1. Introduction
Chengdu is a megacity located on the eastern edge of the Tibetan Plateau, with a population of over 20 million and a vehicle count of over 6 million. Chengdu has been described as a “park city under the snowy peaks” because of the beautiful view of the snowy hills of the Qionglai Mountains on the Tibetan Plateau from the city. Approximately 1200 years ago, Chinese poet Du Fu paid tribute to this place by writing in his famous poem “Jueju, third in a set of four” while he was living in Chengdu in the year 764: “Framed in my window, the thousand autumn snows of the western peaks, tied by my gate, a boat to take me ten thousand miles east to Wu” [
1].
Although the anthropogenic emissions in Chengdu from industry, transportation and residents are massive, the city has seen significant improvements in air quality. The number of days with good and moderate air quality increased from 214 days in 2015 to 299 days in 2021, and the annual average PM2.5 concentration dropped from 57.4 μg/m3 in 2015 to 39.8 μg/m3 in 2021. Yaomei Feng of the Qionglai Mountains on the Tibetan Plateau has also been observed in Chengdu on more days because of improved air quality. The sight of snowy peaks is not only a city attraction, but also a testimony to the improvement of air quality. Since 2021, the Chengdu Meteorological Bureau has made empirical forecasts of snow peak observations based on weather forecasts. However, these empirical forecasts are dependent on individual subjective experience and are relatively unstable, so an objective forecasting system for snow peak sighting was developed in this study to meet current needs.
The ability to observe the snow-capped mountains depends on two main factors: the effect of visibility and the effect of terrain obscuration. Visibility is mainly influenced by pollutant concentrations and meteorological conditions [
2,
3,
4]. Under dry or low PM
2.5 concentrations, PM
2.5 concentrations dominate visibility, and the contribution of humidity becomes increasingly important as PM
2.5 concentrations and humidity increase [
5]. Zhou et al. [
6] constructed a generalized linear model to estimate the site-specific effects of PM
2.5 and relative humidity on visibility based on data from 190 monitoring stations in 182 cities in China. The researchers found that both PM
2.5 and relative humidity had significant nonlinear negative effects on visibility. Liu et al. [
7] also found that organic matter and NH
4NO
3 in PM
2.5 had a great influence on visibility in Chengdu.
Meteorological services on landscapes have been widely carried out in China in recent years, and cloudscapes were forecasted by DING et al. [
8] using multiple models, including logistic regression, support vector machine, decision tree analysis, and empirical prediction. The cloudscape on Emei Mountain was also predicted by Tao et al. [
9]. Chengdu is in a complex topographic area on the eastern edge of the Tibetan Plateau [
10] and is surrounded by Longmen Mountain and Longquan Mountain, making it impossible to observe snow-capped mountains in a part of Chengdu without terrain obstruction; thus, in addition to visibility effects, terrain obstruction is a factor that should be considered when determining whether snow-capped mountains can be observed. The study tested terrain occlusion using high-resolution digital elevation model (DEM) data [
11]. In combination with in-grid multi-point sampling detection, each grid was assigned a terrain occlusion rate, which was used to indicate the probability that the grid might be occluded by terrain, and the probability was applied to the forecast system.
Currently, most research related to snow peak observations relies on remote sensing technology. For instance, D. H. Shugar et al. [
12] employed satellite imagery, seismic records, numeric model results, and eyewitness videos to analysis a massive rock and ice avalanche disaster in 2021. Another study conducted by Ya-Lun S. Tsai et al. [
13] reviewed snow cover remote sensing using spaceborne synthetic aperture radar, emphasizing the importance of snow cover monitoring. In complex landscapes, satellite-based snow cover monitoring has been widely implemented, Kostadinov et al. [
14] developed a lidar-based method to detect snow cover under canopy, and introduced a correction factor to improve satellite-derived fractional snow cover area products. Gascoin et al. [
15] estimated the snow cover in open terrain based on Sentinel-2 data using a normalized difference snow index. Muhuri et al. [
16] assessed the performance of the operational snow detection and fractional snow cover estimation algorithms for high-resolution snow and ice monitoring. However, there is limited research available on predicting whether snow can be observed visually.
