A major barrier in ground-based cloud image segmentation is the lack of high-quality, large-scale, and clear-label public datasets. The performance of deep learning models fundamentally depends on the diversity and precision of the training data. However, existing benchmarks primarily focus on binary classification of clouds versus sky background, neglecting important characteristics such as cloud scale, morphology, and color features.
To address this gap and provide the research community with a more challenging and practical benchmark, we developed and publicly released a new dataset incorporating Complex-Scale variations as well as Radiative and Color attributes (CSRC), which constitutes a core contribution of this paper. The solar radiation effect on clouds is a critical factor in the accuracy of photovoltaic power forecasting. International cloud classification standards are primarily based on the brightness, color, size, and altitude of clouds. Among these, color variations induced by radiative sources play a significant role in the precision of pixel-level classification in ground-based cloud imagery. In this study, we focus particularly on cloud scale and color attributes, and introduce fine-grained segmentation that incorporates radiative influence, re-categorizing clouds into three classes. White and gray clouds are common color patterns in the sky, often covering extensive areas, and are represented by white and gray labels, respectively. Pixels near the sun often appear nearly pure white due to high radiation, with red and blue channel values converging, leading these regions to be classified as overcast. Therefore, it is essential to incorporate fine-grained solar radiation segmentation, which is indicated in red. Black cloud systems, typically associated with heavy rain or hail, result in full sky coverage and are detrimental to both ground-based cloud image segmentation and photovoltaic power prediction tasks. Such black clouds are ignored from our dataset. Cloud-free sky regions are labeled in blue as background.
The images in our CSRC dataset were collected at a meteorological observation station located in Xiqing District, Tianjin, China (geographic coordinates: 117.03° E, 39.10° N). The core acquisition device is an All-Sky Imager (ASI-DC-TK02), which provides a wide-angle field of view exceeding 180° and is housed in a waterproof and dustproof enclosure. The system was configured to automatically capture images at fixed 30-second intervals. All images are stored in RGB color JPG format with a resolution of 1260 × 1260 pixels. Representative samples from the CSRC dataset are shown in Fig.1
在下载数据集前，你应保证已阅读并理解 CSRC 协议中包含的所有内容。下载并将 CSRC 用于你的研究，即表示你同意 CSRC 协议中的所有限制与要求 。
完整数据集云盘链接：
云盘提取密码请联系邮箱：qingxinqazxsw@163.com