The Comprehensive Identification Dataset Development for Dew, Frost, and Icing Phenomena in China (2018?C2024)

ZHU Hualiang¹   ZHANG Miaomiao^1*  HONG Chen¹  WEN Huayang²

1. Anhui Meteorological Information Center, Anhui Meteorological Bureau, Hefei 230031, China;

2. Huaihe River Basin Meteorological Center, Anhui Meteorological Bureau, Hefei 230031, China

Abstract: To achieve automatic observation of dew, frost, and icing phenomena while acquiring timely, comprehensive, and continuous observational data of these weather phenomena occurrences, this paper develops a comprehensive identification algorithm for the 3 weather phenomena at 2,164 surface meteorological stations across China. Using the Bayesian discriminant method, the algorithm utilizes observational data of key meteorological elements, including air temperature, surface temperature, relative humidity, and wind speed. A long-time identification product is hereafter generated, covering the period 2018?C2024 with an hourly temporal resolution. A comparative analysis with manual observation data demonstrates that the proposed product achieves agreement rates of 65.57%, 90.90%, and 95.26% for dew, frost, and icing, respectively, indicating relatively robust performance. Presently, this time-series identification product has been officially incorporated into the operational workflow of meteorological departments across China, effectively substituting manual observations of the aforementioned phenomena at surface meteorological stations. It thereby lays a solid foundation for advancing the automation of comprehensive meteorological observations and facilitating the unattended operation of national-level surface meteorological stations across China. Additionally, the long-time dataset can serve as fundamental data to support weather forecasting, agricultural meteorology, road traffic forecasting, and service provision.

Keywords: dew phenomenon; frost phenomenon; icing phenomenon; Bayesian discrimination; comprehensive identification

DOI: https://doi.org/10.3974/geodp.2026.01.11

Dataset Availability Statement:

The dataset supporting this paper was published and is accessible through the Digital Journal of Global Change Data Repository at: https://doi.org/10.3974/geodb.2025.10.04.V1.

1 Introduction

Dew, frost, and icing are common weather phenomena in daily life, exerting significant impacts on human production and daily activities. For instance, dew serves as a crucial source of water resources and humidity in arid and semi-arid regions, playing an indispensable role in plant growth and development^[1]; Autumn early frost and spring late frost can cause freeze damage to plants (especially crops)^[2,3]; In winter, snowfall and rainfall events are prone to result in road snow accumulation or freezing, which severely hinders people??s travel and daily routines^[4,5]. Therefore, meteorological departments attach great importance to the monitoring and forecasting of these 3 weather phenomena. Timely, comprehensive, and continuous observational data of them are conducive to agricultural meteorological and road traffic forecasting and service provision, and serve as a scientific basis for effective disaster prevention, emergency rescue, and other related initiatives^[6?C10].

In the early days, the monitoring of dew, frost, and icing phenomena relied on manual visual observations^[11]. However, this method has several inherent drawbacks, such as strong subjectivity, low observation frequency, sparse distribution of observation stations, and high investment and maintenance costs^[10,12]. To achieve automated observation of these weather phenomena occurrences, researchers worldwide have primarily conducted experimental research based on tracer-based imaging^[13], fiber optic sensors^[14,15], optical sensors^[16], and microlysimeters^[17]. Nevertheless, due to the high cost of related equipment and suboptimal identification performance, there are currently no mature automated observation devices or corresponding operational products available globally.

Frost, dew, and icing phenomena occur under specific meteorological conditions, being the comprehensive outcomes of variations in multiple meteorological elements. To achieve timely, comprehensive, and continuous automated observation of these weather phenomena occurrences, this study leverages existing observational data to develop an integrated identification algorithm for the 3 weather phenomena based on the Bayesian discriminant method. A corresponding time-series identification product has been generated since 2018, which can effectively substitute manual observations and provide fundamental data support for weather forecasting, agricultural meteorology, and road traffic forecasting and service provision.

2 Metadata of the Dataset

The metadata of The comprehensive identification dataset for dew, frost, and icing phenomena in China (2018?C2024)^[18] is summarized in Table 1. It includes the full name, short name, authors, geographical region, year of the dataset, temporal resolution, data format, data size, data files, data publisher, and data sharing policy, etc.

