Gupta et al. (2026) Optimizing surveillance efficiency with deep learning-driven flood segmentation

Identification

• Journal: Natural Hazards
• Year: 2026
• Date: 2026-02-26
• Authors: Shaurya Gupta, Vinay Dubey, Rahul Katarya
• DOI: 10.1007/s11069-025-07941-6

Research Groups

• Big Data Analytics and Web Intelligence Laboratory, Delhi Technological University, New Delhi, India
• School of Engineering & Technology, Vivekananda Institute of Professional Studies-Technical Campus, New Delhi, Delhi, India

Short Summary

This paper introduces Flood-X, a novel deep learning architecture for pixel-level flood segmentation in ground-level surveillance images, utilizing an Xception-based encoder and a custom lightweight decoder. The model achieves state-of-the-art performance with a mean Intersection over Union (mIoU) of 94% on a combined augmented dataset, outperforming existing methods and demonstrating superior efficiency and accuracy.

Objective

• To develop and evaluate deep learning-based semantic segmentation architectures for accurate pixel-level flood detection in ground-level surveillance images.
• To introduce a novel framework, Flood-X, and compare its performance against existing approaches and a coarse-based network.
• To analyze the effect of different loss functions (Binary Cross-Entropy, Tversky Loss, Log-Cosh Dice Loss) on model convergence and segmentation quality for imbalanced flood datasets.

Study Configuration

• Spatial Scale: Ground-level surveillance images, focusing on local and regional scenes. Input images are resized to 256 pixels × 256 pixels.
• Temporal Scale: Focused on rapid assessment and disaster response for current flood events, enabling real-time deployment for monitoring.

Methodology and Data

• Models used:
  • Flood-X: Novel architecture comprising an Xception-based encoder and a custom lightweight decoder with multi-scale concatenation connections and skip connections.
  • Coarse-based network: An adapted coarse-to-fine network originally for depth estimation, repurposed for binary segmentation.
  • Comparison models: U-Net, SegNet, Mask R-CNN, DeepLabv3+, FASegNet, MoSWIN, ParseNet, Gated-SCNN, Tiramisu, ContextNet, PSPNet.
• Data sources:
  • Combined dataset from Lopez-Fuentes et al. (2017) and Sazara et al. (2019), consisting of 553 original ground-level images of river and flood scenes.
  • Data augmentation performed using the Albumentations library (e.g., flipping, shifting, scaling, rotation, random crop, brightness adjustments, grid distortion, elastic transformation) to expand the dataset to over 2.5 thousand samples.
  • Dataset split: 85% for training (2350 images) and 15% for validation (415 images).
  • Final testing conducted on unseen real-world images sourced from the internet.

Main Results

• Flood-X achieved state-of-the-art performance in ground-level flood segmentation, attaining an mIoU of 94% on the combined augmented dataset.
• Flood-X demonstrated significantly faster convergence, reaching a 90% IOU score in approximately 10 epochs, with minimal overfitting.
• Tversky Loss was identified as the optimal loss function for Flood-X, yielding the highest accuracy and effectively addressing data imbalance, closely followed by Log-Cosh Dice Loss.
• Binary Cross-Entropy loss consistently underperformed, resulting in lower probability predictions near flood boundaries.
• An ablation study confirmed the critical contributions of Flood-X's architectural components: multi-scale concatenation connections (2.75% mIoU drop without them), the Xception encoder (superior efficiency and accuracy compared to standard convolutions), and decoder skip additions (~1.8% mIoU decline without them).
• The lightweight decoder design was validated, as a heavier decoder did not provide significant accuracy gains despite increased computational parameters.

Contributions

• A novel flood segmentation framework (Flood-X) leveraging an Xception-based encoder and a custom decoder optimized for ground-surveillance imagery.
• Comprehensive comparison of Flood-X with a coarse-based network and other state-of-the-art segmentation techniques, demonstrating superior performance.
• Detailed analysis of three loss functions (Binary Cross-Entropy, Tversky Loss, and Log-Cosh Dice Loss) to study their impact on convergence stability and segmentation accuracy for imbalanced flood datasets.
• Extensive quantitative and qualitative evaluations, including results on unseen real-world images, demonstrating the robustness and generalization capability of the proposed Flood-X model for local flood monitoring.

Funding

No funding available.

Citation

@article{Gupta2026Optimizing,
  author = {Gupta, Shaurya and Dubey, Vinay and Katarya, Rahul},
  title = {Optimizing surveillance efficiency with deep learning-driven flood segmentation},
  journal = {Natural Hazards},
  year = {2026},
  doi = {10.1007/s11069-025-07941-6},
  url = {https://doi.org/10.1007/s11069-025-07941-6}
}