PaperlayerView exact revision on arXiv

Source-linked AI summary

Scale-based Approach for Active Wildfire Segmentation on Satellite Imagery

Matheus F. Kovaleski, Cristiano Premebida, João Ruivo Paulo

arXiv:2609.01392v1cs.CV

TL;DR

Active-fire segmentation must handle sparse, imbalanced pixels and changing fire scales, yet standard random splits do not explicitly test scale-shift robustness. The paper introduces an IQR-based protocol using Landsat-8 imagery to compare architectures and SWIR configurations; U-Net generalizes best, SegFormer remains competitive, and SWIR2 is consistently strongest or near-best.

  • Problem

    Wildfire segmentation faces severe fire-background imbalance and changing fire morphology and scale, while random splits do not explicitly evaluate robustness under fire-scale distribution shifts.

  • Method

    The study uses connected-component fire-size analysis and an IQR-derived partition to evaluate U-Net, DeepLabV3+, and SegFormer with different SWIR-based configurations.

  • Results

    U-Net achieves the strongest overall generalization, SegFormer shows competitive robustness, DeepLabV3+ has substantially lower recall, and SWIR2-only consistently provides the strongest or near-best results.

  • Takeaways & Limitations

    Architectural design and spectral configuration both strongly influence robustness in active-fire segmentation under fire-scale distribution shifts.

Abstract

from arXiv · show

Active wildfire mapping from satellite imagery is challenging due to the sparse and highly imbalanced nature of fire pixels, especially in early-stage or low-density fire observations. This work investigates the use of multispectral Landsat-8 imagery for active-fire segmentation under multi-scale wildfire size conditions. We propose a data-driven protocol to characterize fire-region size distributions through connected-component analysis and an interquartile range criterion, enabling the evaluation of model robustness across different local fire-region densities. Three segmentation architectures, U-Net, DeepLabV3+, and SegFormer, are evaluated under different SWIR-based spectral configurations. Results show that U-Net achieves the strongest robustness across the evaluated conditions, SegFormer provides competitive performance, and DeepLabV3+ tends to produce conservative predictions with reduced recall. Across architectures, SWIR2 consistently achieves the strongest or near-best results, highlighting its importance for active-fire segmentation in Landsat-8 imagery. These findings suggest that both spectral band selection and architectural design are critical for robust satellite-based active wildfire mapping trained on low active fire-pixel density images.

I. INTRODUCTION

Wildfire segmentation is difficult because fire regions are sparse, imbalanced, and highly variable in morphology and scale, while conventional random splits may miss robustness under scale shifts. This work evaluates a scale-based protocol, SWIR configurations, and three segmentation architectures.

  • Wildfire segmentation faces severe class imbalance, spectral ambiguity, and substantial variation in fire morphology and spatial extent.
  • Conventional random train-test splits may not adequately assess robustness to distribution shifts caused by fire-scale variability.
  • A data-driven IQR strategy partitions connected fire regions to evaluate generalization from small-scale training regions to substantially larger active-fire regions.
  • The evaluated segmentation architectures are U-Net, DeepLabV3+, and SegFormer.
  • The study analyzes SWIR spectral configurations and compares CNN-based with transformer-based architectures under shifted fire-scale distributions.

II. RELATED WORK

Prior wildfire segmentation research has advanced through multispectral sensing and deep learning, but generalization across fire scales remains insufficiently evaluated. This work addresses that gap with a controlled scale-based evaluation protocol.

  • Traditional wildfire detection methods use thermal and SWIR information but are sensitive to atmospheric effects, noise, and spectral confusion.
  • CNNs learn spatial and spectral patterns directly from data, while U-Net preserves spatial details and DeepLabV3+ captures multi-scale context.
  • Transformer-based models such as SegFormer and FireFormer extend deep-learning approaches for active-fire detection and segmentation.
  • SWIR-inclusive multispectral inputs can improve fire detection, and selected band combinations may outperform using all available spectral channels.
  • Severe fire-background imbalance and variability in fire morphology, spatial distribution, and scale remain central segmentation challenges.
  • The proposed scale-based protocol evaluates model generalization from small-scale to large-scale fire regions instead of relying only on random splits.