3 Methods

3.1 Algorithm

3.1.1 Selection of Relevant Factors

The formation of dew, frost, and icing is influenced by meteorological factors, including temperature, humidity, and wind speed, reflecting the comprehensive effects of variations in these elements. Nevertheless, the magnitude of influence exerted by different meteorological factors on their formation differs significantly. Using surface observation data from more than 600 reference climate stations and basic meteorological stations across China during the period 2003?C2013[1], the correlation coefficients between the occurrence of dew, frost, and

Table 1  Metadata summary of The comprehensive identification dataset for dew, frost, and icing phenomena in China (2018?C2024)

icing correlation coefficients between the occurrence of dew, frost, and icing and key meteorological variables, including air temperature, surface temperature, vapor pressure, and wind speed are calculated. Further correlation analysis between these phenomena and additional meteorological factors indicated that the formation of dew and frost exhibits significant correlations with air temperature, surface temperature, relative humidity, vapor pressure, and wind speed. Specifically, dew formation exhibits a significant positive correlation with air temperature, surface temperature, and dew point temperature, while showing a significant negative correlation with wind speed. Frost formation, in contrast, exhibits a highly significant negative correlation with air temperature, surface temperature, vapor pressure, and wind speed. For icing, its most pronounced correlation is with the minimum air temperature and minimum surface temperature, showing a significant negative correlation. These findings confirm that different meteorological factors exert varying magnitudes of influence on the formation of dew, frost, and icing, with temperature acting as the primary influencing factor. This study thus provides appropriate identification indicators for the development of discriminant models for dew, frost, and icing events.

3.1.2 Bayesian Discriminant Model Construction

The Bayesian discriminant model is a statistical classification and decision-making model grounded in Bayesian theorem. Its core principle entails calculating the posterior probabilities of a sample belonging to distinct categories and assigning the sample to the category with the maximum posterior probability. In the context of weather phenomenon identification, the model first adopts the prior probabilities of target weather phenomena (i.e., dew, frost, and icing)??specifically, the probabilities of the phenomenon occurring or not occurring??and integrates them with the daily observed meteorological element data. It then separately computes the conditional probabilities for 2 scenarios: ??the weather phenomenon is present?? and ??the weather phenomenon is absent??. Finally, by comparing the numerical values of these 2 sets of probabilities, the model determines whether the target weather phenomenon occurs on that day. The training method and discrimination procedures of the Bayesian discriminant model for a specific weather phenomenon (dew, frost, or icing) at an individual meteorological station are outlined as follows^[10,12]:

(1) Using surface meteorological observation data from the meteorological stations spanning the period 2003?C2013 as training samples, the probabilities of the target weather phenomenon occurring and not occurring in the samples are calculated as  and ,  respectively.

(2) Assume that the observed values of relevant meteorological elements on a given day are denoted as , where  represents key variables including daily minimum air temperature, daily minimum surface temperature, and other relevant meteorological factors. Based on the training sample set, the conditional probabilities  (for ) of observing the meteorological element values X are calculated under two mutually exclusive scenarios: the non-occurrence and occurrence of the target weather phenomenon. Here,  denotes the non-occurrence of the target weather phenomenon, and  denotes its occurrence.

(3) According to Bayesian theorem, the posterior probabilities  (for ) of the target weather phenomenon (i.e., dew, frost, and icing) being absent or present on that day are calculated under the condition of the observed meteorological element values X. The calculation equations are provided as follows:

                                                                                                         (1)

If the posterior probability of the target weather phenomenon being absent, , is greater than the posterior probability of it being present, , the phenomenon is determined to be non-existent on that day; otherwise, it is deemed to have occurred. That is:

                                                           (2)

Equation 2 can be simplified as:

                                                                                                                                       (3)

Assuming that the meteorological elements are mutually independent and each follows a normal distribution  under the scenarios of the target weather phenomenon being absent or present, the following holds:

                                                            (4)

In the equation,  and  denote the mean and standard deviation of the j-th feature under the i-th class, respectively. These parameters can be estimated via the maximum likelihood estimation (MLE) method using the training sample set.

(4) Based on the daily observed values of relevant meteorological elements (e.g., daily minimum air temperature and daily minimum surface temperature), the identification of the target weather phenomenon can be accomplished using Equation 4. If , the phenomenon is determined to be absent on that day; otherwise, it is deemed to have occurred.

