AFIRE: Cross-Domain Evidence Integration

• AFIRE is an overloaded term describing distinct frameworks that align heterogeneous evidence, with its meaning dependent on disciplinary context.
• In wildfire initial attack prediction, AFIRE forecasts risk using discovery-time public data with metrics like AUPRC and AUROC, ensuring strict leakage control.
• In neuroimaging and atmospheric modeling, AFIRE facilitates fusion-agnostic fMRI decoding and Bayesian fire arrival time correction using satellite detections.

AFIRE is an overloaded term in contemporary research literature rather than a single established framework. In current arXiv usage, it denotes at least three distinct technical objects: a wildfire initial attack failure prediction problem and benchmark built from public discovery-time environmental data; an agnostic framework for multimodal fMRI response encoding; and a data-assimilation method for incorporating satellite active-fire detections into coupled atmosphere-fire simulation by correcting the fire arrival time field. A closely related orthographic variant, AFiRe, refers to anatomy-driven self-supervised learning for radiographic images (Xu et al., 14 Jun 2026, Yin et al., 6 Oct 2025, Mandel et al., 2014, Liu et al., 15 Apr 2025).

1. Terminological scope and disambiguation

The term AFIRE has no single cross-domain definition. Its meaning depends entirely on disciplinary context.

Usage	Domain	Defining idea
AFIRE as wildfire IA failure prediction	Wildfire ML and risk assessment	Early binary prediction of whether a newly discovered fire will escape initial attack
AFIRE as Agnostic Framework for Multimodal fMRI Response Encoding	Naturalistic neuroimaging	Standardized interface for TR-aligned post-fusion tokens from heterogeneous multimodal encoders
AFIRE as fire-detection assimilation in WRF-SFIRE	Coupled atmosphere-fire modeling	Bayesian correction of the fire arrival time field using MODIS/VIIRS detections
AFiRe	Radiographic SSL	Anatomy-driven fine-grained representation learning in chest radiographs

In the wildfire benchmark literature, AFIRE is used as a task designation: predicting whether a newly discovered wildfire will escape initial attack using only information available at or before discovery time. In neuroimaging, AFIRE is an explicit acronym for Agnostic Framework for Multimodal fMRI Response Encoding. In atmosphere-fire modeling, AFIRE is a named assimilation method operating on the fire arrival time field in WRF-SFIRE (Xu et al., 14 Jun 2026, Yin et al., 6 Oct 2025, Mandel et al., 2014).

A common misconception is to treat AFIRE as a single transferable method. The literature instead uses the term for unrelated constructs that share neither input modality nor objective function. The only reliable way to interpret the term is by domain and citation context.

2. AFIRE in wildfire initial attack failure prediction

In wildfire risk modeling, AFIRE refers to the prediction task formalized in WILDFIREIA, a U.S. national-scale benchmark for wildfire initial attack failure prediction from public, discovery-time environmental and contextual data. The benchmark aligns 38,128 naturally caused FPA-FOD wildfire events with FIRMS/VIIRS thermal detections, gridMET weather and fire-danger variables, LANDFIRE vegetation, fuel, and topography, OpenStreetMap access features, and WorldPop population density. The event unit is a wildfire event with unique fire_id, and the benchmark fixes the event unit, size-based label rule, chronological split, metrics, and forbidden-feature list to prevent leakage (Xu et al., 14 Jun 2026).

The task is formulated over wildfire events rather than pixels or detections. For event $i$, $t_i^d$ is the reported discovery time, $a_i$ is the final burned area, and $\boldsymbol{x}_i(\mathcal{S})$ denotes model input from source subset $\mathcal{S}$, restricted to information available by $t_i^d$. The dataset is written as

$$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$

where $\mathcal{C}$ is the full set of source groups. Because national public data do not consistently provide operational IA outcomes, the benchmark uses a size-threshold proxy: successful IA for fire size $\le 10$ ha with label $y_i=0$, IA failure / escaped fire for fire size $t_i^d$0 ha with label $t_i^d$1, and exclusion of intermediate fires with $t_i^d$2 ha. After filtering, the benchmark covers 2016–2020 in the contiguous United States, with IA failure rates of 7.55% in train, 6.07% in validation, and 8.32% in test, making the problem a rare-event ranking setting (Xu et al., 14 Jun 2026).