III. METHODOLOGY

The methodology uses raw Land8Fire Landsat-8 data and a pipeline combining preprocessing, fire-region analysis, and scale-based partitioning. The resulting splits support controlled evaluation of model generalization.

  • The proposed framework comprises preprocessing and fire-region extraction, statistical fire-size analysis, and scale-based dataset partitioning.
  • Land8Fire is a large-scale, high-resolution multispectral benchmark for semantic segmentation of active wildfires.
  • The dataset derives from Landsat-8 imagery and includes manually validated masks, improved sampling, and wildfire events from multiple world regions.
  • The study uses the raw dataset so patch extraction and scale-based splitting remain under full pipeline control.

B. Data Preprocessing and Scale-Based Splitting Pipeline

The preprocessing pipeline aligns imagery and masks, extracts patches, measures connected fire regions, and partitions samples by an IQR-derived fire-size threshold. This creates a controlled shift from small to large fire structures for evaluation.

  • Landsat-8 images are georeferenced to align them spatially with their ground-truth masks before further processing.
  • Images are divided into non-overlapping 256×256-pixel patches, and connected components in binary fire masks are measured by fire-pixel count.
  • The IQR criterion produces a 16-pixel cutoff, equivalent to approximately 1.44 hectares at 30×30 meters per Landsat-8 pixel.
  • Small-scale and large-scale describe connected fire-region sizes within individual patches rather than operational wildfire size classes.
  • Figure 1 traces preprocessing, connected-component analysis, IQR-based fire-size characterization, partitioning, and evaluation across spectral configurations and models.
  • Patches with largest connected components ≤16 pixels enter training, whereas patches exceeding 16 pixels enter testing.
  • The resulting protocol evaluates model generalization from small to large fire structures under a controlled fire-scale distribution shift.
  • After excluding patches without fire pixels, the dataset contains approximately 12,000 training patches and 400 testing patches.

C. Spectral Band Configurations

The study compares three SWIR-based spectral configurations to assess their contribution to active-fire segmentation and generalization across fire scales. It focuses on consistent 30 m OLI bands while excluding lower-resolution thermal bands.

  • Three configurations compare dual-SWIR, SWIR2-only, and SWIR1-only inputs while retaining Red and NIR bands.The comparison tests individual SWIR contributions, their complementarity, and the effect of spectral richness on scale-shift generalization.
  • SWIR2 targets longer-wavelength sensitivity, whereas SWIR1 provides complementary thermal and reflectance information.Both bands are responsive to active fire, with SWIR2 described as more sensitive to higher temperature ranges.
  • The scale-based examples classify patches by whether any connected fire region exceeds the IQR-derived threshold of 16 pixels.Small-fire rows remain below the threshold, while large-fire rows contain at least one connected region above it.
  • All configurations preserve Red and NIR context while excluding B10 and B11 because their native 100 m resolution requires resampling to 30 m.The design isolates SWIR effects using 30 m OLI bands and avoids potential patch-level spatial uncertainty from resampling.

D. Segmentation Models

The study evaluates U-Net, DeepLabV3+, and SegFormer as complementary architectures differing in spatial-detail preservation, contextual aggregation, and global representation learning. U-Net uses skip connections to retain fine-grained spatial information for sparse, small fire regions.

  • Three architectures—U-Net, DeepLabV3+, and SegFormer—were selected for complementary spatial, contextual, and global representation capabilities.The models support comparison of architectural design trade-offs under wildfire segmentation conditions.
  • U-Net uses an encoder-decoder design with skip connections to preserve fine-grained spatial information during reconstruction.Its localization capability is suited to detecting sparse and small fire regions.

2) DeepLabV3+:

DeepLabV3+ captures multi-scale contextual information through atrous convolutions and Atrous Spatial Pyramid Pooling, while the study compares it with U-Net and SegFormer under multispectral inputs. Its backbone was adapted from three-channel RGB input to four-channel data while preserving pretrained information where possible.