These steps are replicated separately for the dew, frost, and icing phenomena at 2,164 meteorological stations across China. For the Bayesian discriminant models of dew and frost, the selected meteorological elements include air temperature, surface temperature, water vapor pressure, wind speed, and relative humidity. For icing, the selected meteorological elements are air temperature and surface temperature. During the model construction process, Bayesian discriminant models for dew are successfully established at 1,202 meteorological stations, while model establishment failed at 962 stations due to poor-quality dew observation data. For frost, 81 stations lacked historical observation data to serve as modeling samples (attributed to late station establishment or other reasons), making the Bayesian discriminant models for frost unfeasible. Regarding icing discriminant models, 19 stations failed to establish such models due to insufficient historical observation data (caused by late station establishment or other factors). Surface meteorological observation data from 2014 to 2015 were employed as independent samples to validate the discriminative performance of the Bayesian discriminant models. The results showed that the average discriminant accuracy rates of the dew, frost, and icing models reached 86.1%, 91.8%, and 96.9%, respectively. This indicates that the established Bayesian discriminant models are scientifically reasonable and practically acceptable.

3.2 Technical Route

To comprehensively account for the impacts of special weather conditions (e.g., rain and snow), the following steps are adopted for the integrated identification of dew, frost, and icing phenomena:

Step 1: Threshold judgment. Based on the surface observation elements listed in Table 2, determine whether the current time period is favorable for the occurrence of dew, frost, or icing phenomena. If the conditions are not favorable for the occurrence of any of these 3 phenomena, the corresponding weather phenomenon is directly determined to be absent during the current time period; otherwise, proceed to Step 2.

Step 2: Bayesian model judgment. First, confirm whether a Bayesian discriminant model has been constructed for the current meteorological station. If the corresponding Bayesian discriminant model has been established, the model is employed to identify the target weather phenomenon; otherwise, proceed to Step 3.

Step 3: Bayesian model judgment based on reference stations. First, confirm whether there is a qualified reference station for the target meteorological station. If a reference station exists, the Bayesian discriminant model of this reference station is employed to identify the corresponding weather phenomenon; otherwise, proceed to Step 4. The selection

Table 2  Weather conditions unfavorable for the formation of dew, frost, and icing phenomena

Note: The occurrence of any of the aforementioned conditions is unfavorable for the formation of the corresponding weather phenomenon.

criteria for reference stations are specified as follows: (1) The altitude difference from the target station is less than 200 m, and the geographical environment is similar; (2) The correlation coefficient of the daily average air temperature with the target station is greater than 0.9, and the result passes the significance test at the 0.05 level; (3) The straight-line distance from the target station is less than 100 km; (4) A valid Bayesian discriminant model for the corresponding weather phenomenon has been successfully constructed. If multiple candidate reference stations meet the above criteria, the one with the shortest straight-line distance to the target station is selected as the final reference station.

Step 4: Threshold judgment. Due to persistently high (or low) temperatures throughout the year at meteorological stations in coastal areas such as Guangdong, Guangxi, Hainan and high-altitude regions, there are no dew, frost or icing events, or only a few days with such events in the modeling sample. Therefore, these stations were excluded during the model construction process. For stations where neither a Bayesian discriminant model can be constructed nor a reference station model is available in real-time calculations, the element threshold method is directly adopted for judgment, as detailed in Table 3.

Table 3  Discrimination thresholds for dew, frost and icing phenomena

Note: the corresponding weather phenomenon is determined to have formed only when all the above conditions are satisfied simultaneously.

4 Data Results and Validation

4.1 Dataset Composition

This dataset comprises hourly observational records of dew, frost, and icing phenomena from 2,164 surface meteorological stations across China, spanning the period from November 15, 2018, to December 31, 2024. The data are archived in .txt files named after the station numbers, with values delimited by semicolons ( ?? ; ?? ) and missing data indicated by slashes ( ?? / ?? ). The file format and data structure of the physical files are detailed in Table 4.

