[{FWI}^{\left(t1\right)},{FWI}^{\left(t2\right)},{FWI}^{\left(t3\right)},\dots ,{FWI}^{\left(t14\right)}]$$

(13)

To enhance model convergence, the input features were normalized using Min-Max scaling. The spatial relationships between grid cells in the dataset are represented using a graph structure $G=\left(\nu ,\xi \right),$ where $\nu$ is the set of nodes, each corresponding to a grid cell, and $\xi$ is the set of edges representing the spatial connections between these nodes. The graph’s adjacency matrix, i.e., $A\in {\mathbb{R}}_{n\times n}$, where n is the total number of grid cells, encodes the spatial proximity between nodes. To define the edges, a threshold distance $\epsilon$ is applied: nodes i and j are connected if their Euclidean distance $d_{i,j}$ is less than $\epsilon$. The edge weights are computed as the inverse of the Euclidean distance:

$$A_{i,j}=\left\{\begin{array}{c}{\displaystyle \frac{1}{d_{ij}}}, if d_{ij}<\epsilon \\ 0,Otherwise.\end{array}\right.$$

(14)

This ensures that closer grid cells have stronger connections, while distant grid cells do not interact directly. The graph structure captures the spatial relationships to find out the influence of neighboring regions’ fire weather conditions. The degree matrix $D\in {\mathbb{R}}^{n\times n}$ is a diagonal matrix that is used in graph convolution operations to normalize node representations, preventing scale variations across nodes.

The GNN component processes the constructed graph to extract spatial dependencies among grid cells. Each node i is initialized with a feature matrix $X_i\in {\mathbb{R}}^{14\times d}$, where d is the number of meteorological features. The GNN uses graph convolution layers to update the node representations by aggregating information from neighboring nodes. The update rule for the l-th layer of the GNN is given by

$$H_i^{(l+1)}=\sigma ({\sum }_{j\in N\left(i\right)}{\displaystyle \frac{1}{\sqrt{D_{ii}{D}_{jj}}}}{A}_{ij}{H}_j^{\left(l\right)}{W}^{\left(l\right)})$$

(15)

where $H_i^{\left(l\right)}\in {\mathbb{R}}^{d_l}$ is the node representation of node i at layer l, N(i) is the set of neighbors of node i, $A_{ij}$ is the weight of the edge between nodes i and j, D_ii is the degree of node i, $W^{\left(l\right)}\in {\mathbb{R}}^{d_l\times d_{l+1}}$ is the trainable weight matrix for layer l, σ is the activation function, and d_l is the feature dimension at layer l. After L layers of graph convolution, the final node representations $H_i^{\left(L\right)}\in {\mathbb{R}}^{d_L}$ encode the spatial relationships between grid cells. These spatially enhanced features are passed to the TCNN, LSTM, and DeepAR for temporal modeling.

Following the extraction of spatial relationships by the GNN, the spatially enhanced features $H_i^{\left(L\right)}$ are concatenated with the original meteorological features $X_i$ for each grid cell i, creating a unified temporal sequence $Z_i\in {\mathbb{R}}^{14\times (d_L+d)}$, where $d_L$ is the dimension of spatial features from the GNN and d is the number of meteorological features. For each grid cell i, the concatenated input sequence $\left\{Z_i^{\left(1\right)},Z_i^{\left(2\right)},\dots , Z_i^{\left(14\right)}\right\}$ is fed into the LSTM network. At each time step t, the LSTM processes the input $Z_i^{\left(t\right)}$ alongside its internal states, the hidden state $h_{t-1}$ and cell state $c_{t-1}$, to compute the updated states $h_t$ and $c_t$ as sequences, i.e., forget gate (determines which information to discard from the previous cell state), input gate (decides what new information to store in the cell state), cell state update (updates the cell state by combining the old state and new candidate values), and output gate (determines which information to pass to the next hidden state). At the end of the sequence, the hidden state from the final time step ($h_{14})$ encapsulates the temporal dependencies across the 14-day input sequence for grid cell i. This hidden state is passed through a fully connected layer to produce the 14-day FWI forecast:

$${\widehat{Y}}_i=W_{out}\times h_{14}+b_{out}$$

(16)

where ${\widehat{Y}}_i\in {\mathbb{R}}^{14}$ is the predicted FWI values for the 14-day horizon, $W_{out}\in$ ${\mathbb{R}}^{h_d\times d}$ is the output layer weights, $b_{out}\in {\mathbb{R}}^{14}$ is the bias vector, and $h_{14}\in$ ${\mathbb{R}}^{h_d}$ is the final hidden state of the LSTM.

