Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ASCNet: Research on all-sky camera images classification at the Muztagh-ata site

Siqi Wang, Qi Fan, Wenbo Gu, Haozhi Wang, AYZADA Jumahali, Lixian Shen, Daiping Zhang, Liyong Liu, Ali Esamdin

TL;DR

This work targets automated classification of nighttime all-sky camera images to assess cloud coverage at the Muztagh-ata site. It introduces ASCNet, a two-branch architecture that fuses RGB semantic features from a ResNet34 backbone with fine-grained luminance textures from an ASCModule built on depthwise dilated convolutions and SE attention, integrated through a FusionBlock with Efficient Channel Attention. The model employs a staged backbone freezing strategy and Focal Loss to handle class imbalance across five cloud-coverage categories, achieving a peak accuracy of $92.66\%$, precision $83.26\%$, recall $84.25\%$, and F1 $83.67\%$, with a consistency rate of $92.7\%$ against manual labeling on data from 2022–2024. Ablation and model-comparison experiments demonstrate ASCNet's superior discriminative power and robustness, suggesting its practical utility for reducing manual labeling workload and enabling scalable astronomical image classification at multiple sites.

Abstract

Cloud coverage is one of the crucial elements of site testing in astronomy. All-sky camera (ASC) images are beneficial for our research on cloud coverage. In this paper, we propose ASCNet, an innovative model specifically designed for classifying nighttime ASC images collected at the Muztagh-ata site from 2022 March to 2024 June. ASCNet integrates ResNet34 with an ASCModule, which employs Depthwise Dilated Convolution and embeds lightweight Squeeze-and-Excitation attention within its branches to extract fine-grained texture information from the luminance channel. The data set is partitioned by category, with 70% of images assigned to the training set and 30% to the test set. The model's performance is assessed by comparing its predictions on the test set with manually annotated labels, yielding a consistency rate of 92.7%. All evaluation metrics of ASCNet are as follows: Accuracy 92.66%, Precision 83.26%, Recall 84.25%, and F1-Score 83.67%, and both ablation and comparative experiments demonstrate significant superiority over other models. A confusion matrix is utilized to analyze the differences between manual classification and model classification. The statistical results demonstrate the model's excellent classification performance and its robust generalization ability, illustrating that ASCNet has potential for application in future astronomical image classifications.

ASCNet: Research on all-sky camera images classification at the Muztagh-ata site

TL;DR

This work targets automated classification of nighttime all-sky camera images to assess cloud coverage at the Muztagh-ata site. It introduces ASCNet, a two-branch architecture that fuses RGB semantic features from a ResNet34 backbone with fine-grained luminance textures from an ASCModule built on depthwise dilated convolutions and SE attention, integrated through a FusionBlock with Efficient Channel Attention. The model employs a staged backbone freezing strategy and Focal Loss to handle class imbalance across five cloud-coverage categories, achieving a peak accuracy of , precision , recall , and F1 , with a consistency rate of against manual labeling on data from 2022–2024. Ablation and model-comparison experiments demonstrate ASCNet's superior discriminative power and robustness, suggesting its practical utility for reducing manual labeling workload and enabling scalable astronomical image classification at multiple sites.

Abstract

Cloud coverage is one of the crucial elements of site testing in astronomy. All-sky camera (ASC) images are beneficial for our research on cloud coverage. In this paper, we propose ASCNet, an innovative model specifically designed for classifying nighttime ASC images collected at the Muztagh-ata site from 2022 March to 2024 June. ASCNet integrates ResNet34 with an ASCModule, which employs Depthwise Dilated Convolution and embeds lightweight Squeeze-and-Excitation attention within its branches to extract fine-grained texture information from the luminance channel. The data set is partitioned by category, with 70% of images assigned to the training set and 30% to the test set. The model's performance is assessed by comparing its predictions on the test set with manually annotated labels, yielding a consistency rate of 92.7%. All evaluation metrics of ASCNet are as follows: Accuracy 92.66%, Precision 83.26%, Recall 84.25%, and F1-Score 83.67%, and both ablation and comparative experiments demonstrate significant superiority over other models. A confusion matrix is utilized to analyze the differences between manual classification and model classification. The statistical results demonstrate the model's excellent classification performance and its robust generalization ability, illustrating that ASCNet has potential for application in future astronomical image classifications.
Paper Structure (17 sections, 17 equations, 7 figures, 6 tables)

This paper contains 17 sections, 17 equations, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Datasets and Classification Criteria
  3. Image Preprocessing
  4. All-sky Image Classification Criteria
  5. Structure of ASCNet
  6. ResNetBackbone
  7. ASCModule
  8. FusionBlock
  9. Experiment Settings
  10. Pretraining
  11. Evaluation Metrics
  12. Loss Function
  13. Classification Results and Analysis
  14. Ablation Experiment
  15. Comparison Of Different Models
  16. ...and 2 more sections

Figures (7)

  • Figure 1: (a) Original image; (b) The pre-processed image.
  • Figure 2: Five categories of ASC images, (a) Clear; (b) Outer; (c) Inner; (d) Covered; (e) None.
  • Figure 3: Overall structure of ASCNet.
  • Figure 4: The structure of ASCModule.
  • Figure 5: The comparison between the manual labels and the model labels from the test set. In the bar chart, the horizontal coordinate is the five categories of ASC images and the vertical coordinate is the number of occurrences of the categories. Green represents consistency, and red represents inconsistency.
  • ...and 2 more figures