Table 4  Composition of the dataset file content

4.2 Data Results

In accordance with the provisions of the Specifications for Surface Meteorological Observations^[11]: ??If the number of missing observations in a month reaches 7 or more, the month shall be excluded from monthly statistics and treated as missing data; if a year contains 1 or more months excluded from monthly statistics, the year shall be excluded from annual statistics and treated as missing data??. Based on The comprehensive identification dataset for dew, frost, and icing phenomena in China (2018?C2024), annual day-count series of dew, frost, and ice from 2019 to 2024 at 2,164 surface meteorological stations across China were statistically derived (data for 2018 were excluded from annual statistics due to only 1.5 months of records being available). Figures 1?C3 illustrate the spatial distributions of the annual average days of dew, frost, and icing, respectively, which were generated using the thin plate spline interpolation method. As presented in Figure 1, the annual dew days in China generally exhibit a spatial pattern of ??more in the south and fewer in the north??. Southwest China records the longest annual dew days nationwide, with most stations documenting over 250 d??for example, Kunming (271 d), Dali (337 d), and Suichuan County in Jiangxi Province, which has the highest number of dew days (361 d) in the country. In contrast, the northern regions have relatively fewer annual dew days, particularly in Northwest China, where most stations report fewer than 50 d. For instance, Hami and Karamay in Xinjiang both have approximately 20 d of dew per year, with Turpan in Xinjiang recording the minimum of 1 d. Notably, alpine meteorological stations observe fewer dew days compared to surrounding lowland stations??for example, Wutai Mountain Station has an average of only 30 d, and Songshan Mountain Station only 2 d, both significantly lower than those at nearby lowland stations.

As presented in Figure 2, the spatial distribution of frost days exhibits an opposite pattern to that of dew days, characterized by ??more in the north and fewer in the south??. High-value regions of annual frost days are concentrated in the Sichuan-Qinghai border area, the eastern parts of Northeast China (encompassing Heilongjiang, Jilin, and Inner Mongolia), and northern Xinjiang. Among all observational stations, Maduo County in Qinghai Province records the highest annual frost days at 251. Stations including Mohe in Heilongjiang, Seda in Sichuan, Bayanbulak in Xinjiang, and Arxan in Inner Mongolia all document over 200 frost days per year. Frost is absent in areas south of 21??N latitude. Notably, Menghai in Yunnan Province??situated north of 21??N??has the fewest frost days in this latitudinal zone, with an annual record of only 1 frost day.

As presented in Figure 3, the spatial distribution of icing days in China exhibits a pattern of ??more in the north and fewer in the south??, with high-value regions concentrated in Qinghai Province, Xinjiang Uygur Autonomous Region, and Northeast China. For instance, Wudaoliang Station in Qinghai records 282 annual icing days, while Tianshan Daxigou Station in Xinjiang documents 232 d. In Northeast China, both Xinlin and Mohe in Heilongjiang Province have over 220 icing days per year. Icing phenomena are absent in areas south of 23??N latitude. North of 23??N, stations with the fewest icing days??such as Tengchong in Yunnan Province and Zijin in Guangdong Province??record only 1 icing day annually.

Figure 4 illustrates the national average annual sequence of dew days, frost days, and icing days in China from 2019 to 2024. The multi-year national average annual dew days amount to 196, with an average of 54 frost days and 65 icing days per year (due to widespread missing data exceeding 7 d at many stations in July 2020, the statistics of dew days, frost days, and ice days for 2020 are recorded as missing). As presented in the figure, the annual average dew days exhibit an overall increasing trend, rising from 164 d in 2019 to 205 d in 2024, with a cumulative growth rate of 5.48%. Unlike the continuous growth of dew days, the frost day sequence remains relatively stable??recording 45 d both in 2019 and 2021??before increasing to 62 d in 2024. The minimum number of icing days occurred in 2021 (46 d), while in other years, the number of icing days remained around 70 d, showing a relatively gentle variation trend.

4.3 Data Validation

The performance of the comprehensive identification product is evaluated using manual observation data from November 15, 2018, to November 14, 2019. Given that the manual observation data are of daily frequency, whereas the comprehensive identification product adopts an hourly resolution, a temporal aggregation approach is employed to facilitate performance evaluation: if the product detects the occurrence of dew, frost, or icing phenomena at any hour within a single day, the corresponding weather phenomenon is deemed to have occurred on that day. Subsequently, the consistency between the aggregated product results and the manual observation data is quantified. A higher consistency rate indicates superior discriminative performance of the product, while a lower rate reflects inadequate performance.