Unlike recurrent architectures like LSTMs, the TCNN leverages convolutional layers to capture temporal patterns, allowing for faster computation and parallelism. By using dilated convolutions, the TCNN effectively captures long-range dependencies while maintaining computational efficiency. Regarding TCNN, for each grid cell i, the temporal input sequence is represented as

$$Z_i=concat(H_i^{\left(L\right)},X_i)\in {\mathbb{R}}^{14\times (d_L+d)}$$

(17)

where $H_i^{\left(L\right)}\in {\mathbb{R}}^{d_L}$ is the spatially enhanced features from the GNN, and $X_i\in$ ${\mathbb{R}}^{14\times d}$ is the temporal meteorological features. The TCNN uses 1D convolutional layers to process the temporal sequence. For each convolutional layer, a kernel of size k slides across the input sequence, computing feature maps that capture temporal dependencies. The output of the convolution operation at time t is given by

$$F_i^{\left(t\right)}={\sum }_{j=0}^{k-1}{w}_j\times Z_i^{\left(t-j\right)}$$

(18)

where Fi(t) is the feature map at time t, w_j is the weight of the convolutional kernel, $Z_i^{\left(t-j\right)}$ is the input at time t−j. To capture long-term dependencies, dilated convolutions were used. A dilation factor d introduces gaps between the kernel elements, effectively expanding the receptive field of the convolution without increasing the number of parameters. The dilated convolution is defined as

$$F_i^{\left(t\right)}={\sum }_{j=0}^{k-1}{w}_j\times Z_i^{\left(t-j.d\right)}$$

(19)

By stacking multiple dilated convolutional layers with increasing dilation factors, the TCNN captures both short-term and long-term temporal patterns. To improve learning stability and mitigate vanishing gradient issues, residual connections are added between layers:

$$F_{residual}^{\left(t\right)}=F_{input}^{\left(t\right)}+F_{output}^{\left(t\right)}$$

(20)

where $F_{input}^{\left(t\right)}$ is input to the convolutional layer, and $F_{output}^{\left(t\right)}$ is output from the convolutional layer. Residual connections allow the model to directly propagate low-level features to deeper layers, improving the learning of temporal dependencies. The TCNN output at the final layer $F_{final}^{\left(t\right)}$ summarizes the temporal dependencies in the sequence. This aggregated temporal representation is passed through a fully connected layer to generate the 14-day FWI forecast:

$${\widehat{Y}}_i=W_0\times F_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}+b_0$$

(21)

where ${\widehat{Y}}_i\in {\mathbb{R}}^{14}$ is the predicted FWI values for the 14-day horizon, $W_0\in$ ${\mathbb{R}}^{h_d\times 14}$ is the output layer weights, $b_0\in {\mathbb{R}}^{14}$ is the bias vector, and $F_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}{\in \mathbb{R}}^{h_d}$ is the final temporal representation. Following the integration of the TCNN and LSTM with the GNN, the GNN model was then integrated into the temporal modeling framework of DeepAR. Unlike its probabilistic variant, this implementation focuses on deterministic forecasting by using the predicted mean values of the output distributions as the forecasted FWI values.

The DeepAR component used the RNN, i.e., gated recurrent unit (GRU), to model sequential dependencies in the input data. At each time step t, the RNN processes the input $Z_i^{\left(t\right)}$ along with the hidden state from the previous time step $h_{t-1}$ to compute the current hidden state $h_t$. The RNN processes the sequence iteratively, allowing it to capture both short-term and long-term temporal dependencies across the 14-day forecast horizon. At each time step t, the hidden state $h_t$ is passed through a fully connected layer to generate the predicted mean ${\mu }_t$ of the FWI. This approach ensures computational efficiency and straightforward implementation, making it suitable for large-scale FWI forecasting applications.

Apart from the hybrid models, the traditional forecasting model, i.e., RF, was implemented in this study to compare the performance. Thapa et al. [46] stated that RF performed significantly better than the other traditional machine learning models to forecast daily fire radiative energy. Moreover, it is computationally faster than models such as multilayer perceptron, XGBoost, and support vector machines. Following the same structure and input of hybrid models, we trained the RF model for 14 days of forecasting of FWI.

2.3.2. Model Training

Training Pipeline of Hybrid Models

The training pipeline for the hybrid models, i.e., GNN-TCNN, GNN-LSTM, and GNN-DeepAR, was meticulously designed to optimize their ability to forecast the FWI over a 14-day horizon. The input dataset consisted of meteorological variables, including being structured into 14-day sequences for each grid cell with a spatial resolution of 4 km × 4 km. The target variable was the corresponding 14-day FWI values. Input features were normalized using Min-Max scaling to ensure numerical stability during training. The dataset was split into training (2014–2021), validation (2022), and test sets (2023), ensuring no temporal overlap to avoid data leakage. Spatial diversity in the training set was maintained by including grid cells across various geographic regions, while validation and test sets comprised unseen grid cells to evaluate the model’s generalization ability.

Each model’s parameters were initialized using Xavier initialization to ensure stable training dynamics. The training was conducted in mini batches of size 64 to balance computational efficiency and memory usage. For each batch, the GNN processed spatial relationships, and the temporal component, whether the TCNN, LSTM, or DeepAR, handled the temporal dependencies. The models generated 14-day FWI predictions during the forward pass, which were compared with the true FWI values to compute the Index of Agreement (IOA) loss. Gradients of the loss with respect to model parameters were computed via backpropagation, and the parameters were updated using the Adam optimizer with an initial learning rate of 0.001.

