Preamble

Vol. 66 No. 5
Sept., 2025
Acta Astronomica Sinica
doi: 10.15940/j.cnki.0001-5245.2025.05.002

Research on Cloud Monitoring Scheme for Radio Telescope Observatory Sites

ZHANG Hai-long†, LU Deng-rong‡, SUN Ji-xian, LI Ji-bin, ZHANG Xu-guo
(Qinghai Station of Purple Mountain Observatory, Chinese Academy of Sciences, Delingha 817000)

Abstract

In the process of selecting sites for millimeter-wave and submillimeter-wave radio astronomical telescopes, it is essential to fully understand cloud cover information at candidate observatory locations. Therefore, designing an all-sky camera system suitable for field environments is necessary. According to the characteristics of radio telescopes and the specific conditions of field sites, this scheme innovatively utilizes planetary cameras and embedded microcontrollers to develop a full-time, all-sky camera system. The system can operate long-term in the field using solar power, with the most important feature being unmanned, autonomous operation. For data processing, the scheme also innovatively employs deep learning neural network algorithms to extract data feature values, establish a machine learning model library, and automatically统计 cloud cover information at the site. This approach is more efficient and simpler than manual or general image processing algorithms. These studies provide important references for more comprehensive evaluation of millimeter and submillimeter-wave radio observatory sites.

Keywords telescopes: radio, instrumentation: detectors, methods: measurement and evaluation, techniques: image processing, site testing

1 Introduction

Millimeter-wave and submillimeter-wave radio astronomical telescopes have stringent requirements for observatory sites, particularly submillimeter-wave telescopes. In addition to demanding conditions such as low temperature, low atmospheric water vapor content, and large-scale radio-quiet zones, these sites also have certain requirements for altitude, topography, atmospheric effects, and cloud cover [1]. Submillimeter-wave radio telescopes require sites with minimal cloud cover because cloud amount indirectly reflects atmospheric water vapor content. Therefore, cloud cover significantly impacts the quality of data produced by millimeter-wave radio telescopes and can even reduce observation efficiency. In monitoring data at observatory sites, using all-sky cameras to capture cloud images is one of the most common methods for monitoring sky cloud cover [2]. To understand cloud cover information at candidate field sites, the system design draws on the all-sky cloud monitoring scheme from China's SONG (Stellar Observations Network Group) project [3] and other optical site all-sky camera development plans [4]. To achieve automatic all-sky monitoring at sites, the scheme employs a ZWO ASI224MC color planetary camera with 1280×960 pixels and a pixel size of 3.75 μm × 3.75 μm, providing extremely low read noise and high sensitivity. This camera is much smaller than conventional DSLR cameras, facilitating integration, and has low power consumption (maximum 2.8 W), allowing direct power supply and control via USB interface. To monitor cloud cover across all sky regions, the camera is paired with a 5-megapixel fisheye lens with a focal length of 1.55 mm. For rapid processing of large data volumes (at least hundreds of cloud images daily), the scheme uses Python-based deep learning algorithms. Through extensive sample training, various cloud image feature values are extracted to establish a model library, enabling fully automatic statistical analysis of site weather conditions.

2.1 Monitoring Requirements

Monitoring at potential astronomical candidate sites has the basic requirement of unmanned operation to obtain long-term data. Particularly when selecting millimeter-wave and submillimeter-wave radio telescope sites, candidate locations are typically in high-altitude, hypoxic, uninhabited areas without any support infrastructure. This requires the monitoring system to be fully automated, including automatic monitoring, automatic camera parameter adjustment based on light conditions, automatic data saving, and automatic data transmission.

The all-sky camera monitoring system must monitor sky cloud cover changes in real time and save data locally. In environments with network connectivity, the system supports remote login to display cloud cover data. Cloud cover data can also be saved locally in image format according to sampling time for subsequent processing. The sampling interval can be modified; based on comprehensive analysis of image size, storage space, and cloud change timescales, the scheme selects a 5-minute interval for saving one cloud image. Due to long-term unmanned operation, high reliability is required, and candidate site environments can reach minimum temperatures of -35 °C, necessitating guaranteed normal operation in low-temperature conditions. The monitoring control system must be integrated, lightweight, and low-power to adapt to field installation, operation, and maintenance.

