From Next Day Wildfire to Next Days Wildfire: A New Multi-Continent Dataset of Active-Fire Spread for Sequence-to-Sequence Forecasting

Jose Guillen; Renzo Zanca; Valentin Barriere

cs.CVOct 28, 2026

From Next Day Wildfire to Next Days Wildfire: A New Multi-Continent Dataset of Active-Fire Spread for Sequence-to-Sequence Forecasting

Authors: Jose Guillen, Renzo Zanca, Valentin Barriere

Organizations: CENIA, Macul, Chile · Universidad de Chile, DCC, Santiago, Chile

Abstract

Benchmarks such as WildfireSpreadTS (WSTS) have enabled data-driven forecasting of wildfire spread from multimodal satellite time series, but they suffer from limitations: size restricts them to large fires that burn for weeks, and they are from similar environments. We introduce AFSeq, a multimodal dataset of 1,194 recent fire events (2023 to 2026) from 8 different countries across three regions, dominated by the short, small fires that existing benchmarks underrepresent. Our dataset is the largest and most diverse to date for time-series-based wildfire spread prediction. We also introduce UTAE-AR, an extension of a classical U-Net with Temporal Auto Encoder, enhanced with recurrent modules before the decoder. This allows the model to not only predict the next day but the whole sequence in an autoregressive way. Moreover, our model can feed its own predicted fire mask back one day at a time, producing the day-by-day forecasts that single-step models cannot. Our results show that UTAE-AR surpasses state-of-the-art on both WSTS, WSTS+ and AFSeq for next-day accuracy while having 21.5% of their parameters, and being able to predict confidently at a longer horizon.

Figures & tables

	AFSeq	WSTS [ 10 ]	WSTS+ [ 16 ]
Coverage
Years covered	2023-2026	2018-2021	2016-2023
Regions	US, S. America, S. Europe	Western US	US
Size
Fire events	$1,194$	607	$1,005$
Total daily obs.	$13,842$	$13,607$	$24,462$

Table 1: AFSeq compared with the two public SITS benchmarks.

Figure 2: Overview of UTAE-AR architecture.

Model	Seq in	Seq out	Params	AP
Persistence	✗	✗	$\emptyset\hskip 9.24994pt$	$.124\pm.013$
Single-frame
U-Net	✗	✗	0.43 M	$.242\pm.032$
ConvNeXt	✗	✗	203 M	$.254\pm.037$
Temporal
ConvLSTM	✓	✗	0.19 M	$.302\pm.028$

Table 2: Next-day prediction on AFSeq. AP on the test partition, mean and standard deviation over three seeds. Single-frame models see day

t

only; temporal models see the five-day history.

Model	Veg	Multi	All	Params
Res50-Unet [ 16 ]	$.349\pm.109$	$.351\pm.105$	$.351\pm.112$	32.5 M
Res18-UTAE [ 16 ]	$.354\pm.113$	$.363\pm.129$	$.350\pm.117$	14.7 M
UTAE-AR	$\mathbf{.429\pm.105}$	$\mathbf{.445\pm.103}$	$\mathbf{.441\pm.106}$	3.16 M

Table 3: Next-day prediction on WSTS+ under its published four-fold protocol. AP on the 128-pixel crop, reported as mean and standard deviation across folds.

Model	Trained on	AFSeq	WSTS+	Avg.
UTAE-AR	WSTS+	$.238\pm.027$	$\mathbf{.429\pm.105}$	.334
	AFSeq	$\mathbf{.354\pm.140}$	$.343\pm.096$	.349
	Both	$.346\pm.010$	$.416\pm.118$	.381
Res18-UTAE	WSTS+	$.234\pm.012$	$.354\pm.113$	.294
	AFSeq	$.317\pm.012$	$.305\pm.087$	.311
	Both	$\mathbf{.331\pm.07}$	$\mathbf{.406\pm.111}$	.369

Table 4: Cross-dataset evaluation: AP at

t{+}1

. Rows indicate the training set used; columns indicate the test set. WSTS+ columns correspond to the mean of the 4 folds. AFSeq columns correspond to the mean of the 3 splits. The row labeled Both trains on the union of the two training sets of Veg .

Figure 3: Multi-day forecast performance across horizons of UTAE-AR.

Figure 4: Encoder representations across benchmarks. Joint UMAP of the 512-dimensional deepest-stage features. With WSTS+ training only, unseen AFSeq samples mostly fall within WSTS+ neighborhoods. Joint training reveals a distinct AFSeq mode while preserving substantial overlap.