Based on numerical air quality forecast technology, a snow peak sighting probability (SPSP) forecast system was developed using R [
17,
18,
19,
20,
21,
22,
23,
24,
25]. Weather and air quality models, including the Weather Research and Forecasting (WRF) model version 3.5 [
26] and the Community Multiscale Air Quality (CMAQ) model version 5.0.2 [
27], were used to provide meteorological and environmental predictions over the region. A threshold methodology was implemented to determine if snow peaks could be seen from certain locations in Chengdu. Monte-Carlo simulation was used for parameter optimization.
3. Results
3.1. Base Model Evaluation
Meteorological and air quality modeling performances were important for snow peak sighting forecasts, and hindcast results for January, April, July, and October 2020 were used to assess the performance of WRF and CMAQ.
Normalized mean bias (
NMB), root mean square error (
RMSE), and coefficient of correlation (
R) were estimated.
NMB was used as a normalization to facilitate a range of modeling result magnitudes; the closer to 0 the better.
RMSE was used to provide information on the standard deviation of the model prediction error [
42].
R was used to measure the degree of linear association between observations and model results, and normally the absolute values of
R should be larger than 0.4 [
43,
44]. These metrics were defined in Equations (5) to (7), where
M represents modeled values and
O stands for observed ones:
Observations including 10 m wind speed, relative humidity, atmospheric pressure and 2 m temperature from 13 national meteorological stations were used to evaluate the performance of the WRF model, as shown in
Figure 4.
In terms of the spatial distribution of correlation coefficients for meteorological items, the performances of the model on 2 m temperature and atmospheric pressure were relatively good for all sites over Chengdu. However, model results of relative humidity were better for southern sites, and owing to the complex terrain over western sites, the performance of 10 m wind speed were worse in those sites.
Table 3 shows the statistical results of meteorological fields of 13 national stations in January, April, July, and October, with approximately 35,000 observations. The NMB value of T2 was 3.66% and that of PRES was 0.12%, the NMB value of WS was over 40%, which indicated that the modeled values were obviously high; and the NMB value of RH was about −10%, which might be due to the overestimation of T2. The RMSE values of T2 and PRES were around 2 hPa. The R values of T2 and PRES were more than 0.95. Owing to the complex terrain over Chengdu, the performance of WRF on winds might be affected, and station observations of WS were more sensitive to terrain and surrounding objectives.
Observations from 35 municipal and national environmental monitoring stations were used to evaluate the performance of CMAQ on PM
2.5, PM
10, NO
2, and O
3, and results are shown in
Figure 5.
Considering the spatial distribution of R values of pollutants, we can infer that the performance of CMAQ was consistent over most of the monitoring sites.
Statistical results of the spatial average of seven national stations are listed in
Table 4. A total of 2426 observed values were used for each variable. The NMB values of PM
2.5 and PM
10 were over −25%, indicating underestimations on particulate matters, and the NMB value of NO
2 and O
3 were significantly lower than that of PM. The RMSE values of PM
2.5, PM
10, NO
2, and O
3 were 26.75, 40.68, 22.06, and 29.66 μg/m
3, respectively. The R value of NO
2 was obviously lower than those of PM
2.5, PM
10, and O
3.
3.2. Terrain Occlusion Results
Chengdu is surrounded by two mountains: the Longmen Mountains are partly located in the west of Chengdu, with an average altitude of 1381 m, and stand in front of Yaomei Feng, while Longquan Mountain is located in central eastern Chengdu and has an average altitude of approximately 1000 m. Most of the terrain occlusions for Chengdu are due to the Longmen Mountains.
As shown in
Figure 6, owing to the presence of mountains, Yaomei Feng can only be seen from a part of Chengdu. Terrain occlusion rates were lower than 50% for only 49.3% of the model grids, and were mostly distributed in the Chengdu Plain between the Longmen Mountains and Longquan Mountain. Terrain occlusion rates were used to adjust the snow peak sighting forecast results for each model grid.
3.3. Threshold Values Acquired by Monte-Carlo Simulation
Random numbers were generated according to the range of each variable, and each random number was used to determine if a certain snow peak could be seen from each model grid. SPSP values for 2020 were calculated and compared with social media posts to estimate the performance of random numbers, and the best was selected as listed in
Table 5.