2.2 System Composition

The all-sky camera monitoring system consists of an embedded microcontroller (Raspberry Pi 4B), an astronomical camera produced by ZWO (ASI120MM-S), a heating control module, a fisheye lens (FS15520FEMP), an acrylic hemisphere, a housing, and a DC power supply (including solar panels, batteries, charging controller, etc.), as shown in Figure 1 [FIGURE:1].

The entire all-sky camera chamber is wrapped in acrylic plates, integrating the embedded microcontroller and planetary camera inside. The embedded controller uses a Raspberry Pi microcontroller with 5 V DC power supply and maximum power consumption of 6–7 W, running a Linux operating system with 128 GB storage space. The planetary camera is powered and communicated via the Raspberry Pi's USB port, with maximum power consumption of 2.86 W.

Considering camera and lens protection, light transmittance, processing and installation, and overall weight, the all-sky camera housing structure uses acrylic material (organic glass) throughout. Acrylic plates have excellent light transmittance (over 92%) and weather resistance, with strong adaptability to natural environments. Performance remains unchanged after long-term exposure to sunlight, wind, and rain, demonstrating good aging resistance. For processing, acrylic is suitable for both mechanical processing and thermoforming, and cutting is convenient. The acrylic hemisphere and housing are available as ready-made products requiring no complex processing, while other support plates can be cut from acrylic sheets according to installation requirements. For field installation and operation, the structural design principle is to achieve miniaturization, integration, and lightweight. The 3D model of the all-sky camera is shown in Figure 2 [FIGURE:2].

When used in winter, the large temperature difference between inside and outside the acrylic housing can cause frost formation on the inner wall of the hemisphere. Therefore, gaps are left at the bottom contact surface during hemisphere installation, and ventilation holes are opened around the perimeter to facilitate temperature balance and eliminate frost formation on the inner wall, as shown in Figure 3 [FIGURE:3].

To improve thermal insulation, the entire acrylic housing is also wrapped with thermal insulation cotton to reduce heat loss. The actual device is shown in Figure 5 [FIGURE:5].

The camera's minimum operating temperature is -5 °C. Considering that winter temperatures at field sites can reach -35 °C, flexible polyimide (PI) film heaters are installed for the camera, using 12 V DC heating for thermal insulation. Through heater operation and heat generated by the Raspberry Pi controller itself, the camera temperature is maintained above -5 °C. Temperature control uses a simple temperature switch chip installed near the camera. Heating is controlled based on camera temperature: below 0 °C the temperature switch closes and the heater powers on; above 0 °C the temperature switch remains open and the heater powers off, as shown in Figure 4 [FIGURE:4].

3.1 All-Sky Camera Installation and Monitoring

The all-sky camera was initially tested from early December 2023 to mid-February 2024, installed at the Qinghai Observatory Station site of Purple Mountain Observatory, Chinese Academy of Sciences. In March 2024, it was installed at the Delingha Snow Mountain Ranch radio multi-band candidate site of Purple Mountain Observatory for long-term monitoring, as shown in Figure 6 [FIGURE:6].

The camera control program is written in Python, calling the camera's built-in function library to achieve automatic photography at set time intervals and save images. The program can also control camera exposure time and gain parameters in real time based on changes in sky brightness. Common algorithms for camera gain and exposure time control include average brightness method, weighted mean method, and brightness histogram, with the average brightness method being the most common and employed in this scheme. The specific control process is shown in Figure 7 [FIGURE:7].

The control program calculates the mean brightness of each captured image using the median function from Python's numpy library. Based on measured results, a brightness range is set: values below the minimum are considered too dark, requiring increased exposure time and gain, while values above the maximum are considered too bright, requiring decreased gain and exposure time. The program then sets new gain and exposure parameters, acquires the adjusted image, and recalculates mean brightness for iterative adjustment until the image's average brightness falls within the set range.

