Temporal Modelling for Burn Scars on Sentinel-3

Luca Barco; Edoardo Arnaudo; Andrea Bragagnolo; Claudio Rossi; Paolo Garza

cs.CVSep 28, 2026

Temporal Modelling for Burn Scars on Sentinel-3

Authors: Luca Barco, Edoardo Arnaudo, Andrea Bragagnolo, Claudio Rossi, Paolo Garza

Organizations: Politecnico di Torino, Torino, Italy · Fondazione LINKS, Torino, Italy

Abstract

Rapid and accurate burn scar delineation from satellite imagery is essential for post-fire damage assessment. Sentinel-3 OLCI, with daily revisit and 21 spectral bands, suits rapid mapping, yet most pipelines treat acquisitions independently, leaving the pre/post-fire change signal unexploited. We present a dataset of 246 wildfire activations (2016-2025) from the Copernicus Emergency Management Service, with Sentinel-3 OLCI temporally paired acquisitions. We benchmark spatial and temporal (ConvLSTM-augmented) variants of three backbones (U-Net, SegFormer, ConvNeXt-UPerNet) under two input modes and spectral configurations. Temporal modeling improves segmentation only when pre-fire frames are included, and a 5-band subset matches the full 21-band OLCI configuration.

Figures & tables

Fig. 1 : Geographic distribution of the dataset events: 146 train, 50 validation, and 50 test events.

Model	Bands	Input	F1		IoU
Model	Bands	Input	2D	3D	2D	3D
ConvNeXt–UPerNet	5	Post-fire only	$55.86\pm 0.79$	$58.82\pm 1.02$	$47.66\pm 0.55$	$49.69\pm 0.94$
	5	Pre- and post-fire	$58.71\pm 1.65$	$59.80\pm 1.25$	$50.26\pm 0.92$	$51.39\pm 0.92$
	21	Post-fire only	$55.99\pm 0.56$	$56.92\pm 1.30$	$48.02\pm 0.45$	$48.01\pm 1.00$
	21	Pre- and post-fire	$57.26\pm 1.81$	$59.98\pm 1.42$	$49.17\pm 1.43$	$51.17\pm 1.21$
U-Net	5	Post-fire only	$63.37\pm 0.83$	$62.73\pm 1.08$	$54.75\pm 0.86$	$53.74\pm 0.79$

TABLE I : Average F1 and IoU on the fire class for all model configurations, spectral settings, temporal conditioning modes, and backbone architectures, evaluated on the test set (mean

\pm

standard deviation over three seeds; best value per column in bold). For 2D models, the pre- and post-fire input is the channel-wise concatenation of the pre-fire and the final post-fire acquisition (bi-temporal early fusion).

Panel of Figure 3 — Fig. 2 : Qualitative results. S3 OLCI input timeseries with related predictions using full-band and band-subset configurations. In each timestep column, red pixels overlaid on the S3 input indicate predicted fire pixels; the final GT column shows the ground-truth burned area the model is expected to predict. (a) Successful delineation of the growing burned area. (b) Failure case: small burned area with limited valid pixel coverage. For EMSR457, five of the eight acquisitions are shown; $t_{0}$ is the pre-fire acquisition.

Explore similar work

A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires
Sep 14, 2026Shreyan Mitra, Mohammadreza Narimani, Parastoo FarajpoorWildfire Sentinel-2
California High School, San Ramon, CA, 94583, USA · Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA
Scale-based Approach for Active Wildfire Segmentation on Satellite Imagery
Sep 1, 2026Matheus F. Kovaleski, Cristiano Premebida, João Ruivo PauloWildfire Detection Wildfire Prediction
Institute of Systems and Robotics, Department of Electrical and Computer Engineering, University of Coimbra, Portugal
CanadaFireSat: Toward high-resolution wildfire forecasting with multiple modalities
Jun 10, 2025Hugo Porta, Emanuele Dalsasso, Jessica L. McCarty +1Wildfire Prediction Multimodal Deep Learning
EPFL, Route des Ronquos 86, Sion, 1950, Wallis, Switzerland · Université Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, Grenoble, 38000, France · NASA Ames Research Center, Earth Science Division„ Moffett Field, California, 94035, USA

Sep 14, 2026, eess.IVJ/K move · Enter open · S save

A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires

Shreyan Mitra, Mohammadreza Narimani, Parastoo Farajpoor

California High School, San Ramon, CA, 94583, USA · Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA

Wildfire Sentinel-2 Satellite Imagery Vision Datasets

Open paper AskPDF arXiv

This article describes an open image dataset for developing and evaluating active-fire segmentation methods in satellite imagery. The dataset contains 2,148 image-mask pairs from 25 California wildfires, with acquisitions spanning July 2020 to August 2026. Each image is a 512x512-pixel, three-channel composite derived from Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m spatial sampling. A fixed linear rendering is applied throughout the dataset. Corresponding masks distinguish background, SWIR-rule active fire and invalid observations. The masks were generated from shortwave-infrared brightness and near-infrared contrast, followed by constrained neighborhood growth. The release includes chip-level metadata and an incident-disjoint partition containing 18 training, three validation and four test fires. Among the image pairs, 841 contain active-fire labels; these labels occupy 0.0766% of all grid cells. A mask-blind analyst review covers 233 test chips and provides a separate assessment of the rule-generated labels at chip and connected-component levels. Reference training and evaluation code accompanies the data, including a ResNet-34 U-Net implementation with validation-based checkpoint and threshold selection. The archived images, masks, metadata and review annotations support research on rare-class segmentation, learning from algorithmic labels and transfer across fire incidents. The versioned dataset is deposited on Zenodo, with preparation and reuse software maintained in a public GitHub repository.

Abstract

Explore similar work

A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires

Scale-based Approach for Active Wildfire Segmentation on Satellite Imagery

CanadaFireSat: Toward high-resolution wildfire forecasting with multiple modalities