Source-linked AI summary
Temporal Modelling for Burn Scars on Sentinel-3
Luca Barco, Edoardo Arnaudo, Andrea Bragagnolo, Claudio Rossi, Paolo Garza
TL;DR
Burn-scar pipelines often treat satellite acquisitions as independent snapshots, leaving pre/post-fire change information unused. The paper introduces TMB-S3 and evaluates spatial and recurrent segmentation with different temporal inputs. Temporal modeling improves segmentation when pre-fire frames are included, and a five-band subset matches the full-band configuration under that conditioning.
Problem
Most burn-scar pipelines treat acquisitions as independent snapshots, leaving the pre/post-fire spectral contrast unused despite Sentinel-3 OLCI’s daily revisit.
Source: §I · p. 1Method
The paper introduces TMB-S3 and compares spatial baselines with ConvLSTM temporal variants using post-fire-only or pre- and post-fire sequences.
Source: D. Temporal Conditioning Modes · p. 3, B. Results and Discussion · p. 4, §VI · p. 6Results
With pre- and post-fire frames, ConvLSTM improves over the bi-temporal SegFormer baseline by +2.14 and +2.51 F1 points for full-band and band-subset inputs, respectively; the five-band subset matches full-band across architectures.
Source: B. Results and Discussion · p. 4Takeaways & Limitations
Under temporal pre- and post-fire conditioning, a five-band surface subset matches the full-band configuration across the three architectures.
Source: B. Results and Discussion · p. 4Geographically close events from the same cluster may fall in different splits, creating possible spatial leakage.
Source: §VI · p. 6
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
The paper asks whether explicit temporal modeling improves Sentinel-3 OLCI burn-scar segmentation, addressing a snapshot-based approach that leaves pre/post-fire change information unused. It introduces a paired-acquisition dataset and finds that temporal gains depend on including a pre-fire reference.
I. INTRODUCTION
- Most burn-scar pipelines treat acquisitions as independent snapshots, leaving the spectral contrast between pre-fire vegetation and post-fire charred surface unexploited.Smoke, cloud cover, and atmospheric variability can obscure that contrast in a single acquisition.Source: §I · p. 1
- The paper introduces TMB-S3, a dataset of 246 European wildfire activations from 2016–2025 with temporally paired Sentinel-3 OLCI acquisitions.The dataset is derived from the CEMS archive and uses geographically stratified splits.Source: §I · p. 1
- Temporal modeling yields consistent gains only when pre-fire acquisitions are included; post-fire-only recurrent modeling provides no consistent benefit.The pre/post-fire ConvLSTM variants outperform the other settings of their respective architectures.Source: §I · p. 1
II. RELATED WORKS
Prior burn-area mapping largely uses single or paired snapshots, while Sentinel-3 OLCI and its daily temporal sampling remain underused for learned segmentation. The paper addresses a dataset gap with 246 CEMS wildfire activations and temporally paired imagery.
II. RELATED WORKS
- Burn-area mapping mainly uses single or paired snapshots, and whether recurrent models benefit segmentation or require a pre-fire reference remained unstudied.Sentinel-2 dominates high-resolution mapping, while Sentinel-3 OLCI was largely unexploited for learned burned-area segmentation and no large temporally paired OLCI benchmark existed.Source: §II · p. 1
III. DATASET
- The dataset draws on 246 European CEMS wildfire activations from 2016–2025, with fire polygons, sensor footprints, validity masks, event metadata, and WorldCover information.Spatially consistent AOIs are merged and intersected with sensor footprints before tiling.Source: §III · p. 2
- Resampling 10 m fire masks to the 300 m OLCI grid by logical union raises burned pixels to 21.0% of valid pixels, versus 15.7% by fractional area.This recall-favoring rule inflates burned area by 34%, mostly through a one-pixel boundary dilation; the effect is strongest for small fires.Source: §III · p. 2
- Each bbox has one pre-fire acquisition and at least one post-fire acquisition, retrieved from metadata-driven windows and filtered for invalid pixels.The median pre-fire lead time is 31 days (IQR 8–161), so seasonal or phenological changes may affect the contrast at the longest lead times; only 18 bboxes (2.3%) exceed 30 days after the event, where residual growth or regrowth may cause label noise.Source: §III · p. 2
- Event-level splits use geographic clusters and event-size bins, yielding 146 train, 50 validation, and 50 test events.The stratification covers all regions and fire sizes, but nearby events from one cluster can fall in different splits; Fig. 1 shows the event geography by split.Source: §III · p. 2, Figure 1 · p. 2
A. Problem Formulation
The task is binary pixel-wise burn-area segmentation from temporally ordered Sentinel-3 acquisitions, with predictions evaluated against burned-area masks. The study compares spatial backbones with temporal variants that process ordered frames recurrently.