The brightness range must be determined based on actual site conditions, which can be assisted by long-term measurements using sky brightness sensors, night sky monitors, and other instruments to calculate setting parameters. In this scheme, based on specific conditions such as presence or absence of sun and moon, clear skies, cloudy conditions, and various weather scenarios, optimal empirical brightness ranges were obtained through multiple parameter adjustments. Figure 8 shows cloud images captured during daytime with sun, clear sky, cloudy, and overcast conditions. Figure 9 shows cloud images captured at night with moon, clear sky, cloudy, and overcast conditions.

During testing, it was found that direct sunlight reflects the camera's bottom image onto the top of the hemisphere, as shown in the top-left photo of Figure 8. Later, black anti-reflection cloth was considered for placement at the bottom of the camera lens to eliminate this reflection.

In traditional cloud image processing, the grayscale values of sun and moon regions are sometimes very similar to cloud features, causing data processing algorithms to easily misidentify sun and moon influence areas as clouds. Therefore, image data requires preprocessing. Solutions include time segmentation method and difference method [5]. The basic idea is that in adjacent photos, the position changes of the sun and moon are smaller than cloud changes, so the difference method can remove them. The time segmentation method identifies and removes sun and moon trajectories by recognizing that their grayscale values change little over multiple consecutive frames. These two methods can effectively eliminate sun and moon influence but complicate data processing. For applications not requiring accurate analysis of cloud size and coverage, deep learning neural network algorithms offer greater advantages, as discussed below.

3.2 Data Processing and Statistical Analysis

All-sky camera data is saved in image format, and the data volume is extremely large, requiring efficient algorithms for processing and statistics. Cloud image analysis has mostly relied on manual statistics [5], with some algorithmic processing. However, algorithm modeling generally requires professional knowledge of image processing, and image processing algorithms are typically cumbersome, resulting in unsatisfactory data processing efficiency. Therefore, this scheme innovatively employs image recognition technology from the field of artificial intelligence. The basic approach involves first extracting image features, then establishing a model library through machine learning, and finally classifying and统计 target images based on this foundation. The traditional image recognition process is shown in Figure 10 [FIGURE:10].

With the development of AI technology, image recognition has become an important field of artificial intelligence, with increasingly stringent requirements that reveal many shortcomings of traditional image recognition, such as weak adaptability, susceptibility to noise interference, and dependence on manual design [6]. Particularly in complex application environments like autonomous driving, computers must quickly identify various objects. Deep learning is well-suited for such scenarios, as extensive sample training and long-duration training significantly improve image recognition accuracy and efficiency.

Deep learning research can be traced back to 1943, when American psychologist McCulloch and mathematician Pitts proposed the M-P neuron model while exploring artificial neural networks [7]. Deep learning mainly includes the following algorithms: Deep Belief Network (DBN) [8-10], Recurrent Neural Network (RNN) [11], and Convolutional Neural Network (CNN) [12-15]. A typical neural network structure is shown in Figure 11 [FIGURE:11], including an input layer, several hidden layers, and an output layer. Hidden layers contain numerous neurons with corresponding weight connections and activation functions. The number of neurons and hidden layers indicates network complexity—more numerous networks have stronger adaptability and more significant nonlinear effects.

The CNN structure is similar to traditional neural networks, also designed based on brain neurons. Its essence is connecting simple neurons, using each low-level neuron to extract image information, then combining many such neurons to extract very complex image information, ultimately obtaining high-level semantic features of images [6]. According to relevant experimental results, training models designed with CNN can achieve 96% accuracy on datasets [6], providing fundamental support for better development of image recognition technology.

The data processing flow is shown in Figure 12 [FIGURE:12]. In this scheme, CNN is first used to generate a machine learning model library, dividing numerous sample cloud images into four groups (represented by numbers: 0: overcast; 1: cloudy; 2: clear; 3: snow). The data processing program converts all cloud images to grayscale, obtains pixel matrices, and uses deep learning CNN tools for feature extraction to obtain feature matrices W for the four sample groups of overcast, clear, cloudy, and snow conditions. After extensive iterative training with large samples, the training model parameter library is finally obtained. The cloud image recognition process also converts cloud images to grayscale, then calls the training model library, extracts cloud image feature values through CNN, and finally classifies them into the predefined four groups based on feature values to determine weather conditions.