To prevent overfitting, dropout regularization with a rate of 0.2 was applied, and early stopping was implemented, terminating training if the validation loss did not improve for 10 consecutive epochs. The dropout rate of 0.2 was empirically determined during preliminary experiments to achieve optimal model performance. Lower rates (e.g., 0.1) resulted in overfitting, while higher rates (e.g., 0.3 or above) led to underfitting, degrading the model’s predictive accuracy. A dropout rate of 0.2 provided the best balance, effectively regularizing the model while preserving its ability to capture complex spatial–temporal dependencies in the data. Additionally, a cosine annealing scheduler was employed to dynamically adjust the learning rate during training, gradually reducing it to 1 × 10⁻⁶. Validation was performed after every epoch, and performance was monitored using root mean squared error (RMSE) and IOA metrics. This robust pipeline ensured that the models captured complex spatial–temporal dependencies while maintaining generalization performance.

Hyperparameters and Computational Setup

The hyperparameters for the GNN-TCNN, GNN-LSTM, and GNN-DeepAR models were fine-tuned through grid search to optimize their performance. The optimal configurations for each model are summarized in

Table 1. Hyperparameter configurations for the GNN-TCNN, GNN-LSTM, and GNN-DeepAR.

Hyperparameter	GNN-TCNN	GNN-LSTM	GNN-DeepAR
Learning Rate	0.001	0.001	0.001
Batch Size	64	64	64
Optimizer	Adam	Adam	Adam
Dropout Rate	0.2	0.2	0.2
GNN Layers	3	3	3
GNN Hidden Dimension	64	64	64
Temporal Layers	3 (Conv1D)	2 (LSTM)	2 (GRU)
Temporal Hidden Dim.	128	128	128
Kernel Size (TCNN)	333	-	-
Dilation Factor (TCNN)	1, 2, 41, 2, 41, 2, 4	-	-
Sequence Length	14 days	14 days	14 days
Epochs	100	100	100

.

The hyperparameters of the RF model were optimized to achieve a balance between predictive performance and computational efficiency. A grid search approach identified the optimal configuration, with 500 trees ensuring sufficient ensemble diversity for stable predictions. The maximum tree depth was set to 20 to prevent overfitting while maintaining adequate model complexity. The minimum samples required to split an internal node were set to 10, while the minimum samples per leaf node were configured to 5, ensuring a balance between model flexibility and robustness. A random state of 42 was used to ensure reproducibility by controlling the bootstrapped subsampling and feature selection during training. These hyperparameter settings were chosen for their ability to minimize overfitting and maintain reliable performance across spatially and temporally diverse datasets.

The training process was implemented in PyTorch v2.3.0, with PyTorch Geometric used for the GNN component. The training was conducted on an NVIDIA A30 GPU, manufactured by NVIDIA Corporation (Santa Clara, USA), with 12 GB of memory, ensuring efficient processing of the large-scale spatial–temporal dataset. This computational setup enabled the models to handle the complexity of the data while maintaining computational efficiency. By combining optimized hyperparameters, regularization techniques, and GPU acceleration, the models were effectively trained to forecast FWI with high accuracy and generalizability.

Model Evaluation

The performance of the hybrid models, i.e., GNN-TCNN, GNN-LSTM, and GNN-DeepAR, was rigorously evaluated using RMSE and IOA. RMSE quantifies the average magnitude of prediction errors, providing an interpretable measure of model accuracy. IOA, on the other hand, evaluates the degree of agreement between observed and predicted values, capturing both error magnitude and variance. The use of these metrics ensures a rigorous and comprehensive assessment of the model’s predictive capabilities, particularly for continuous variables like the FWI. The mathematical expressions of RMSE and IOA are expressed as below:

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\displaystyle \frac{1}{14}}{\sum }_{i=1}^{14}{(Y_i^{\left(t\right)}-{\widehat{Y}}_i^{\left(t\right)})}^2}$$

(22)

$$IOA=1-{\displaystyle \frac{{\sum }_{i=1}^n{\sum }_{t=1}^{14}{(Y_i^{\left(t\right)}-{\widehat{Y}}_i^{\left(t\right)})}^2}{{\sum }_{i=1}^n{\sum }_{t=1}^{14}{(\left|Y_i^t-\overline{Y}\right|+\left|{\widehat{Y}}_i^{\left(t\right)}-{\overline{Y}}_i\right|)}^2}}$$

(23)

where n is the total number of grid cells, $Y_i^t$ and ${\widehat{Y}}_i^{\left(t\right)}$ are the observed and predicted FWI values for grid cell i at time t, respectively, and Y_i is the mean observed FWI value for grid cell i over the 14-day forecast horizon. The models were evaluated on an independent test set comprising unseen grid cells to ensure generalization across spatially diverse regions.