A. Problem Formulation
- Each event is represented by an ordered sequence containing one pre-fire acquisition and one or more post-fire acquisitions, and the model predicts pixel-wise burn probabilities against binary masks.Inputs include either 21 or 5 OLCI bands plus a static WorldCover channel.Source: A. Problem Formulation · pp. 2–3
- Spatial models use a single tensor built from at most two acquisitions, whereas temporal models process frames in order with a recurrent encoder before segmentation.The temporal mode preserves the sequence structure while the spatial mode predicts the burned area at the final acquisition.Source: A. Problem Formulation · p. 3
B. Segmentation Architectures
- U-Net, SegFormer, and ConvNeXt-UPerNet serve as spatial baselines and as backbones for temporal variants.Each temporal variant wraps the backbone with the temporal encoder.Source: B. Segmentation Architectures · p. 3
- U-Net is a lightweight encoder-decoder trained from scratch, with three downsampling stages and a 256-channel bottleneck.Its channel widths progress from 32 to 64 to 128 before the bottleneck.Source: B. Segmentation Architectures · p. 3
- SegFormer is trained from scratch with three encoder stages using overlapping patch embeddings, spatial-reduction attention, and feedforward layers with depthwise convolution.The stages have depths (2, 2, 2) and embedding dimensions (32, 64, 160).Source: B. Segmentation Architectures · p. 3
- ConvNeXt-UPerNet combines an ImageNet-pretrained ConvNeXt-Tiny encoder with a UPerNet segmentation head.Its first convolution is adapted to each input-channel count by cyclically repeating RGB weights and rescaling them by 3/Cin.Source: B. Segmentation Architectures · p. 3
C. ConvLSTM Temporal Encoder
A ConvLSTM accumulates spatial context across temporally ordered acquisitions before passing frame representations to the segmentation backbone. The sequence provides a pre-fire landscape reference and post-fire evidence, while only final-frame predictions receive direct loss and evaluation.
C. ConvLSTM Temporal Encoder
- A single ConvLSTM cell with 64 hidden channels and a 3 × 3 kernel processes the sequence to accumulate spectral evidence and follow fire evolution.Post-fire acquisitions can differ in burn extent, cloud cover, or viewing angle; when included, the first pre-fire frame supplies a landscape reference.Source: C. ConvLSTM Temporal Encoder · p. 3
- The pre/post indicator was omitted because appending it changed pre- and post-fire F1 by less than 0.8 points in preliminary U-Net experiments.The pre-fire acquisition is always first when included, so the recurrence can identify it by position.Source: C. ConvLSTM Temporal Encoder · p. 3
- At each step, the hidden state is concatenated with the original frame and forwarded to the backbone; loss and evaluation use only the final post-fire acquisition.Earlier frames contribute through the recurrent forward pass but receive no direct gradient; their per-timestep predictions are used only for qualitative visualization.Source: C. ConvLSTM Temporal Encoder · p. 3
- Training sequences are right-padded to the longest batch sequence, with loss computed only at each sequence’s last valid step.Because the ConvLSTM is unidirectional and padding follows real frames, padded steps cannot affect valid hidden states; validation and test use batch size 1 without padding.Source: C. ConvLSTM Temporal Encoder · p. 3
D. Temporal Conditioning Modes
The section compares post-fire-only inputs with pre- and post-fire sequences, and contrasts recurrent processing of the full sequence with spatial models using a single input or a concatenated pair.
D. Temporal Conditioning Modes
- Post-fire-only mode removes the pre-fire reference, leaving the model to detect burned areas from post-onset spectral signatures alone.This is an ablation on the same sequences, not a separate event subset.Source: D. Temporal Conditioning Modes · p. 3
- Pre- and post-fire mode prepends one pre-event acquisition to post-fire imagery, providing a spectral-change cue for burn-scar delineation.The pre-fire acquisition is taken before the event begins; post-fire acquisitions start at fire onset.Source: D. Temporal Conditioning Modes · p. 3
- Temporal models process the full sequence recurrently, whereas spatial models use the final acquisition or concatenate the pre-fire reference with the final acquisition.Spatial models omit intermediate acquisitions; in the paired mode, the land-cover channel is included once.Source: D. Temporal Conditioning Modes · pp. 3–4
E. Training Objective
The training objective is pixel-wise masked binary crossentropy, with invalid observations excluded from the loss and no class re-weighting.