CNN feature extraction is a crucial component in both training model library generation and cloud image recognition, forming the core of image recognition. Its basic structure is shown in Figure 13 [FIGURE:13].

As shown in Figure 13, the CNN image feature extraction employs multiple classic network layers: Convolution, Max pooling, ReLU (Rectified Linear Unit) activation, and Fully connected layers. The convolutional layer functions to extract image features and is the core of CNN. Pixel matrix data of images undergoes convolution operations with convolutional kernels containing weight parameters. The convolution operation proceeds left to right, top to bottom in the pixel matrix according to formula (1) to obtain the image feature matrix, as shown in Figure 14 [FIGURE:14].

In formula (1), the leftmost matrix is the image pixel matrix, denoted as a. The weight values of the convolutional kernel are the final results to be obtained through image recognition, namely the model library. Initial values are random, and after training iterations, the machine learning model library is finally obtained.

The pooling layer's main function is to select features extracted by convolutional layers, choosing image information not disturbed by position. Second, it reduces feature dimensions, decreasing the number of feature variables and thereby reducing computation. Common pooling operations include max pooling and mean pooling. In Figure 15 [FIGURE:15], the extracted feature matrix is divided according to bold-lined boxes, with the maximum value extracted from each region to form a new feature matrix—this is the max pooling operation.

In neurons, input Input undergoes a series of weighted summations and then acts on another function called the activation function. Similar to neuron-based models in the human brain, the activation function ultimately determines whether to transmit signals and what content to transmit to the next neuron. Common activation functions include Sigmoid, Tanh, and ReLU. The first two are prone to overfitting during operation, so the ReLU function is generally used, with the form:

The fully connected layer serves a classification function in the neural network model and is the final layer. During training model library generation, training efficiency can be improved by optimizing parameters such as the number of CNN layers, number of convolutional kernels, whether to perform padding operations, the Dropout random deactivation ratio for neural network model regularization, and the number of data traversal epochs. Among these, the number of convolutional layers and traversal epochs have greater impact.

Figure 16 [FIGURE:16] shows the influence of neural network convolutional layer numbers on training effect, where Train_loss and Train_acc represent neural network loss function values and accuracy, respectively. It can be seen that with the same sample data volume, more convolutional layers increase the risk of loss function jumps and overfitting. Figure 17 [FIGURE:17] shows training effects when Epochs are set to 10, 30, 60, and 100 iterations. Accuracy reaches over 90% when data traverses 30 times. Excessive traversal makes loss function values non-smooth, risks overfitting, and increases computer processing time.

From March 4 to 17, 2024, 4089 data groups were measured at the candidate site, and weather statistics were obtained through the training model. Figures 18 [FIGURE:18] and 19 [FIGURE:19] show model training and weather data statistics, respectively.

After obtaining the training model, to evaluate its quality, four groups of real sample sets (gold standard) were extracted from the sample data, including overcast, cloudy, clear, and snow conditions. These real sets were then input into the model for analysis. The open-source artificial neural network library keras written in Python was used, along with keras's built-in model evaluation function Evaluate and the confusion matrix commonly used in evaluation classification algorithms.

Analysis results show: the built-in evaluation function evaluates model accuracy at 94%, and the confusion matrix evaluates each classification set's accuracy at 93.7%, as shown in Figure 20 [FIGURE:20]. Figure 21 [FIGURE:21] compares manual statistics and neural network algorithm statistics. The maximum difference lies between cloudy and overcast conditions, mainly because sample sizes were small during model training, with each sample set containing fewer than 100 data groups. However, in processing time, the machine learning algorithm is much more efficient than manual processing. For the aforementioned 4089 data groups, the algorithm can complete statistics in about 1 minute, while manual processing would take at least 1 hour. For large-scale batch processing, machine learning algorithms are significantly more efficient.