Sensitivity Analysis

In the context of FWI forecasting, permutation sensitivity analysis was adapted to evaluate the impact of each feature on the predictive performance of the model across spatial grids. For the grid-based dataset, let X = {x_i,j,k} represent the input features for each grid cell (i,j) where k indexes the features. Let Y = {y_i,j} represent the true FWI values for the grid cells. The baseline performance of the model (M) is given by

$$L={\displaystyle \frac{1}{n}}\sum _{(i,j)}\mathcal{L}\left(y_{i,j},{\widehat{y}}_{i,j}\right)$$

(24)

where ${\widehat{y}}_{i,j}$ = M (x_i,j,1, x_i,j,2, …, x_i,j,k) is the predicted FWI for grid cell (i,j); $\mathcal{L}$ is the loss function, and n is the total number of grid cells. For a specific feature f_k, its values are permuted across all grid cells to produce $X_{perm}^{f_k}$, while other features remain unchanged. The model was evaluated on this permuted dataset:

$$L_{perm}^{f_k}={\displaystyle \frac{1}{n}}\sum _{(i,j)}\mathcal{L}\left(y_{i,j},{\widehat{y}}_{i,j}^{f_k}\right)$$

(25)

where ${\widehat{y}}_{i,j}^{f_k}=M$($x_{i,j,1},x_{i,j,2},\dots ,{\tilde{x}}_{i,j,k},\dots , x_{i,j,m}$) and ${\tilde{x}}_{i,j,k}$ represent the permuted values of f_k for grid cell (i,j). The importance of f_k is measured by the change in loss between the baseline and the permuted dataset.

3. Results

3.1. Descriptive Statistics of FWI Across CONUS Regions (2014–2023)

The daily FWI from 2014 to 2023 exhibited substantial spatial and temporal variability across the nine climatic regions of the CONUS. Figure 3 illustrates the spatial distribution of three key FWI metrics, i.e., mean, maximum, and minimum. Descriptive statistics, i.e.,

Table 2. Summary statistics of the FWI across nine climatic regions in the CONUS (2014–2023). The table includes the mean, standard deviation (Std), maximum (Max), and minimum (Min) FWI values for each region. Additionally, the date, latitude, and longitude corresponding to the maximum FWI values are provided, highlighting spatial and temporal variability in fire weather conditions.

Region	Mean	Std	Max	Min	Date of Max FWI	Latitude	Longitude
Northeast	2.65	3.86	57.44	0.0	6 September 2016	39.008	−77.188
Northern Rockies	19.64	16.60	194.23	0.0	7 July 2020	42.248	−108.112
Northwest	19.29	18.19	171.64	0.0	7 September 2020	47.0	−118.768
Ohio Valley	5.47	7.35	83.08	0.0	8 October 2016	35.264	−89.688
South	16.60	18.65	385.19	0.0	15 December 2021	38.0	−101.272
Southeast	5.85	8.09	276.41	0.0	10 September 2017	25.508	−81.184
Southwest	33.88	23.25	279.99	0.0	15 December 2021	38.0	−102.064
Upper Midwest	5.97	6.91	84.11	0.0	14 September 2017	41.204	−95.908
West	36.60	24.06	204.24	0.0	8 September 2020	35.264	−114.988

, revealed that the West region has the highest mean FWI (36.60) and a large standard deviation (24.06), reflecting persistent and highly variable fire weather conditions. Similarly, the Southwest shows a high mean FWI (33.88) with considerable variability (standard deviation of 23.25), aligning with its arid climate and high susceptibility to wildfire activity. In contrast, regions such as the Northeast (mean FWI of 2.65) and the Ohio Valley (mean FWI of 5.47) reported significantly lower FWIs, consistent with their relatively moist climates and lower fire risk.

Regions like the Northern Rockies and the Northwest exhibited moderately high mean FWIs (19.65 and 19.29, respectively), with standard deviations around 16, highlighting consistent but less extreme fire weather conditions. The South displayed a mean FWI of 16.60, yet its extreme variability (maximum FWI of 385.19) indicated sporadic but severe fire events. The Southeast and Upper Midwest had low average FWIs (5.85 and 5.97, respectively), but their occasional peaks (276.41 in the Southeast and 84.11 in the Upper Midwest) suggested the presence of localized, high-impact fire weather conditions.

To quantify the occurrence of extreme fire weather, FWI values exceeding 90 were identified as extreme events. The Southwest recorded the highest number of extreme events (3,343,223), followed by the West (1,987,358) and the South (1,070,693). The Northern Rockies and Northwest exhibited 387,491 and 132,629 extreme events, respectively, indicating moderate but notable fire weather risks. In contrast, the Southeast reported only 475 extreme events.

3.2. Maximum FWI Days and Regional Peaks

Analyzing the days with the maximum FWI in each region provides further insights into extreme fire weather conditions. The highest recorded FWI, 385.19, occurred on 15 December 2021, in the South. This extreme event, observed in late fall, highlights the potential for critical fire weather conditions during non-traditional fire seasons under anomalous meteorological conditions such as prolonged droughts and dry winds [47]. The same day, an FWI of 279.99 in the Southwest was also observed, which is indicative of widespread extreme fire weather across multiple regions [47].