Figure 5: Qualitative next-day predictions on AFSeq from the United States, South America, and Europe. Green, red, and yellow denote true positives, false positives, and false negatives, respectively. Black boxes highlight regions where UTAE-AR better recovers the observed fire spread than Res18-UTAE.

Channel(s)	Product	Native res.	Cadence
elevation, slope, aspect	SRTM	30 m	static
landcover (17 one-hot)	MODIS MCD12Q1 (IGBP)	500 m	annual
ndvi, evi	MODIS MOD13Q1	250 m	16 d
et	MODIS MOD16A2	500 m	8 d
lst	MODIS MYD11A1 (Aqua) *	1 km	daily
dew, rh, wind_speed, wind_dir, pressure	ERA5-Land hourly	9 km	daily

Table S5: Input channels and their source products. * When Aqua does not see the ground, LST falls back to ERA5-Land skin temperature, with a mean error of 4.0 °C against real Aqua LST, compared with 11.2 °C for a daily mean.

			Region			Start year
Seed	Split	$n$	S. Am.	S. Eur.	U.S.	2023	2024	2025	2026
5	Train	835	317	223	295	89	336	357	53
	Val	179	68	48	63	10	80	79	10
	Test	180	68	48	64	17	77	73	13
7	Train	835	317	223	295	77	351	356	51
	Val	179	68	48	63	18	71	77	13

Table S6 : Split composition per seed, by region and by start year. Region counts are identical across seeds by construction because the splits are stratified by region. The year balance instead results from random sampling and is not explicitly constrained.

Region	Events	Area (ha)	Fire days	Fire px	Persistence
		median	median (max)	median	AP
South America	453	1,088	10 (49)	7,017	0.128
S. Europe	319	625	2 (25)	1,969	0.095
United States	422	1,327	5 (42)	4,227	0.122

Table S7: Per-region coverage of AFSeq. Fire days counts only days with active detections; fire px is the median number of active-fire pixels per event. Persistence is the AP at

t{+}1

of copying the day

t

mask, measured over every event of the region.

Figure S6 : IGBP land cover within the event windows of each region, aggregated into fuel types.

Figure S7 : Terrain and weather conditions on the day of the fire by country. Boxes show the central 50% of the observations, with the median indicated by the line inside each box.

Figure S8 : Comparison between AFSeq and WSTS+. (a) Final fire size by range; (b) cumulative distribution on a log scale; (c) number of days with active fire.

Model	Params (M)	Predicted days	GFLOPs	Latency (ms)	ms / day	Peak mem. (MiB)
UTAE-AR (ours)	3.16	5	124.5	21.3	4.3	139
Res18-UTAE	14.7	1	11.4	14.5	14.5	120
UTAE	0.89	1	38.8	8.2	8.2	210

Table S8 : Inference cost per event on one A100 with batch size 1. UTAE-AR predicts the full five-day horizon in each forward pass, while Res18-UTAE and UTAE predict only

t{+}1

. UTAE-AR has 4.7

\times

fewer parameters than Res18-UTAE, although it requires more FLOPs.

	fold 0	fold 1	fold 2	fold 3	mean $\pm$ std
Persistence	0.134	0.067	0.182	0.233	0.154 $\pm$ 0.071
UTAE-AR, Veg	0.378	0.309	0.495	0.539	0.429 $\pm$ 0.105
UTAE-AR, Multi	0.410	0.318	0.508	0.549	0.445 $\pm$ 0.104
UTAE-AR, All	0.406	0.311	0.501	0.550	0.441 $\pm$ 0.106

Table S9 : UTAE-AR on WSTS+ per fold, for the three feature sets of Tab. 3 . Test AP at

t{+}1

on the 128-pixel crop, with five days of history as in the published protocol. The fold ordering is identical across feature sets and follows the persistence floor (

r=+0.98

Model	partition 5	partition 7	partition 23	mean $\pm$ std
Persistence	0.120	0.114	0.139	0.124 $\pm$ 0.011
U-Net	0.216	0.232	0.278	0.242 $\pm$ 0.032
ConvNeXt	0.215	0.258	0.289	0.254 $\pm$ 0.037
ConvLSTM	0.273	0.305	0.328	0.302 $\pm$ 0.028
UTAE	0.268	0.290	0.336	0.298 $\pm$ 0.035
Res18-UTAE	0.309	0.320	0.335	0.321 $\pm$ 0.013

Table S10: AFSeq test AP per partition, pooled at

t{+}1

Figure S9 : Next-day AP by fire size in AFSeq and WSTS+. Fires are grouped by final burned area into

<

1k ha, 1k-5k ha, and

>

5k ha. Error bars show the standard deviation across three splits-seeds for AFSeq and four folds for WSTS+.