The data contract is explicitly discovery-time. FIRMS/VIIRS contributes only discovery-day thermal evidence; D+1 and later detections are excluded. gridMET supplies a D-4:D discovery window, that is, five days ending on discovery day, including temperature, humidity, wind, energy release component, burning index, and fuel moisture. LANDFIRE contributes 30 m static landscape variables, including elevation, slope, aspect, EVT, EVC, EVH, FBFM40, canopy fuel variables, fuel disturbance, and fuel vegetation layers. OpenStreetMap provides drivable-road density and fire-station proximity, while WorldPop supplies prior-year annual population density to avoid future leakage. Missing VIIRS detections are zero-filled and retained through zero-valued thermal features plus a detection indicator (Xu et al., 14 Jun 2026).

Leakage prevention is a central design principle. The benchmark removes final fire size, containment timestamps, MTBS identifiers, post-discovery detections, and label-derived variables from model inputs. Preprocessing statistics are fit on the training split only, and the chronological split is fixed at 2016–2018 for training, 2019 for validation, and 2020 for testing. Four model-ready representations are provided: tabular $t_i^d$3, temporal $t_i^d$4 plus static features, spatial $t_i^d$5, and spatiotemporal $t_i^d$6, with the spatial artifact defined as a $t_i^d$7 event-centered grid in EPSG:5070 (Xu et al., 14 Jun 2026).

Sixteen models are evaluated across tabular, temporal, spatial, and spatiotemporal families, including Logistic Regression, XGBoost, MLP, GRU, TCN, Transformer, ResNet18-UNet, ResNet50-UNet, Swin-UNet, SegFormer, ConvLSTM, ConvGRU, ResNet3D, PredRNN-V2, UTAE, and SwinLSTM. The main prediction target is

$t_i^d$8

Because IA failure is rare, the primary metric is AUPRC, with additional reporting of AUROC, Recall@5%, F1, Brier score, and ECE (Xu et al., 14 Jun 2026).

The main result is that public discovery-time data provide useful but incomplete signal. XGBoost achieves the best overall performance with AUPRC = 53.3%, AUROC = 87.1%, Recall@5% = 38.0%, F1 = 49.5%, Brier = 9.1, and ECE = 13.5. Leave-one-source-out ablations identify FIRMS/VIIRS as the least redundant source, with the largest performance drop when removed; weather adds smaller but consistent value; and vegetation, fuel, topography, access, and population are more redundant in the full-input setting. When dynamic sources are removed, fuel is the strongest static predictor. Longer weather histories do not consistently improve results, with best performance usually coming from 1–2 days of history. The benchmark also defines a containment-duration auxiliary task with $t_i^d$9 and log target $a_i$0; here the best model again is XGBoost, but $a_i$1 is only about 0.18, indicating that discovery-time public features explain only a small fraction of containment-time variance (Xu et al., 14 Jun 2026).

3. AFIRE as an agnostic framework for multimodal fMRI response encoding

In naturalistic neuroimaging, AFIRE stands for Agnostic Framework for Multimodal fMRI Response Encoding. It is designed for whole-brain prediction from multimodal stimuli such as video, audio, and text, under conditions where multimodal inputs are heterogeneous, upstream fusion styles differ, and inter-subject variability is pronounced. The framework is intended to be fusion-agnostic and subject-aware, addressing the incompatibility of latent formats produced by image-aligned, Transformer-fused, or LLM-centric backbones (Yin et al., 6 Oct 2025).

AFIRE operates as a plug-in interface that ingests time-aligned post-fusion tokens from varied multimodal encoders, standardizes them into a common representation, and exposes them to a downstream decoder. Upstream features are sampled at 2 Hz, whereas fMRI is collected at TR resolution. AFIRE aggregates the 2 Hz features inside each TR bin to form

$a_i$2

It can ingest time-aligned per-layer features from heterogeneous backbones, merge modalities through a configurable operator into a single token stream, and optionally apply a lightweight temporal module consisting of one or two self-attention layers or a temporal MLP with positional encoding and normalization. The architectural consequence is that the decoder is decoupled from the specific upstream fusion design (Yin et al., 6 Oct 2025).