Compared with other algorithms, such as grayscale value aggregation algorithms [5], convolutional neural networks have advantages in complexity, efficiency, and accuracy for specific applications (where precise cloud calculation is not required). The comparison results are shown in Table 1 [TABLE:1]. Of course, for applications requiring precise cloud calculation, neural networks would need more refined classification and larger training samples, possibly even complex image processing algorithms. However, for radio telescope astronomical site selection, precise cloud data is not needed—only long-term statistical data reflecting the general atmospheric conditions above the site is required, to be combined with other environmental data to reflect site weather conditions. For this reason, convolutional neural network algorithms have significant advantages.

4 Summary

All-sky camera monitoring systems play an important role in real-time cloud cover information monitoring and constitute an important component of optical telescope and radio telescope observatory site monitoring systems. This system aims at the main characteristics of millimeter-wave and submillimeter-wave radio telescope site selection, where candidate sites are basically in high-altitude, harsh environments, uninhabited areas without any support conditions. The design scheme must meet requirements for simple power supply (solar power), simple installation, and unmanned automatic operation. In the future, the BeiDou system's short message communication can be used to directly achieve remote monitoring. Centered on these special requirements, the system hardware innovatively selects cost-effective planetary cameras to obtain sky images. By using low-power, miniaturized embedded controllers that automatically adjust camera parameters based on actual sky brightness, unmanned operation of the monitoring system is achieved.

In data processing, deep learning convolutional neural network algorithms are pioneeringly introduced. Based on Python language, deep learning algorithm software is written. Through extensive sample machine training, feature values of various cloud images are extracted, enabling effective processing of massive data (tens of thousands scale), achieving batch data processing and fully automatic statistical analysis of site weather conditions.

In subsequent data processing algorithm research, main research contents can include: image preprocessing, eliminating misjudgment in sun and moon regions, eliminating reflections; researching the definition and calculation of whole-sky cloud cover under convolutional neural network algorithms; increasing neural network model training sample data to make models more reliable; conducting remote data transmission experiments, transmitting processed data results via BeiDou short message to achieve remote monitoring of weather conditions at field sites.

References

[1] ZHANG Hai-long, YAO Qi-jun, SUN Ji-xian, et al. Qinghai Science and Technology, 2023, 3: 23
[2] PENG Huan-wen, XIN Yu-xin, HE Shou-sheng. Astronomical Research & Technology, 2015, 12: 89
[3] TIAN Jian-feng, DENG Li-cai, YAN Zheng-zhou. Acta Astronomica Sinica, 2016, 57: 366
[4] FAN Liang, LEI Cheng-ming. Acta Astronomica Sinica, 2016, 57: 461
[5] WEI Shi-ya, QIU Bo, CAO Zi-huang. Astronomical Research & Technology, 2019, 16: 85
[6] ZHANG Yu-hong, BAI Ren-xiang, MENG Fan-jun. New Technology & New Process, 2021, 1: 52
[7] ZHOU Fei-yan, JIN Lin-peng, DONG Jun. Chinese Journal of Computers, 2017, 40: 1229
[8] GUO Li-li, DING Shi-fei. Computer Science, 2015, 42: 28
[9] LIU Fang-yuan, WANG Shui-hua, ZHANG Yu-dong. Computer Engineering and Applications, 2018, 54: 11
[10] ZHENG Yin, CHEN Quan-qi, ZHANG Yu-jin. Journal of Image and Graphics, 2014, 19: 175
[11] CONG Shuang, DAI Yi. Computer Applications, 2004, 24: 18
[12] ZHANG Rong-lei, TIAN Ai-kui, TAN Hao. Journal of Shandong University of Technology (Natural Science Edition), 2018, 32: 48
[13] LUO Jia-jie, SHI Jia-lin. Science and Technology Innovation Herald, 2019, 24: 130
[14] XU Feng, LU Jian-gang, SUN You-xian. Information and Control, 2003, 32: 344
[15] CHANG Liang, DENG Xiao-ming, ZHOU Ming-quan. Acta Automatica Sinica, 2016, 42: 1300