FireSenseNet: A Dual-Branch CNN with Cross-Attentive Feature Interaction for Next-Day Wildfire Spread Prediction

Jinzhen Han; JinByeong Lee; Hak Han; YeonJu Na; Jae-Joon Lee

cs.CVApr 9, 2026

FireSenseNet: A Dual-Branch CNN with Cross-Attentive Feature Interaction for Next-Day Wildfire Spread Prediction

Authors: Jinzhen Han, JinByeong Lee, Hak Han, YeonJu Na, Jae-Joon Lee

Organizations: Department of Civil, Architectural & Environment Engineering, Sungkyunkwan University, Suwon, South Korea · AI Urban Disaster Prevention Research Center, Advanced Institute of Convergence Technology (AICT), Suwon 16229, Republic of Korea · Department of Fire Safety Engineering, Jeonju University, Republic of Korea

Abstract

Accurate prediction of next-day wildfire spread is critical for disaster response and resource allocation. Existing deep learning approaches typically concatenate heterogeneous geospatial inputs into a single tensor, ignoring the fundamental physical distinction between static fuel/terrain properties and dynamic meteorological conditions. We propose FireSenseNet, a dual-branch convolutional neural network equipped with a novel Cross-Attentive Feature Interaction Module (CAFIM) that explicitly models the spatially varying interaction between fuel and weather modalities through learnable attention gates at multiple encoder scales. Through a systematic comparison of seven architectures -- spanning pure CNNs, Vision Transformers, and hybrid designs -- on the Google Next-Day Wildfire Spread benchmark, we demonstrate that FireSenseNet achieves an F1 of 0.4176 and AUC-PR of 0.3435, outperforming all alternatives including a SegFormer with 3.8* more parameters (F1 = 0.3502). Ablation studies confirm that CAFIM provides a 7.1% relative F1 gain over naive concatenation, and channel-wise feature importance analysis reveals that the previous-day fire mask dominates prediction while wind speed acts as noise at the dataset's coarse temporal resolution. We further incorporate Monte Carlo Dropout for pixel-level uncertainty quantification and present a critical analysis showing that common evaluation shortcuts inflate reported F1 scores by over 44%.

Figures & tables

Figure 1: FireSenseNet pipeline overview. Given 12-channel geospatial inputs from Day

t

, dual-branch encoders with modality-specific kernel sizes process fuel/terrain (4 ch,

3{\times}3

) and meteorological (8 ch,

5{\times}5

) features independently. Cross-Attentive Feature Interaction Modules (CAFIM) fuse the two branches via learned spatial attention at three encoder scales. A U-Net decoder with multi-scale skip connections produces the Day

t{+}1

fire spread prediction, with Monte Carlo Dropout enabling pixel-level uncertainty estimation.

Figure 2: Visualization of a representative test sample showing all 12 input channels and the ground truth fire mask. Static fuel/terrain features (top-left four panels) exhibit fine-grained spatial structure, while dynamic meteorological variables (remaining panels) vary smoothly across the patch. The ground truth (bottom-right) shows fire pixels (red), non-fire pixels (green), and unknown pixels (gray, labeled $-1$ ). This heterogeneity motivates the dual-branch design of FireSenseNet.

Figure 3: FireSenseNet architecture. Fuel/terrain and meteorological inputs are processed by independent encoder branches with modality-appropriate kernel sizes. At each spatial scale, a Cross-Attentive Feature Interaction Module (CAFIM) generates a spatial attention map that gates the contribution of each modality before routing to the U-Net decoder. MC Dropout before the prediction head enables uncertainty quantification.

Model	Params	GFLOPs	Latency (ms)
FireSenseNet (CAFIM)	3.01M	2.52	3.1
Baseline CNN	1.15M	0.87	1.3
SegFormer	11.31M	2.44	6.8
Small Transformer	1.86M	0.60	4.6
Regularized Transformer	1.86M	0.60	4.6
Hybrid CNN+Trans	5.47M	2.20	4.1

Table 1: Computational complexity comparison. GFLOPs measured for a single

64\times 64

input; inference time averaged over 100 forward passes on a single GPU.

Model	Type	Params	Prec.	Rec.	F1	AUC-PR
FireSenseNet (CAFIM)	CNN	3.0M	0.3818	0.4609	0.4176	0.3435
Hybrid CNN+CAFIM+Trans	Hybrid	5.9M	0.3550	0.4449	0.3949	0.3155
Baseline CNN	CNN	1.2M	0.3307	0.4850	0.3933	0.3354
Hybrid CNN+Trans	Hybrid	5.5M	0.3365	0.4629	0.3897	0.3172
SegFormer	Trans	11.3M	0.2960	0.4287	0.3502	0.2699
Small Transformer	Trans	1.9M	0.2902	0.4378	0.3490	0.2677

Table 2: Performance comparison on the Next-Day Wildfire Spread test set. All metrics computed under strict next-day-only prediction with

-1

pixels excluded. Best results in bold.

Figure 4: Architecture spectrum: F1 score vs. model complexity. FireSenseNet achieves the highest F1 (0.4176) with only 3.0M parameters, while Transformer-based architectures cluster at the bottom despite higher parameter counts. The dual y-axis reveals an inverse relationship between F1 and reliance on Transformer components.

Figure 5: Qualitative prediction comparison on four test samples. FireSenseNet (CAFIM) produces spatially precise predictions matching ground truth geometries, while the SegFormer generates diffuse, over-spread probability maps. The Baseline CNN captures fire locations but with less boundary precision.