The decoder used within AFIRE is MIND, the Mixture-of-Experts Integrated Decoder. MIND contains $a_i$3 parallel MLP experts

$a_i$4

where $a_i$5 can be, for example, 1000 Schaefer parcels. For token $a_i$6, prediction is given by

$a_i$7

This is a sparse, subject-aware late fusion over experts. The gating mechanism, SADGate (Subject-aware Dynamic Gating), combines a Token Router and a Subject Prior Router (Yin et al., 6 Oct 2025).

For subject $a_i$8 and token $a_i$9, the Token Router forms

$\boldsymbol{x}_i(\mathcal{S})$0

then computes routing logits $\boldsymbol{x}_i(\mathcal{S})$1 and token-dependent probabilities $\boldsymbol{x}_i(\mathcal{S})$2. The Subject Prior Router introduces a global expert-logit vector $\boldsymbol{x}_i(\mathcal{S})$3 and a subject-expert bias matrix $\boldsymbol{x}_i(\mathcal{S})$4, yielding

$\boldsymbol{x}_i(\mathcal{S})$5

These are combined by elementwise multiplication,

$\boldsymbol{x}_i(\mathcal{S})$6

followed by sparse Top-$\boldsymbol{x}_i(\mathcal{S})$7 selection and renormalization: $\boldsymbol{x}_i(\mathcal{S})$8 The paper characterizes this as delivering conditional computation, sparsity, stability, and interpretability (Yin et al., 6 Oct 2025).

Training is end-to-end with parcel-wise reconstruction loss described as MSE, a Switch-style load-balancing regularizer, and an L2 penalty on the subject-bias matrix. The system is evaluated on the Algonauts 2025 naturalistic fMRI benchmark using three multimodal backbones: TRIBE, ImageBind, and Qwen2.5-Omni. Reported metrics are Pearson $\boldsymbol{x}_i(\mathcal{S})$9, Spearman $\mathcal{S}$0, $\mathcal{S}$1, and Inter-Subject Generalization (ISG). Across all three backbones, MIND improves over both a plain MLP decoder and an MMoE decoder. The reported gains over baseline are approximately +0.017 $\mathcal{S}$2, +0.019 $\mathcal{S}$3, +0.011 $\mathcal{S}$4, +0.054 ISG for TRIBE; +0.090 $\mathcal{S}$5, +0.082 $\mathcal{S}$6, +0.038 $\mathcal{S}$7, +0.065 ISG for ImageBind; and +0.095 $\mathcal{S}$8, +0.075 $\mathcal{S}$9, +0.034 $t_i^d$0, +0.059 ISG for Qwen2.5-Omni. The paper summarizes the average gains as +0.067 Pearson $t_i^d$1, +0.065 Spearman $t_i^d$2, +0.028 $t_i^d$3, and +0.063 ISG (Yin et al., 6 Oct 2025).

Ablations compare token-only, prior-only, and combined routing; the combined router performs best across all backbones, indicating complementarity between transient token-dependent context and persistent subject preference. Interpretability analyses report similar parcel-wise high-correlation regions across TRIBE, ImageBind, and Qwen2.5-Omni when all are decoded with MIND, and time-resolved expert weights that differ across subjects and across time. The authors state that expert patterns correlate with content type and subject-specific response tendencies (Yin et al., 6 Oct 2025).

4. AFIRE as satellite-fire data assimilation in WRF-SFIRE

In coupled atmosphere-fire modeling, AFIRE is a method for assimilating MODIS and VIIRS active-fire detections into WRF-SFIRE. Its defining move is indirect assimilation: instead of treating satellite detections as a literal fire/no-fire mask, it corrects the fire arrival time field

$t_i^d$4

which records the time at which fire reaches each location. The method is motivated by the properties of existing active-fire products: detections are sparse and incomplete, false negatives are common, detection squares are coarse relative to fire-behavior models, geolocation errors can reach about 1.5 km, and cloud cover can suppress detection entirely. The paper therefore argues that such products should be used to improve fire modeling in a statistical sense only (Mandel et al., 2014).