Other significant maximum FWIs included 276.41 in the Southeast on 10 September 2017, coinciding with severe drought conditions, and 204.24 in the West on 8 September 2020, during an active wildfire season driven by heatwaves and low humidity. In the Northern Rockies and Northwest, the highest FWIs (194.23 and 171.64, respectively) occurred in July 2020, reflecting their susceptibility to peak fire weather during summer months. The Ohio Valley and Upper Midwest reported their maximum FWIs (83.08 and 84.11) in fall 2016 and late summer 2017, respectively, suggesting localized fire risk spikes driven by transient dry weather. The occurrence of extreme events during non-traditional fire seasons, such as winter in the South and fall in the Southeast, highlighted the influence of anomalous climatic conditions, including prolonged droughts and heatwaves, in shaping fire weather dynamics.

3.3. Seasonal Variations in Fire Weather Index Across the CONUS

Seasonal variability in the FWI provides critical insights into the temporal distribution of wildfire risk across the nine climatic regions of the CONUS. The monthly averaged FWIs from 2014 to 2023 revealed clear seasonal trends, highlighting both regional patterns and inter-annual fluctuations. This section analyzed these seasonal variations using heat maps and spatial distribution maps.

3.3.1. Monthly and Annual Trends Across Regions

Figure 4 illustrates the seasonal progression of FWI values in each region, highlighting the interplay between climatic conditions and wildfire risks. The Southwest and West exhibited the highest FWIs during the summer months, peaking in June, July, and August, with values exceeding 50 in the Southwest and 40 in the West. Notably, these regions also experience high variability, with inter-annual fluctuations reflecting the influence of anomalous climate events such as droughts and heatwaves [48].

In contrast, the Northeast and Ohio Valley showed significantly lower FWIs throughout the year, with minor peaks during the fall months (September and October) due to drier conditions following summer [49]. The Southeast followed a similar pattern, with low FWIs across most months, except for occasional spikes in late summer and early fall, driven by transient dry spells. The Northern Rockies and Northwest exhibited moderate FWIs, peaking in July and August, which correspond to their dry summer periods [50]. The South stands out for its distinct seasonal peak in August, where FWIs exceeded 30 in some years, driven by high temperatures and prolonged dry periods [51]. However, this region also showed notable inter-annual variability, with significant spikes in FWIs observed during years of extreme weather events.

The heatmaps in Figure 4 also capture inter-annual variability in the FWI, revealing significant differences in wildfire risk between years. For instance, the Southwest experienced particularly high FWIs during the summer months of 2018 and 2020, corresponding to widespread drought conditions [52,53]. Similarly, the West recorded elevated FWIs in 2017 and 2021, years marked by severe wildfires and prolonged heatwaves [54,55]. The Northern Rockies and Northwest also display notable variability, with some years (e.g., 2015 and 2020) showing substantially higher FWIs during summer months compared to others. The studies [56,57] showed similar inter-annual variability in CONUS.

3.3.2. Spatial Distribution of Seasonal Fire Weather

Figure 5 provides spatial maps of the monthly averaged FWI distribution across the CONUS, further emphasizing the seasonal patterns. During winter (December to February), FWI values are generally low across most regions, with only the Southwest showing moderate activity due to its persistently dry climate, which was also verified [47]. By spring (March to May), the FWI begins to rise in the Southwest and West, reflecting the gradual onset of dry conditions.

The summer months (June to August) were marked as the peak wildfire season, with widespread high FWI values across the Southwest, West, and Northern Rockies. The spatial maps revealed that the Southwest consistently records the highest FWIs, particularly in Arizona and New Mexico, where values frequently exceed 40. The West, including California and Nevada, also showed significant fire weather activity during this period. Meanwhile, the Northern Rockies and Northwest displayed moderate FWI values, driven by their dry and warm summer conditions [56].

By fall (September to November), FWI values gradually decline in most regions, although some areas, such as the South and Southeast, exhibit late-season spikes due to lingering drought conditions and high temperatures [52]. The spatial maps highlight that while the Southwest and West remain active during this period, the intensity of fire weather diminishes compared to the summer months.

3.4. Hotspots of High FWI and Risk Clustering Across the CONUS

The spatial distribution of high FWI values reveals distinct regional hotspots and risk levels across the CONUS. Figure 6 highlights the locations with the highest FWI values, predominantly concentrated in the Southwest and parts of the West. These areas, including Arizona, New Mexico, and Southern California, exhibit consistently elevated FWI values exceeding 55, reflecting prolonged periods of high temperatures, low precipitation, and favorable wind conditions that drive wildfire risks. The clustering of high FWI values in these arid and semi-arid regions aligns with the descriptive statistics and extreme event counts, which show the Southwest and West as persistent hotspots for extreme fire weather. Figure 7 further classifies these spatial patterns into distinct risk levels using k-means clustering of FWI values. Regions are categorized as low risk, medium risk, or high risk based on their historical FWI distributions. The Southwest and West dominate the high-risk category (red), emphasizing their elevated susceptibility to frequent and severe wildfire events. The medium-risk category (blue) spans areas in the Northern Rockies, Northwest, and parts of the South, reflecting their transitional fire conditions with periodic extreme events. The eastern and central regions, including the Northeast, Midwest, and Southeast, fall predominantly into the low-risk category (green), corresponding to their low average FWIs and infrequent extreme fire weather episodes.