  • 2) DeepLabV3+:: DeepLabV3+ uses atrous convolutions and Atrous Spatial Pyramid Pooling to capture multi-scale contextual information.Its large receptive fields support representation of spatially complex fire structures.
  • 2) DeepLabV3+:: The comparison frames DeepLabV3+ around contextual multi-scale aggregation alongside U-Net’s local detail preservation and SegFormer’s global attention-based learning.These design trade-offs are evaluated for generalization from small to large fire events.
  • 2) DeepLabV3+:: DeepLabV3+ was modified to accept four-channel multispectral inputs by changing its first convolutional projection layer.The original three-channel weights were retained, and the added channel was initialized from their mean.
  • 2) DeepLabV3+:: The input adaptation incorporates additional spectral information without significantly altering the original architecture or training dynamics.A similar pretrained-preserving extension was applied to SegFormer for comparison.

E. Evaluation Metrics

The evaluation uses standard segmentation metrics and compares conventional random splitting with the proposed scale-based split. Results are averaged across independent runs under shared training settings to assess generalization across fire scales.

  • Performance was measured using Precision, Recall, F1-score, Intersection over Union, and Matthews Correlation Coefficient.These metrics provide multiple views of binary fire-versus-background segmentation quality.
  • The experiments compare random splitting with scale-based splitting to test generalization under a controlled fire-scale distribution shift.The comparison examines whether conventional evaluation protocols reliably estimate cross-scale performance.
  • Models used shared learning rate, batch size, weight decay, and training-iteration settings whenever possible.Reported results are means and standard deviations over 5 independent runs with different random seeds.
  • The reference split allocated 70% of data to training, 15% to validation, and 15% to testing without considering fire size or spatial characteristics.This setup represents a conventional training scenario and excludes scale-based filtering.

C. Scale-Based Setup

The scale-based setup evaluates generalization from small fire regions to substantially larger regions through a controlled distribution shift. It compares this protocol with conventional evaluation across three testing conditions.

  • The scale-based split trains on small fire regions and tests on patches containing large fire regions.
  • All three architectures use the same spectral configurations and preprocessing pipeline for consistent comparison.
  • The baseline split represents a conventional training scenario, while the scale-based split introduces a controlled distribution shift.
  • Small-fire testing measures performance on data similar to the training distribution.
  • Large-fire testing measures generalization when fire regions are significantly larger than those seen during training.

F. Results

Across scale-based wildfire segmentation tests, U-Net is the most robust, SegFormer is competitive, and DeepLabV3+ is conservative with lower recall. SWIR2-only and Dual-SWIR configurations generally outperform SWIR1-only.

  • U-Net: U-Net achieves the strongest overall performance, with F1-scores above 92% and IoU values above 86% for Dual-SWIR and SWIR2-only on baseline and large-fire tests.On small-fire tests, precision and IoU decrease while recall remains consistently high.
  • DeepLabV3+: DeepLabV3+ shows substantially lower performance than U-Net, especially in recall, producing conservative predictions that miss many actual fire pixels.Its strongest results use Dual-SWIR or SWIR2-only, while SWIR1-only performs considerably worse.
  • SegFormer: SegFormer achieves intermediate performance, with F1-scores above 85% and IoU values above 74% for Dual-SWIR and SWIR2-only on baseline and large-fire tests.It maintains a more balanced precision-recall trade-off than DeepLabV3+ and remains competitive under shifted fire scales.
  • Cross-Model Analysis: U-Net provides the most robust performance across evaluated conditions, particularly for recall, IoU, F1-score, and MCC.Its encoder-decoder structure with skip connections preserves fine spatial details for sparse, fragmented, and imbalanced fire masks.
  • Cross-Model Analysis: SWIR2-only consistently outperforms SWIR1-only, while Dual-SWIR generally provides the most stable results across architectures.The results identify SWIR2 as especially important for active wildfire mapping under scale-based distribution shifts.

V. CONCLUSION

The conclusion presents a scale-based partitioning strategy for wildfire segmentation and reports that robustness depends on architecture and spectral configuration. U-Net generalizes best, SegFormer remains competitive, DeepLabV3+ has lower recall, and SWIR2-only is consistently strong or near-best.

  • The proposed partitioning strategy derives fire-scale splits from the interquartile range of connected fire-region sizes in the Land8Fire dataset.
  • U-Net achieves the strongest overall generalization, SegFormer shows competitive robustness, and DeepLabV3+ exhibits substantially lower recall.
  • SWIR2-only consistently provides the strongest or near-best results across architectures in Landsat-8 active wildfire segmentation.The scale-based protocol is described as more challenging and informative than conventional random train-test splits.
Loading 2609.01392v1…