Table 5 presents the overall evaluation results of the comprehensive identification product. For the dew identification product, the number of consistent cases where both national manual observations and the product recorded dew occurrence was 151,476, while the number of consistent cases where both recorded no dew was 432,211, resulting in an overall consistency rate of 65.57% for dew identification. The number of dew occurrences detected by the product is significantly higher than that recorded by manual observations. This discrepancy is partly attributed to the fact that dew often forms first on winter nights and subsequently freezes into frost; in early morning observations, only frost is recorded, with dew omitted. On the other hand, according to meteorological observational standards, dew formed by melted frost is not counted as valid dew. In winter, mixed occurrences of frost and dew are common, and dew formation during frost melting also contributes to the underrecording of dew in manual observations. For the frost identification product, the number of consistent cases where both manual observations and the product recorded frost occurrence was 95,204, and the number of consistent cases where both recorded no frost was 714,061, with an overall consistency rate of 90.90%. Similarly, the product detected more frost occurrences than manual observations. This is primarily due to the nocturnal formation characteristic of frost: after 2013, nighttime observation periods were canceled at national general meteorological stations, leading to the underrecording of nighttime frost in manual observations. Regarding the icing identification product, the number of consistent cases where both manual observations and the product recorded icing occurrence was 176,285, and the number of consistent cases where both recorded no icing was 671,718, achieving an overall consistency rate of 95.26%. Compared with manual observations, the product exhibits fewer underdetection and misdetection events, and can effectively reflect the actual occurrence of icing phenomena.

Table 5  Overall evaluation results of the comprehensive identification product for dew, frost, and icing Phenomena

5 Discussion and Conclusion

This product is primarily generated through a comprehensive identification process, which integrates Bayesian discriminant models (constructed via the Bayesian discriminant method) for dew, frost, and icing phenomena with surface meteorological observation elements (e.g., temperature, humidity, and wind speed). Verification and evaluation results indicate that the overall consistency rates for the comprehensive identification of dew, frost, and icing are 65.57%, 90.90%, and 95.26%, respectively. However, the identification accuracy of certain stations remains unsatisfactory, with relatively frequent missed detections and false alarms. This is partly attributed to the use of manual observation data for both model training and validation: factors such as strong subjectivity in manual recording and low observation frequency undermine the reliability of the datasets. On the other hand, during the construction of the Bayesian discriminant models, limitations including the insufficient selection of influencing factors, inadequate data volume, and uneven data distribution have led to compromised generalization ability of the trained models. To address the these issues, optimizations can be implemented from 3 aspects: (1) Integrate automated observation devices (e.g., intelligent video-based weather phenomenon monitors) to collect real-scenario data, supplementing small-sample categories and edge-case data; (2) Strengthen dataset cleaning and anomaly detection, utilizing clustering algorithms to identify and filter outliers, thereby improving data reliability; (3) Replace the Bayesian models with moderately complex models (e.g., deep learning models) and conduct multi-dimensional performance evaluation using metrics such as precision, recall, and F1-score to refine model performance.

This product is now officially operational nationwide across China??s meteorological departments, replacing manual observations of dew, frost, and icing phenomena. It can provide fundamental support for weather forecasting, agrometeorological services, road traffic forecasting, service provision, and other related fields. However, the product is limited by its low spatial resolution, which is insufficient to meet the requirements of refined meteorological services. In future work, combined with surface gridded analysis data, efforts will be devoted to developing national gridded analysis products for dew, frost, and icing phenomena.

Author Contributions

Wen, H. Y. contributed to the overall design of the dataset development. Zhu, H. L. designed the models and algorithms, and conducted the evaluation and validation of the dataset product. Zhang, M. M. and Hong, C. collected and preprocessed the observational data of weather conditions conducive to the formation of dew, frost, and icing phenomena. Zhu, H. L. and Zhang, M. M. wrote the data paper.

Conflicts of Interest

The authors declare no conflicts of interest.