Channel	Group	$\Delta$ F1	Interpretation
PrevFireMask	Fuel	$-0.2096$	Dominant predictor; fire persistence
ERC	Weather	$-0.0189$	Integrated fire potential index
NDVI	Fuel	$-0.0170$	Vegetation fuel availability
Elevation	Fuel	$-0.0058$	Topographic channeling
Wind Speed	Weather	$+0.0011$	Acts as noise (coarse resolution)

Table 3: Feature importance via channel masking on FireSenseNet.

\Delta

F1 indicates the change when the channel is replaced by its global mean.

Figure 6: (a) CAFIM ablation: removing CAFIM and replacing it with simple concatenation reduces F1 by 7.1%. (b) Channel-wise feature importance via systematic ablation. PrevFireMask dominates (

\Delta

F1 =

-0.21

); wind speed is the only channel whose removal improves performance (green bar), indicating it acts as noise at this temporal resolution.

Figure 7: CAFIM attention maps at three encoder scales for representative test samples. The learned spatial attention $\alpha$ determines the relative contribution of fuel features ( $\alpha\to 1$ , yellow) vs. weather features ( $\alpha\to 0$ , dark). At fine scales, CAFIM focuses sharply on fire perimeters; at coarse scales, it captures broader susceptibility patterns.

Model	Clean F1	Inflated F1	Inflation
FireSenseNet (CAFIM)	0.4176	0.6033	+44.4%
Hybrid CNN+CAFIM+Trans	0.3949	0.5733	+45.2%
SegFormer	0.3502	0.5252	+50.0%

Table 4: Impact of evaluation protocol on reported F1. “Clean” uses next-day-only prediction with

-1

exclusion; “Inflated” includes previous-day fire in the target and sets

-1\to 0

Figure 8: Clean vs. inflated evaluation protocols. Including previous-day fire in the target and relabeling unknown pixels as non-fire inflates reported F1 by 44–50%, disproportionately benefiting weaker architectures and masking genuine performance differences.

Figure 9: Monte Carlo Dropout uncertainty estimation for three test samples. From left to right: previous-day fire mask, ground truth, mean prediction (20 stochastic passes), and pixel-level uncertainty ( $\sigma$ ). Uncertainty concentrates at fire perimeter boundaries, providing actionable information for prioritizing ground-truth verification during emergency response.

Explore similar work

Modular Deep Learning Mechanisms for Auditable Next-Day Wildfire Spread Prediction
Sep 15, 2026Miguel Esparza, Aydin Ayanzadeh, Ahmad Mousavi +1Wildfire Forecasting Spatiotemporal Forecasting
Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843, USA
FireGraphNet: Sub-daily wildfire spread prediction with directed graph neural networks
Oct 13, 2026Aiyshwariya Paulvannan Kanmani, Miguel Esparza, Ali MostafaviWildfire Forecasting Spatiotemporal GNNs
Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77840, United States of America
Interpretable Patch-Based Deep Learning for Wildfire Spread Prediction from Ensemble Simulations
Sep 16, 2026Marcin Lawenda, Aleksandra Krasicka, David Caballero +2Neural Surrogate Modeling Wildfire Forecasting
Poznan Supercomputing and Networking Center, Poznań, Poland · MeteoGrid, Madrid, Spain · Technical University of Czestochowa, Częstochowa, Poland

Sep 15, 2026J/K move · Enter open · S save

Modular Deep Learning Mechanisms for Auditable Next-Day Wildfire Spread Prediction

Miguel Esparza, Aydin Ayanzadeh, Ahmad Mousavi, Ali Mostafavi

Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843, USA

Wildfire Forecasting Spatiotemporal Forecasting Modular Neural Networks

Open paper AskPDF arXiv

Next-day wildfire prediction requires models whose forecasts can be evaluated alongside the assumptions and historical evidence used in their computation. Although deep learning can learn spatial patterns from remote-sensing data, predictive performance alone does not establish physical fidelity or operational trustworthiness. This study investigates three modular augmentations for next-day active-fire prediction: wind- and slope-conditioned attention biases, physics-feature retrieval-augmented output correction, and fire conditioned dual-stream gating. The attention biases expose prescribed directional preferences, while the retrieval module selects historical tiles using a nine-dimensional environmental and fire-state descriptor and applies a learned correction to a frozen model's logits. The modules are evaluated across five backbones on the Next Day Wildfire Spread benchmark, using staged ablations, directional audits, retrieval perturbations, calibration measures, and computational comparisons. The three-seed mean F1 score and area under the precision--recall curve (AUC-PR) of a SwinUNETR model with all three augmentations are 0.4216 and 0.3673. Then, a mixed ensemble (two augmented architectures and one non-augmented architecture) model achieves 0.4292 and 0.3790. Benefits vary across architectures, and retrieval-related improvements in AUC-PR do not consistently translate into higher F1. The constructed wind bias aligns closely with input wind, but its alignment with observed next-day fire displacement is much weaker, distinguishing prior inspectability from predictive physical fidelity. The study contributes a framework for exposing and evaluating selected domain-informed components within wildfire prediction models. Together, the results presented show that predictive performance, operational trustworthiness, and computational practicality need not be competing objectives.

Abstract

Explore similar work

Modular Deep Learning Mechanisms for Auditable Next-Day Wildfire Spread Prediction

FireGraphNet: Sub-daily wildfire spread prediction with directed graph neural networks

Interpretable Patch-Based Deep Learning for Wildfire Spread Prediction from Ensemble Simulations