3.4. Snow Peak Sighting Probability
Annual snow peak sighting probability was calculated and visualized for 2020 as shown in
Figure 7, and terrain occlusion rates were applied so the probability in the west regions were relatively low. Significant differences were shown in the spatial distribution of snow peak sighting probability, and higher values appeared in the west near Longmen Mountains, including Dujiangyan, Wenjiang, Qionglai, and Pujiang. High probability is also shown in some regions of highly urbanized areas such as Jinniu, Qingyang, and Wuhou. It was easier to see Yaomei Feng from places on top of Longquan Mountain in Longquanyi.
As shown in
Figure 8, monthly results indicated that the best time for snow peak sighting was May because it had the highest probability for Yaomei Feng to be visible from Chengdu, followed by August and July. Thus we can conclude that the best snow peak sighting seasons for Chengdu are late spring and summer. Similar spatial distributions were found in different months, and SPSP values were higher for western grids in general.
3.5. Comparison with Social Media Posts
Owning to lack of detailed time and location information, posts collected from social media might not be a good reference for snow peak sighting hindcast evaluation, but no better source of observations are available at present. As shown in
Figure 9, the number of days is shown as the height of bars, and averaged SPSP values are shown as fill colors. Compared with social media posts, the model results were higher in every month, especially in January, April, and December. Significant over-estimation was also found in May. SPSP values were higher from May to September, indicating that snow peaks could be seen from more sites in those months.
Figure 10 shows a detailed comparison between modelled and posted days with sighting records on Yaomei Feng from May to September 2020. The variation of total sighting days in each month fitted the observations well, but more false positives were found, especially in April, May, and December.
As listed in
Table 6, the accuracy of the model was 78.9%, and the recall value was 57.1%, but the precision was only 24.4% because of a large number of false positives. However, some true positives might be counted as false when using social media posts, so it was important to build snow peak observation stations for further model optimization.
4. Summary and Conclusions
The development of a threshold-method-based snow peak sighting probability forecast system was introduced in this paper. The system was based on a novel threshold method; meteorological and environmental fields provided by numerical models including WRF and CMAQ were used as input to the snow peak sighting forecast system. Threshold values used in the forecast system were determined through Monte Carlo simulations. A WRF-CMAQ hindcast of 2020 was carried out to establish the forecast system. The paper provided an idea to develop a snow peak sighting forecast model based on limited observations.
The following conclusions could be drawn:
A logical formula considering meteorological and environmental was proposed as the core of the snow peak sighting forecast system.
The spatial and temporal distribution of snow peak sighting probabilities were analyzed in this paper, showing that the best regions for snow peak sighting in Chengdu were in the plain between Longquan and Longmen Mountains, and the best periods for snow peak sighting were between May to September.
Estimations showed that the accuracy of the system was 78.9%, while the precision was 24.4%, and a lack of dependable observations caused difficulty in determining the cause of false positives.
The system was still in an early stage when this study was conducted, and a seven-day operational forecast has been carried out in Chengdu since May 2022. An automatic photo system was designed with a telescope equipped with a Raspberry Pie Camera, which was used to provide observations for model optimization before official stations were built. Additionally, an app named “Seeing the Snow Peaks” was developed for citizens to access the forecast products and upload their photos with locations, which could also be used as observations in the future. Although the system was designed to carry out forecasts for multiple snow peaks, Monte Carlo simulations were found necessary for peaks other than Yaomei Feng if observations were available. Utilizing the technique described in this paper to identify optimal observation sites and time periods for glaciers or snow can prove advantageous for cryosphere observations. Additionally, by developing a threshold method that takes into consideration the meteorological and environmental conditions above the snow peaks, this technique may serve as a reference for satellite observations of snow-covered peaks.
Optimizations on the model could be done in the future with technologies including machine learning and accumulated visibility calculation, and hindcasts for different years should be conducted to discuss the differences of spatial and temporal distribution of snow peak sighting probability among years. Even with limited observations, the methodology in this paper could be applied to any snow peak sighting points, and forecasts could be provided for snow peak tourism lines and snow resorts.