Fold	Train years	Val	Test	Persistence	UTAE-AR
0	2020, 2021	2019	2018	0.192	0.511
1	2019, 2021	2020	2018	0.192	0.495
2	2019, 2020	2021	2018	0.192	0.485
3	2020, 2021	2018	2019	0.112	0.374
4	2018, 2021	2020	2019	0.112	0.357
5	2018, 2020	2021	2019	0.112	0.356

Table S11 : UTAE-AR on the original WSTS 12-fold protocol. Persistence is the copy-the-day-

t

-mask floor of the held-out year; folds sharing a test year share it by construction. AP is pooled at

t{+}1

on the full test year.

Held out	Model	LORO	Standard split	$\Delta$
South America	Persistence	0.128	–	–
	UTAE-AR (ours)	0.206	0.364 $\pm$ 0.051	$-0.158$
	Res18-UTAE	0.181	0.322 $\pm$ 0.091	$-0.141$
S. Europe	Persistence	0.095	–	–
	UTAE-AR (ours)	0.243	0.279 $\pm$ 0.053	$-0.036$
	Res18-UTAE	0.193	0.202 $\pm$ 0.017	$-0.009$

Table S12: Leave-one-region-out (LORO) evaluation on AFSeq. LORO models are trained on two regions and evaluated on the held-out region. Standard split reports performance when the held-out region is represented during training, and

\Delta

is the difference between LORO and standard-split AP.

Figure S10 : Input importance across the forecast horizon. (a) Relative AP drop when each input group is replaced by its dataset mean. (b) The same results shown across horizons, with input groups ordered by their importance at $t{+}5$ .

Trained on	Future covariates at inference	$t{+}1$	$t{+}2$	$t{+}3$	$t{+}4$	$t{+}5$
Ground-truth future covariates	Ground-truth future covariates	0.368	0.318	0.274	0.236	0.222
	GFS forecast issued on day $t$	0.368	0.268	0.236	0.198	0.192
	Day- $t$ values frozen	0.368	0.222	0.154	0.118	0.104
GFS forecast	GFS forecast issued on day $t$	0.360	0.286	0.243	0.200	0.181
	Ground-truth future covariates	0.360	0.281	0.229	0.187	0.168
	Day- $t$ values frozen	0.360	0.192	0.126	0.090	0.076

Table S13: AP across forecast horizons under different future-covariate settings on AFSeq (seed 5). The history is always observed and only future covariates change.

Figure S11 : The ten highest-AP UTAE-AR predictions at $t{+}1$ on the AFSeq test set, ordered from left to right and top to bottom. Green denotes true positives, red false positives, and yellow false negatives.

Figure S12 : The ten lowest-AP UTAE-AR predictions at $t{+}1$ on the AFSeq test set, ordered from left to right and top to bottom. Green denotes true positives, red false positives, and yellow false negatives. These cases often correspond to fragmented fires with activity distributed across disconnected regions.

Wildfire Forecasting

Wildfire forecasting predicts where fires will start, how they will spread, and their likely intensity or impacts using environmental and fire observations.

See all 28 papers

FireSenseNet: A Dual-Branch CNN with Cross-Attentive Feature Interaction for Next-Day Wildfire Spread Prediction
Apr 9, 2026Jinzhen Han, JinByeong Lee, Hak Han +2Wildfire Forecasting Convolutional Neural Networks
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
Set Prediction for Next-Day Active Fire Forecasting
May 11, 2026Yuchen Bai, Georgios Athanasiou, Xin Yu +4Wildfire Forecasting Set Prediction
Max Planck Institute for Biogeochemistry, Jena, Germany · Orion Lab, National Technical University of Athens · University of Utah, Salt Lake City, USA +4
ShearFuse-UNet: Hadamard, DCT, and Shearlet Transform Fusion for Next-Day Wildfire Spread Prediction
Jun 12, 2026Ene Meco, Yingyi Luo, Emadeldeen Hamdan +2U-Net Wildfire Forecasting
Department of Electrical and Computer Engineering, University of Illinois Chicago, Chicago, IL 60607 USA · US Forest Service, Pacific Wildland Fire Science Laboratory, Seattle, WA 98103 USA

Apr 9, 2026J/K move · Enter open · S save

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

Jinzhen Han, JinByeong Lee, Hak Han, YeonJu Na+1

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

Wildfire Forecasting Convolutional Neural Networks Cross-Attention

Open paper AskPDF arXiv

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%.

Abstract

Wildfire Forecasting

Explore similar work

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

Set Prediction for Next-Day Active Fire Forecasting

ShearFuse-UNet: Hadamard, DCT, and Shearlet Transform Fusion for Next-Day Wildfire Spread Prediction