References

[1]        Ma, B., Tian, J. C., He, J. Y., et al. Mechanism of dew formation in the arid zone of central Ningxia and its impact on surface soil moisture [J]. Advances in Water Science, 2022, 33(6): 955?C966.

[2]        Wang, Y., Qiu, X. L., Li, Y. X., et al. Risk assessment of forest and fruit yield reduction caused by frost damage based on phenological period: a case study of Hebei Province [J]. Chinese Journal of Agrometeorology, 2022, 4(10): 821?C831.

[3]        Zhang, B., Sun, S. S., Ding, L. G., et al. Hazard analysis and zoning of spring tea frost damage in Guizhou [J]. Journal of Meteorology and Environment, 2023, 39(5): 99?C105.

[4]        Bao, L. L., Cheng, P., Wang, X. Y., et al. Road icing early warning in Gansu Province based on Logistic regression and neural network [J]. Journal of Arid Meteorology, 2024, 42(1): 137?C145.

[5]        Zhang, H. F., Lu, S., Shen, J. J., et al. Spatiotemporal variation characteristics of road icing in Shaanxi and its risk early warning model [J]. Journal of Arid Meteorology, 2020, 38(5): 878?C885.

[6]        Song, P., Che, J. H., Guo, T. T., et al. Low-temperature climatic characteristics and SVM prediction model of the highway pavement around Jiaozhou Bay [J]. Journal of Marine Meteorology, 2023, 43(3): 80?C87.

[7]        Wang, K. X., Bao, Y. X., Zhu, C. Y., et al. Application of random forest regression in winter pavement temperature prediction [J]. Meteorological Monthly, 2021, 47(1): 82?C93.

[8]        Zou, L. J., Liu, S., Lu, Q. J. Pavement temperature model and icing potential based on neural network [J]. Highway, 2022, 67(10): 409?C414.

[9]        Zhang, Q. K., Xiang, Y., Ji, Z. M., et al. Climatic characteristics and trend analysis of icing phenomena in Anhui Province in recent 55 years [J]. Journal of Natural Disasters, 2020, 29(6): 218?C226.

[10]     Hua, L. S., Wen, H. Y., Zhu, H. L., et al. Discussion on automatic observation model of frost formation based on Bayesian discrimination method [J]. Meteorological Monthly, 2015, 41(8): 964?C969.

[11]     China Meteorological Administration. Specifications for Surface Meteorological Observations [M]. Beijing: China Meteorological Press, 2003: 21?C27.

[12]     Wen, H. Y., Zhu, H. L., Ma, W. Z., et al. Correction of icing phenomenon data series based on Bayesian discrimination method [J]. Meteorological Monthly, 2021, 47(9): 1113?C1121.

[13]     Shi, C. C. Pavement icing monitoring and early warning system based on machine vision [J]. Electronic Design Engineering, 2025, 33(6): 34?C38.

[14]     Zhao, X. K., Hu, Z., Zhang, J. P., et al. Research progress on intelligent monitoring of pavement icing based on optical fiber sensing technology [J]. Journal of Jilin University (Engineering and Technology Edition), 2023, 53(6): 1566?C1579.

[15]     Xie, Q. Z., Wang, L., Ge, J. Y., et al. Design of polymer optical fiber icing sensor and method for detecting pavement ice layer thickness [J]. Journal of China & Foreign Highway, 2023, 43(4): 59?C67.

[16]     Ma, S. Q., Wu, K. J., Chen, D. D., et al. Design of automatic observation system for weather phenomena [J]. Meteorological Monthly, 2011, 37(9): 1166?C1172.

[17]     Ran, B., Zhang, Z. Y., Yang, J. B., et al. Formation law of condensed water on Artemisia ordosica in the Mu Us Sandy Land and its impact on water balance [J]. Transactions of the Chinese Society of Agricultural Engineering, 2023, 39(8): 111?C119.

[18]     Zhu, H. L., Zhang, M. M., Hong, C., et al. The comprehensive identification dataset for dew, frost, and icing phenomena in China (2018?C2024) [J/DB/OL]. Digital Journal of Global Change Data Repository, 2025. https://doi.org/10.3974/geodb.2025.10.04.V1.

[19]     GCdataPR Editorial Office. GCdataPR data sharing policy [OL]. https://doi.org/10.3974/dp.policy.2014.05 (Updated 2017).