Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation

Yijie Li, Hewei Wang, Shaofan Wang, Yee Hui Lee, Muhammad Salman Pathan, Soumyabrata Dev

TL;DR

The paper tackles cloud segmentation from ground-based sky images and introduces UCloudNet, a residual U-Net enhanced with deep supervision to achieve higher accuracy with reduced training time. By embedding residual connections in the encoder and employing auxiliary loss branches, the model improves feature extraction and accelerates convergence. Experiments on the SWINySEG dataset show that UCloudNet with $k=4$, deep supervision, and learning-rate decay attains state-of-the-art performance across multiple metrics, while requiring significantly fewer iterations. This work advances practical, real-time cloud segmentation by balancing accuracy and training efficiency, with potential extensions to multi-class segmentation and cloud depth estimation.

Abstract

Recent advancements in meteorology involve the use of ground-based sky cameras for cloud observation. Analyzing images from these cameras helps in calculating cloud coverage and understanding atmospheric phenomena. Traditionally, cloud image segmentation relied on conventional computer vision techniques. However, with the advent of deep learning, convolutional neural networks (CNNs) are increasingly applied for this purpose. Despite their effectiveness, CNNs often require many epochs to converge, posing challenges for real-time processing in sky camera systems. In this paper, we introduce a residual U-Net with deep supervision for cloud segmentation which provides better accuracy than previous approaches, and with less training consumption. By utilizing residual connection in encoders of UCloudNet, the feature extraction ability is further improved.

UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation

TL;DR

The paper tackles cloud segmentation from ground-based sky images and introduces UCloudNet, a residual U-Net enhanced with deep supervision to achieve higher accuracy with reduced training time. By embedding residual connections in the encoder and employing auxiliary loss branches, the model improves feature extraction and accelerates convergence. Experiments on the SWINySEG dataset show that UCloudNet with , deep supervision, and learning-rate decay attains state-of-the-art performance across multiple metrics, while requiring significantly fewer iterations. This work advances practical, real-time cloud segmentation by balancing accuracy and training efficiency, with potential extensions to multi-class segmentation and cloud depth estimation.

Abstract

Recent advancements in meteorology involve the use of ground-based sky cameras for cloud observation. Analyzing images from these cameras helps in calculating cloud coverage and understanding atmospheric phenomena. Traditionally, cloud image segmentation relied on conventional computer vision techniques. However, with the advent of deep learning, convolutional neural networks (CNNs) are increasingly applied for this purpose. Despite their effectiveness, CNNs often require many epochs to converge, posing challenges for real-time processing in sky camera systems. In this paper, we introduce a residual U-Net with deep supervision for cloud segmentation which provides better accuracy than previous approaches, and with less training consumption. By utilizing residual connection in encoders of UCloudNet, the feature extraction ability is further improved.
Paper Structure (13 sections, 2 equations, 5 figures, 1 table)

This paper contains 13 sections, 2 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Architecture
  4. Double Convolution Block (DCB)
  5. Down Sample Block and Up Sample Block
  6. Deep supervision
  7. Parameters control and loss function
  8. Experiments & Results
  9. Dataset and Training Configurations
  10. Metrics
  11. Quantitative Analysis
  12. Qualitative Analysis
  13. Conclusion

Figures (5)

  • Figure 1: The architecture of the UCloudNet model. The procedure between the output of model and the segmentation mask has been omitted in this figure.
  • Figure 2: The structure of ‘Double Convolution Block’ in encoder, decoder, ‘Down Sample Block’, and ‘Up Sample Block’ (from left to right).
  • Figure 3: Results of cloud segmentation for day-time (1-6 columns) and night-time (7-12 columns).
  • Figure 4: PR curve of UCloudNet with different training configurations on full SWINySEG ground-based cloud segmentation dataset.
  • Figure 5: Loss curve of the final output and auxiliary outputs.