Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CP2M: Clustered-Patch-Mixed Mosaic Augmentation for Aerial Image Segmentation

Yijie Li, Hewei Wang, Jinfeng Xu, Zixiao Ma, Puzhen Wu, Shaofan Wang, Soumyabrata Dev

TL;DR

CP2M tackles overfitting in remote sensing segmentation with limited annotations by introducing a two-phase augmentation: Enhanced Mosaic Augmentation and Clustered Patch Mixed Augmentation. The method uses Connected Component Labeling to extract distinct class instances as patches and pastes them into images with corresponding label updates, controlled by a mosaic probability $p_{mosaic}$. On the ISPRS Potsdam dataset, CP2M yields higher accuracy and mIoU than both the baseline and Mosaic alone, demonstrating improved generalization and robustness for high-resolution aerial imagery. This augmentation strategy expands data diversity while mitigating labeling costs, offering practical benefits for remote sensing segmentation tasks.

Abstract

Remote sensing image segmentation is pivotal for earth observation, underpinning applications such as environmental monitoring and urban planning. Due to the limited annotation data available in remote sensing images, numerous studies have focused on data augmentation as a means to alleviate overfitting in deep learning networks. However, some existing data augmentation strategies rely on simple transformations that may not sufficiently enhance data diversity or model generalization capabilities. This paper proposes a novel augmentation strategy, Clustered-Patch-Mixed Mosaic (CP2M), designed to address these limitations. CP2M integrates a Mosaic augmentation phase with a clustered patch mix phase. The former stage constructs a new sample from four random samples, while the latter phase uses the connected component labeling algorithm to ensure the augmented data maintains spatial coherence and avoids introducing irrelevant semantics when pasting random patches. Our experiments on the ISPRS Potsdam dataset demonstrate that CP2M substantially mitigates overfitting, setting new benchmarks for segmentation accuracy and model robustness in remote sensing tasks.

CP2M: Clustered-Patch-Mixed Mosaic Augmentation for Aerial Image Segmentation

TL;DR

CP2M tackles overfitting in remote sensing segmentation with limited annotations by introducing a two-phase augmentation: Enhanced Mosaic Augmentation and Clustered Patch Mixed Augmentation. The method uses Connected Component Labeling to extract distinct class instances as patches and pastes them into images with corresponding label updates, controlled by a mosaic probability . On the ISPRS Potsdam dataset, CP2M yields higher accuracy and mIoU than both the baseline and Mosaic alone, demonstrating improved generalization and robustness for high-resolution aerial imagery. This augmentation strategy expands data diversity while mitigating labeling costs, offering practical benefits for remote sensing segmentation tasks.

Abstract

Remote sensing image segmentation is pivotal for earth observation, underpinning applications such as environmental monitoring and urban planning. Due to the limited annotation data available in remote sensing images, numerous studies have focused on data augmentation as a means to alleviate overfitting in deep learning networks. However, some existing data augmentation strategies rely on simple transformations that may not sufficiently enhance data diversity or model generalization capabilities. This paper proposes a novel augmentation strategy, Clustered-Patch-Mixed Mosaic (CP2M), designed to address these limitations. CP2M integrates a Mosaic augmentation phase with a clustered patch mix phase. The former stage constructs a new sample from four random samples, while the latter phase uses the connected component labeling algorithm to ensure the augmented data maintains spatial coherence and avoids introducing irrelevant semantics when pasting random patches. Our experiments on the ISPRS Potsdam dataset demonstrate that CP2M substantially mitigates overfitting, setting new benchmarks for segmentation accuracy and model robustness in remote sensing tasks.
Paper Structure (13 sections, 3 equations, 5 figures, 1 table)

This paper contains 13 sections, 3 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. CP2M Pipeline
  4. Enhanced Mosaic Augmentation
  5. Clustered Patch Mixed Augmentation
  6. Experiments & Results
  7. Experiments Settings
  8. Model Architecture and Objective Function
  9. Metrics
  10. Qualitative Evaluation
  11. Quantitative Evaluation
  12. Visualization of CP2M Generated Samples
  13. Conclusion

Figures (5)

  • Figure 1: Overall pipeline of Clustered-Patch-Mixed Mosaic (CP2M)
  • Figure 2: MobileNetV2-UNet Model Architecture
  • Figure 3: Qualitative comparison between baseline and CP2M. class mapping: impervious surfaces (C1), buildings (C2), low vegetation (C3), trees (C4), cars (C5), clutter/background (C6).
  • Figure 4: Impact of the probability of using Mosaic on model performance (Accuracy and mIoU). We use $\mathrm{p\_mosaic}$ as a threshold to control the proportion of normal and Mosaic samples in training. $\mathrm{p\_mosaic}$ is 0 means Mosaic is not used, $\mathrm{p\_mosaic}$ is 1 means Mosaic is used for every sample.
  • Figure 5: CP2M Augmentation Process: (A) Original Mosaic Images, (B) Mosaic Labels, (C) Patch Images, (D) Patch Labels, (E) Connected Component Labeling Results with Color-coded Instances, (F) Binary Masks Selection, (G) Augmented Output Images, (H) Final Augmented Output Labels.