Scholarlink

<p>Accurate atmospheric visibility estimation is critical for meteorological monitoring, air quality assessment, and transportation safety. However, the lack of public datasets presents a significant challenge for Deep Learning (DL) research. To address this gap, this paper introduces VisHKO, a new dataset comprising approximately 11,000 weather images categorized into five classes according to the visibility levels. The images were automatically acquired from the Hong Kong Observatory (HKO) website at hourly intervals between 6:00 a.m. and 6:00 p.m. and annotated using ground truth labels derived from official visibility curves. We provide a comprehensive description of the image acquisition and annotation protocols and benchmark the dataset using various DL architectures, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Notably, ViT achieved higher accuracy compared to other DL architectures. The outcomes of all evaluations were then thoroughly discussed. The dataset and its detailed acquisition and annotation processes are publicly accessible, serving as a valuable benchmark for researchers in the field of visibility estimation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

VisHKO: A new weather image dataset for deep learning-based atmospheric visibility estimation

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Kabira Ait Ouadil,
  • Soufiane Idbraim,
  • Nidhal Carla Bouaynaya,
  • Giuseppina Carannante

摘要

Accurate atmospheric visibility estimation is critical for meteorological monitoring, air quality assessment, and transportation safety. However, the lack of public datasets presents a significant challenge for Deep Learning (DL) research. To address this gap, this paper introduces VisHKO, a new dataset comprising approximately 11,000 weather images categorized into five classes according to the visibility levels. The images were automatically acquired from the Hong Kong Observatory (HKO) website at hourly intervals between 6:00 a.m. and 6:00 p.m. and annotated using ground truth labels derived from official visibility curves. We provide a comprehensive description of the image acquisition and annotation protocols and benchmark the dataset using various DL architectures, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Notably, ViT achieved higher accuracy compared to other DL architectures. The outcomes of all evaluations were then thoroughly discussed. The dataset and its detailed acquisition and annotation processes are publicly accessible, serving as a valuable benchmark for researchers in the field of visibility estimation.