Let $t_i^d$5 be the forecast fire arrival time, $t_i^d$6 the analysis to be estimated, $t_i^d$7 the satellite detection data, and $t_i^d$8 the overpass time. AFIRE casts the assimilation as a Bayesian inverse problem with posterior

$t_i^d$9

The likelihood uses a local log-likelihood function $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$0, where $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$1 is the time elapsed since fire arrival relative to the overpass. In discrete form,

$$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$2

Inside detection squares, recent arrival is favored; outside detection squares, non-arrival or much earlier arrival is favored. Detections at different locations are assumed independent. The prior is Gaussian,

$$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$3

where $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$4 is a discretized negative power of the Laplacian, so the prior penalizes rough corrections and favors smooth deviations from forecast. The ignition point is constrained to remain fixed (Mandel et al., 2014).

The corresponding maximum-a-posteriori estimate minimizes

$$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$5

subject to the ignition constraint. The paper interprets this as a generalized nonlinear least-squares or generalized Tikhonov regularization problem. Although the raw gradient contains the forcing derived from the likelihood term, the method instead computes a smoothed descent direction using the $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$6-inner product. Enforcing the ignition-point constraint leads to a saddle-point system, and the solution is obtained using multiplication by $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$7, which acts as a smoothing operator and can be applied efficiently, for example with FFTs (Mandel et al., 2014).

The optimization is iterative: initialize with the forecast $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$8, compute forcing from the detections, compute a smooth descent direction $$\mathcal{D} = \{(\boldsymbol{x}_i(\mathcal{C}), y_i)\}_{i=1}^{n},$$9, perform a line search along $\mathcal{C}$0, choose the minimizing step, and repeat if necessary. In the examples reported, often one iteration was enough to obtain a satisfactory improvement. The practical importance of AFIRE lies in its coupled workflow. After the corrected arrival-time field is estimated, the model can be rerun from a time in the past, regenerate heat fluxes implied by the modified fire history, spin up the atmospheric model to restore fire-atmosphere consistency, and then continue the coupled simulation from the updated state at the overpass time (Mandel et al., 2014).

The method is demonstrated on the Woods Hollow fire, using 151 MODIS detections, and on the Barker Canyon fire using VIIRS detections. In both cases, the reported analysis contour better matches the actual perimeter than the forecast. The authors emphasize that AFIRE “fills in” fire arrival time between sparse detection squares and supports cyclic corrections at successive satellite overpasses (Mandel et al., 2014).

5. The orthographic variant AFiRe in radiographic self-supervision

A distinct term, AFiRe, appears in medical imaging and should not be conflated with AFIRE. AFiRe stands for Anatomy-driven self-supervised learning for enhancing Fine-grained Representation in radiographic image analysis. It is a radiography-specific self-supervised framework for chest X-rays whose central claim is that generic image-level contrastive learning underrepresents local anatomy, spatial relationships, density patterns, and subtle lesions (Liu et al., 15 Apr 2025).

AFiRe uses a siamese ViT architecture with a student network and a teacher network updated by EMA,

$\mathcal{C}$1

where $\mathcal{C}$2 follows a cosine schedule from 0.99 to 1. Pretraining uses normal CXRs and synthetic abnormal CXRs generated by Synthetic Lesion Masks (SLMs). The two self-supervised tasks are token-wise anatomy-guided contrastive learning and pixel-level anomaly-removal restoration, with overall objective

$\mathcal{C}$3

SLM is introduced because traditional augmentations such as rotation, cropping, affine transforms, and flipping can corrupt anatomical consistency, while paste-based methods can create unrealistic lesion boundaries. SLM samples a highlighted region using Otsu binarization, resizes it to lesion sizes $\mathcal{C}$4 with $\mathcal{C}$5, shapes it as irregular ovals, and places the resulting texture-shape compositions on the image while preserving the surrounding anatomy (Liu et al., 15 Apr 2025).

The token-wise contrastive component treats each ViT token as a local anatomical unit. It defines token probabilities for student and teacher outputs, uses Sinkhorn-Knopp for teacher assignments, separates normal and abnormal tokens via the patchified SLM mask, and introduces a spatial-aware prototype matrix

$\mathcal{C}$6

so that each token position has its own prototype distribution. The restoration component replaces abnormal latent tokens with a trainable mask token and reconstructs the normal image with a weighted MSE that emphasizes anomalous regions (Liu et al., 15 Apr 2025).