Together, these figures illustrate the spatial heterogeneity in fire weather risks across the CONUS, providing a visual confirmation of the descriptive and statistical analyses. Regions in the high-risk category, particularly the Southwest and West, require enhanced forecasting models and resource allocation to mitigate wildfire impacts, while medium-risk regions benefit from monitoring and periodic interventions during fire seasons.

3.5. Model Results

The performances of the GNN-TCNN, GNN-LSTM, and GNN-DeepAR were evaluated using the IOA and RMSE across a 14-day forecast horizon (Figure 8 and Figure 9). The GNN-TCNN exhibited the best overall performance, with a mean IOA of 0.885 and a mean RMSE of 1.325. On Day 1, the GNN-TCNN model achieved the highest IOA of 0.95 and the lowest RMSE of 1.21, demonstrating exceptional accuracy in modeling the spatial and temporal patterns of the FWI. The GNN-LSTM model followed closely, with an IOA of 0.93 and an RMSE of 1.25, indicating robust short-term predictive capabilities. The GNN-DeepAR, while slightly trailing, achieved a competitive IOA of 0.92 and an RMSE of 1.30, reflecting its potential to capture immediate dependencies. These results highlight the GNN-TCNN’s ability to effectively model spatial and temporal dependencies in FWI forecasting. The GNN-LSTM followed with a mean IOA of 0.852 and an RMSE of 1.590, demonstrating robust performance, particularly for short-term predictions. The GNN-DeepAR showed comparatively lower performance, achieving a mean IOA of 0.8225 and the highest RMSE of 1.755, indicating challenges in maintaining accuracy over extended time horizons. These results highlight the GNN-TCNN’s ability to effectively model spatial and temporal dependencies in FWI forecasting.

Like the performance of Day 1, the GNN-TCNN consistently outperformed the other models across both short-term (days 1–7) and long-term (days 8–14) forecasts. For the first week, the GNN-TCNN and GNN-LSTM achieved similar IOA values exceeding 0.9, reflecting strong predictive accuracy. In contrast, the GNN-DeepAR achieved IOA values around 0.87, which, although lower, indicated reasonable performance. Over the longer horizons, the GNN-TCNN maintained its accuracy, with IOA values stabilizing around 0.87. The GNN-LSTM exhibited a slight decline, with IOA values dropping to 0.84, while the GNN-DeepAR showed a more pronounced decrease, with IOA values averaging 0.81. The RMSE trends mirrored these patterns. The GNN-TCNN maintained the lowest RMSE values throughout the forecast period, underscoring its precision and stability. The GNN-LSTM displayed slightly higher RMSE values, which increased gradually over the longer forecast horizons. The GNN-DeepAR exhibited the highest RMSE values, which became more pronounced as the forecast horizon extended, reflecting its limitations in capturing long-term dependencies.

The performance of the four models, GNN-TCNN, GNN-LSTM, GNN-DeepAR, and RF, for wildfire forecasting was evaluated across nine regions of the CONUS. The GNN-TCNN model consistently outperformed the other models, achieving the highest IOA values and the lowest RMSE across all regions and forecast days. For instance, in the Northeast region, the GNN-TCNN achieved an IOA of 0.94 and an RMSE of 1.20 on Day 1, which gradually declined to an IOA of 0.74 and an RMSE of 3.50 by Day 14. In contrast, the GNN-LSTM, GNN-DeepAR, and RF exhibited slightly lower IOA values, with RF consistently showing the poorest performance. This trend was consistent across all regions.

In addition to predictive performance, the computational efficiency of the models was assessed. The GNN-TCNN required an average training time of 3.2 h, compared to 2.8 h for the GNN-LSTM, and 1.5 h for the GNN-DeepAR. Despite its longer training time, the GNN-TCNN’s superior accuracy and stability justify its computational cost. The reduced training time of GNN-DeepAR reflects its relatively simpler architecture; however, its lower performance indicates that this model may not be ideal for applications requiring high accuracy and reliability in terms of FWI forecasting. These results place the GNN-TCNN among the most robust frameworks for wildfire forecasting compared to previous studies (Figure 10).