E. Training Objective
- Models are trained with pixel-wise masked binary crossentropy on the predicted and binary fire masks, excluding invalid pixels through a binary mask.The loss applies no class re-weighting; burned-pixel-centred crops partially mitigate fire/background imbalance.Source: E. Training Objective · p. 4
- The objective uses no class re-weighting, while burned-pixel-centred cropping partially mitigates fire/background imbalance.Source: E. Training Objective · p. 4
- A binary validity mask determines which pixels contribute to the loss.Source: E. Training Objective · p. 4
A. Experimental Setup
The benchmark compares three backbones across spatial and temporal variants, two input-conditioning modes, and two spectral configurations. Performance is assessed on final post-fire acquisitions using fire-class F1 and IoU.
A. Experimental Setup
- The benchmark comprises 72 runs across three backbones, 2D and 3D variants, two conditioning modes, two spectral configurations, and three random seeds.The spectral settings are all 21 OLCI bands or a five-band blue-to-near-infrared subset.Source: A. Experimental Setup · p. 4
- Training uses random 64 × 64 crops, balanced between burned-pixel-centred and uniformly placed crops, with AdamW for 100 epochs.The checkpoint with the best validation F1 is retained; validation is used only for checkpoint selection.Source: A. Experimental Setup · p. 4
- Fire-class F1 and IoU are averaged unweighted across 118 test bboxes, with predictions evaluated on final post-fire acquisitions at a fixed 0.5 threshold.False alarms in fire-free scenes and outside CEMS-mapped areas are not assessed; findings also hold for per-event scores.Source: A. Experimental Setup · p. 4
B. Results and Discussion
Temporal modeling improves burn-scar segmentation consistently only when pre-fire acquisitions are included. Under that conditioning, the five-band subset matches the full-band configuration, while U-Net and SegFormer outperform ConvNeXt-UPerNet in the reported comparisons.
The pre-fire reference helps spatial models, the full sequence helps further
- Bi-temporal early fusion improves on single post-fire input by +1.27 to +2.85 F1, and full-sequence ConvLSTM adds +1.09 to +3.60 F1 over bi-temporal baselines.Intermediate post-fire acquisitions may help track burn evolution and compensate for partial occlusion, but recurrence and fusion mechanism are confounded.Source: B. Results and Discussion · p. 4
Band-subset matches full-band for temporal pre+post models
- For temporal pre+post models, band-subset versus full-band F1 is 68.24 vs 66.65 (U-Net), 68.24 vs 67.64 (SegFormer), and 59.80 vs 59.98 (ConvNeXt), all within one standard deviation.The authors hypothesize that the five-band subset retains broadband surface-reflectance sampling while discarded bands add correlated dimensions without proportional burn-discriminative power.Source: B. Results and Discussion · pp. 4–5
Temporal encoding requires a pre-fire reference to be effective
- 67.64±0.91 F1 (SegFormer, full-band) and 68.24 F1 (band-subset) are the best overall results, with pre/post ConvLSTM variants outperforming their architecture’s other settings.Improvements over the strongest spatial baseline, bi-temporal SegFormer, are +2.14 and +2.51 F1 points, respectively.Source: B. Results and Discussion · p. 4, Table I · p. 5
Qualitative analysis
- Figure 2 shows both models tracking a growing scar across acquisitions, but predictions for a small coastal burn with limited valid-pixel coverage are scattered and fail to consistently overlap the scar.The authors suggest that scars spanning only a handful of 300 m pixels offer too little spectral contrast.Source: Figure 2 · p. 5, B. Results and Discussion · pp. 5–6
Architecture comparison
- U-Net and SegFormer are comparable under band-subset temporal pre+post conditioning, while SegFormer leads U-Net by 0.99 F1 under full-band conditioning, within one standard deviation.ConvNeXt-UPerNet remains the weakest architecture in the reported comparisons.Source: B. Results and Discussion · p. 5
VI. CONCLUSIONS
TMB-S3 supports evaluating paired Sentinel-3 OLCI acquisitions, and the conclusions find temporal modeling beneficial when pre- and post-fire imagery are available, with several scope limitations.
VI. CONCLUSIONS
- ConvLSTM modeling yields consistent gains on matched pre- and post-fire acquisitions, but no consistent benefit with post-fire-only sequences.Source: §VI · p. 6
- A bi-temporal early-fusion baseline also benefits spatial models from a pre-fire reference, while the full post-fire sequence brings further consistent improvement.Source: §VI · p. 6
- TMB-S3 comprises 246 European wildfire events with temporally paired Sentinel-3 OLCI acquisitions.Source: §VI · p. 6
- The study is limited by possible spatial leakage across splits, European geographic extent, no fire-free-scene evaluation, and no attention-based temporal encoders.Source: §VI · p. 6