AFiRe is pretrained on 811,170 CXR images total, including 81,117 normal images from MIMIC-CXR-JPG and 9 synthetic abnormal versions per normal image. Evaluation is conducted on ChildCXRs, NIH ChestX-ray8, CheXpert, and SIIM-ACR Pneumothorax. With ViT-B as backbone, a lightweight ViT decoder with 8 layers and 512 embedding dimension, and an MLP projection head with 256 hidden dimensions, pretraining runs for 800 epochs with AdamW, batch size 64, 8 NVIDIA A6000 GPUs, and 6 days. Reported downstream performance includes NIH mean AUC of 63.2 at 1% labels, 76.1 at 10%, and 83.2 at 100%; CheXpert mean AUC of 78.8, 86.3, and 89.6 at the same label fractions; and SIIM-ACR Dice of 87.7 at 1% labels and 92.4 at 100%. In anomaly detection, AFiRe reports AUC 89.4, ACC 83.9, F1 87.6 on ChildCXRs; AUC 79.9, ACC 72.4, F1 73.8 on CheXpert; and AUC 68.2, ACC 66.8, F1 60.4 on NIH (Liu et al., 15 Apr 2025).

The relevance of AFiRe to AFIRE is terminological rather than methodological. The shared letter sequence can cause bibliographic confusion, but the radiographic framework is unrelated to wildfire prediction, fire-data assimilation, or multimodal neuroimaging.

6. Comparative structure, domain boundaries, and interpretive significance

The three principal AFIRE usages occupy different positions in the modeling pipeline. The wildfire benchmark in WILDFIREIA addresses early-warning risk ranking at discovery time, using public environmental and contextual variables and strict leakage control. The WRF-SFIRE AFIRE method addresses state correction in a coupled physical simulation after satellite overpass, using Bayesian assimilation of noisy detections. The neuroimaging AFIRE framework addresses representation standardization and subject-aware decoding for whole-brain response prediction from multimodal stimuli (Xu et al., 14 Jun 2026, Mandel et al., 2014, Yin et al., 6 Oct 2025).

These usages also differ in what is being predicted or estimated. In wildfire AFIRE benchmarking, the primary output is a binary risk score approximating $\mathcal{C}$7. In WRF-SFIRE AFIRE, the latent state of interest is the continuous fire arrival time field $\mathcal{C}$8. In fMRI AFIRE, the output is voxel- or parcel-level brain activity reconstructed from TR-synchronous multimodal tokens. Their evaluation regimes are correspondingly distinct: AUPRC and related rare-event metrics for wildfire IA failure; MAP objective minimization and perimeter correction in coupled simulation; and Pearson $\mathcal{C}$9, Spearman $\le 10$0, $\le 10$1, and ISG in neuroimaging (Xu et al., 14 Jun 2026, Mandel et al., 2014, Yin et al., 6 Oct 2025).

A second misconception is that the two wildfire-related AFIRE usages are simply variants of the same framework. They are not. WILDFIREIA formalizes AFIRE as predicting whether a newly discovered wildfire will escape initial attack from discovery-time public data, whereas the WRF-SFIRE method assimilates MODIS/VIIRS detections into an ongoing atmosphere-fire simulation by modifying arrival times and replaying heat flux history. One is a supervised benchmark over historical events; the other is a model-based inverse problem embedded in online simulation (Xu et al., 14 Jun 2026, Mandel et al., 2014).

A plausible implication is that AFIRE has become a convenient acronymic label for work on heterogeneous evidence integration under operational constraints. In the wildfire benchmark, heterogeneous public data are aligned under a discovery-time contract. In neuroimaging AFIRE, heterogeneous multimodal backbones are normalized into a common post-fusion token interface. In WRF-SFIRE AFIRE, sparse and uncertain detections are converted into a smooth correction of a latent physical state. This suggests a family resemblance at the level of problem structure—alignment of incompatible evidence streams—without implying any shared algorithmic lineage.

For scholarly use, the term AFIRE therefore requires explicit qualification. In bibliography, methods, and discussion sections, the relevant paper id is not optional but essential to disambiguation: (Xu et al., 14 Jun 2026) for wildfire initial attack failure prediction, (Yin et al., 6 Oct 2025) for multimodal fMRI response encoding, (Mandel et al., 2014) for satellite-fire assimilation in WRF-SFIRE, and (Liu et al., 15 Apr 2025) for the distinct variant AFiRe in radiographic self-supervision.