In FWI forecasting, feature importance elucidates the contribution of individual variables to the predictive capacity of the model. This understanding is essential for identifying the key drivers of wildfire risk and optimizing forecasting methodologies. We employed permutation for the sensitivity analysis for the best model in this case, i.e., the GNN-TCNN. Due to its simplicity, scalability, and alignment with the evaluation of global feature importance in complex models like the GNN-TCNN, we chose the permutation method over LIME and SHAP analysis. Permutation importance evaluated the effect of a feature on the model’s performance by randomly shuffling its values and observing the corresponding change in the model’s output metrics. This method directly quantifies the loss in predictive accuracy caused by the perturbation, providing a clear and interpretable measure of a feature’s contribution. Temperature emerged as the most influential variable, with an importance score of 0.35 (Figure 11), where RH ranked second (0.25). Wind speed (0.20) contributed to fire spread dynamics, while precipitation (0.10) reflected short-term moisture replenishment. Temporal variables such as month (0.08), day (0.05), and week (0.02) provided seasonal and diurnal context for FWI predictions.

The integration of the forecasted FWI into fire danger rating systems provides a crucial implication of this study, underscoring its utility in real-world wildfire risk management. Using the forecasted FWI, a fire danger rating map was generated (Figure 12), categorizing the CONUS region into levels of fire risk: No Danger, Low, Moderate, High, Very High, Extreme, and Very Extreme. This spatial classification of fire danger allows for actionable insights, enabling stakeholders to identify areas at immediate risk and allocate resources effectively. For instance, the fire danger rating for 30 June 2023 illustrates the practical implications of FWI forecasting. This approach exemplifies how predictive modeling can transition into operational tools, bridging the gap between research and application.

4. Discussion

The findings revealed significant spatial and temporal variability in the FWI across the CONUS, driven by regional climatic and environmental factors. Arid regions such as the Southwest and West exhibited persistently high FWI values and frequent extreme events, highlighting their vulnerability to prolonged droughts, elevated temperatures, and favorable wind conditions. Conversely, the Northeast and Ohio Valley reported lower mean FWI values, consistent with their relatively moist climates, though episodic peaks indicate the influence of transient meteorological anomalies, such as droughts and heatwaves, on localized wildfire risk. The seasonal analysis represented the pronounced wildfire risk in the Southwest and West, particularly during the summer months when climatic conditions converge to create ideal fire weather [46]. These regions not only exhibited the highest FWIs but also experienced significant inter-annual variability, reflecting the influence of large-scale climate patterns such as the El Niño–Southern Oscillation (ENSO) and prolonged drought events [58]. The moderate but consistent summer peaks in the Northern Rockies and Northwest highlight their susceptibility during dry seasons, while the low but episodic risks in the Northeast, Southeast, and Ohio Valley emphasize the importance of monitoring localized conditions.

Therefore, this study integrates hotspot identification and risk clustering, which offers a comprehensive understanding of fire-prone regions, emphasizing the need for the tailored forecasting of wildfire potential. Among the implemented models for forecasting the FWI, the GNN-TCNN consistently outperformed other models, showcasing its ability to effectively model spatial and temporal dependencies in FWI forecasting. Its superior accuracy, reflected in the highest IOA and lowest RMSE across forecast horizons, establishes the GNN-TCNN as a robust framework for wildfire prediction. The model’s performance stability over long-term forecasts highlights its scalability and suitability for operational applications. Comparatively, the GNN-LSTM demonstrated strong short-term accuracy but showed a gradual decline in performance over extended horizons, underscoring its limitations in capturing long-term dependencies. The GNN-DeepAR, while computationally efficient, exhibited the lowest accuracy, reflecting challenges in modeling the complexity of wildfire dynamics across the CONUS. Regional variations further emphasize the need for appropriate forecasting approaches. The Southwest and other regions with complex wildfire dynamics revealed sharper declines in IOA, highlighting the challenges of predicting in areas with highly variable conditions. In contrast, more stable regions like the Northeast showed slower performance degradation, suggesting opportunities for regional optimization.

These results place the GNN-TCNN among the most robust frameworks for wildfire forecasting compared to previous studies. For instance, Son et al. [23] utilized a hybrid CFS-SR model that achieved RMSE values between 1.0 and 1.5 for test regions in the Western U.S. at a similar 4 km resolution. While the CFS-SR model successfully integrated physical and statistical methods, the GNN-TCNN’s ability to directly model spatial and temporal dependencies led to superior consistency across the entire CONUS, addressing a broader range of meteorological and climatic conditions.

Chen et al. [25] employed LSTM-based models for short-term fire danger forecasting in Northeastern China, achieving an accuracy of 87.5%. However, the reliance on recurrent layers limited their capacity to model long-term dependencies effectively. In contrast, this study’s GNN-TCNN maintained robust performance across both short-term and long-term horizons, with IOA values exceeding 0.87 even for extended forecasts. The spatial generalizability of the GNN-TCNN also surpasses that of Chen et al. [36], whose findings were geographically confined.

Shmuel et al. [26] explored machine learning models like RF and XGBoost for wildfire danger classification, focusing on feature importance and decision tree-based methods. Their models achieved high accuracy for localized scenarios, but the lack of spatial dependencies in tree-based methods made them unsuitable for regional or national-scale forecasting. Compared to their RMSE of approximately 1.8, the GNN-TCNN’s lower RMSE of 1.325 highlights the advantage of leveraging spatial and temporal relationships inherent in the data.

Thapa et al. [46] applied data-driven methods, including LSTM and Random Forest, to predict daily fire radiative energy (FRE). While their LSTM model excelled in capturing temporal dynamics, the RMSE values ranged between 1.5 and 2.0, reflecting challenges in adapting to varying fire weather conditions. This study’s GNN-TCNN outperformed Thapa et al.’s models in both accuracy and scalability, particularly in regions with high FWI variability, such as the Southwest U.S. The GNN-TCNN in this study demonstrated superior long-term robustness, with an RMSE increase that was more gradual over the 14-day horizon, further validating its ability to capture extended temporal dependencies effectively.

While the GNN-TCNN demonstrated superior performance, its higher computational cost compared to simpler models like RF and GNN-DeepAR necessitates careful consideration for large-scale applications. Future work should address these limitations by exploring ensemble methods, integrating additional data sources such as satellite-derived fire indices, and incorporating probabilistic forecasting to improve decision-making under uncertainty. This study’s reliance on meteorological variables also highlights the need for expanding datasets to regions with sparse observations.

In terms of the model’s explainability, this study utilized the permutation approach to explore the importance of features. Among the features, temperature was the most influential variable. Its dominance stems from its direct impact on fuel dryness and flammability. Elevated temperatures accelerate moisture evaporation from vegetation and soil, creating conditions conducive to wildfire ignition and propagation. This variable’s strong influence underscores the critical role of thermal dynamics in fire weather prediction [59]. RH, on the other hand, ranked second among the covariates. It affects the moisture content within live and dead vegetation, a critical determinant of fuel availability for combustion. Low humidity levels increase the propensity for fire ignition and intensification, highlighting its relevance in accurately predicting the FWI under varying atmospheric conditions [60]. Following RH, WS was the third most influential factor. Its significance lies in its dual role in facilitating fire spread and supplying oxygen to flames. Higher wind speeds enhance the rate of fire spread, intensify burning, and complicate fire suppression efforts. This variable is integral to capturing the dynamic aspects of wildfire behavior in FWI forecasting [61]. Precipitation, contributing a lower importance score comparatively, provides information on short-term moisture replenishment. While precipitation can temporarily mitigate fire risk by wetting fuels and increasing soil moisture, its impact is often overshadowed by persistent dry spells that precede most wildfires. This variable remains relevant for capturing episodic reductions in fire risk [61]. Temporal variables, such as month, day, and week, contribute contextual information. The month captures seasonal variations that significantly influence fire weather conditions, such as peak fire seasons in the summer. Day-specific features account for diurnal cycles in temperature and humidity, which can influence daily fire risk. Weekly trends offer a broader temporal perspective but lack the granularity required for detailed short-term forecasting.

This study provides several unique contributions compared to the existing literature. Unlike the regionally focused studies by Son et al. [23] and Chen et al. [25], this study delivers a comprehensive framework for forecasting FWI across the entire CONUS at a 4 km resolution. Additionally, the integration of the GNN with advanced temporal models (TCNN, LSTM, DeepAR) allows for the better modeling of spatial–temporal dependencies, which are often overlooked in tree-based methods like those used by Shmuel et al. [26] and Thapa et al. [46]. Moreover, the GNN-TCNN model achieves superior accuracy and robustness compared to purely recurrent or convolutional models, providing a balance between computational efficiency and predictive performance. Future research should focus on extending forecast horizons and improving model generalizability to address the evolving challenges of wildfire prediction in a changing climate.

5. Conclusions

Wildfires pose a significant challenge to ecosystems, public health, and economies, necessitating robust and precise forecasting models for effective risk management and mitigation strategies. This study addressed this critical need by implementing three advanced deep learning models i.e., GNN-TCNN, GNN-LSTM, and GNN-DeepAR for forecasting the FWI across the CONUS. This study introduced a fire danger rating system to translate forecasted FWI values into actionable risk categories, providing a practical framework for wildfire prevention and response.

The results demonstrated that the GNN-TCNN model achieved the highest predictive accuracy, with a mean IOA of 0.885 and a mean RMSE of 1.325 over a 14-day forecast horizon. The GNN-LSTM, while effective for short-term predictions, recorded a mean IOA of 0.852 and an RMSE of 1.590, with declining performance over longer horizons. The GNN-DeepAR model, characterized by its computational efficiency, delivered the lowest accuracy, with a mean IOA of 0.822 and an RMSE of 1.755. Spatial analyses identified the Southwest as the most fire-prone region, with a mean FWI of 33.88 and peaks exceeding 279, emphasizing the need for targeted management in this area.

This study underscores the necessity of leveraging advanced forecasting tools for wildfire risk management and lays the groundwork for future research to integrate socio-economic and health impact assessments with fire danger forecasts. The findings highlight the potential for these models to inform policy decisions and enhance preparedness strategies, offering a critical step toward mitigating the growing threat of wildfires in a changing climate.