Meteorological data

3.2.1. Meteorological stations

Four meteorological stations were installed in winter 2017, which included three stations along forest road #100 in Chilan and a station located at forest road #110 in Mingchih (Figure 3.1). Three stations were mounted on the ground and the other one on the flux tower at ~3 m above ground. All stations were positioned without any canopy obstructed.

The stations were situated along an elevation gradient from 1151 to ~1810 m a.s.l.

(namely MC, 9K, 14.5K and 30K, Figure 3.1). The spatial extent of the stations was 15.5 km, and the slope aspect of them were different from each other (Table 3.1). The station sensors collected temperature (T, ºC), relative humidity (RH, %), rainfall (mm), and photosynthesis photon flux density (PPFD, but called PAR hereafter)( PAR Photon Flux Sensor Model S110, Decagon Devices, Inc, WA, USA) (μmol m^-2s^-1) with the time frequency of two minutes. PAR was commonly defined as wavelengths between 400 and 700 nm of solar radiation, where was the primary spectral regions for photosynthesis. We downloaded the data and checked instruments condition in the field once a month since November 2017, and 2+ years of data have been recorded. In this study, our analyzed datasets were time span from 2018/1/1 to 2019/12/31.

Table 3.1 Geographical information of four meteorological stations. From MC to 30K, the location shift from north east to south west and the elevation increase gradually.

Plot Elevation

3.2.2. Meteorological data gap filling

It was impractical to obtain continuous time-series meteorological data through a long period of time in forests. For instance, the high humidity and diurnal temperature range would challenge battery performance of data logger; the wild animals especially formosan reeve's muntjac in our research region liked to chew the sensor cables; some unknown reasons caused records cannot download. Data missing was inevitable.

We set a rule that if time period of missing data was less than 10 minutes, the gaps were directly applied linear interpolation for T, RH and PAR data. If the gaps were longer than 10 minutes, they were set as NA, and would not be analyzed.

3.3. Himawari-8 satellite data

3.3.1. Himawari-8 data download and retrieved

Himawari-8 (H-8 hereafter) is a Japanese geostationary meteorological satellite, which aboard Advanced Himawari Imagers (AHIs) instrument. The sensor can acquire 16 observation bands, including three visible, three near-infrared, and 10 shortwave infrared bands, with spatial and temporal resolutions of 0.5–2 km and 10 minutes, respectively (Bessho et al. 2016). With these fine resolution, the data can monitor the fluid atmospheric activities. In this study, we used the PAR product (details see section 3.5.1), which can freely download from the Japan Aerospace Exploration Agency (JAXA) P-Tree system (https://www.eorc.jaxa.jp/ptree/registration_top.html). The raw H-8 PAR products were

1985), Heliosat (Cano et al. 1986), random forest (Hou et al. 2020), and neural network (Takenaka et al. 2011). The PAR product provided by JAXA was based upon Frouin and Murakami (2007) and may meet our demand. We batch downloaded daily per 10 minutes data during daytime of Taiwan from the beginning of 2018 to the end of 2019. For each image, the PAR pixel values of geographically corresponding meteorological stations were retrieved with Matlab (MATLAB v. R2018b). After that, we can get two years’ time series PAR of four stations. The values at 10:40 local time everyday were continuing missing due to technical problem. The missing value sometimes also occurred randomly.

Because the missing time period was short, we applied liner interpolation to make it a completed two year PAR time series data.

3.3.2. Daytime period setup and Himawari-8 PAR performance

Only daytime period was analyzed because photosynthesis can solely occur during daytime, and fog events may disturb this forest growth cycle. In fact, daytime periods were different between each day especially the pronounced seasonal variance so it was not suitable to define a universal daytime period. For convenience's sake, a universal daytime period was selected first, and then the unique daytime of each day was defined by solar zenith angle lower than 90 degree based on the universal period. The universal period was defined by in-situ to H-8 PAR ratio lay between 0 and 1, which happened mostly during 6 a.m. to 6 p.m. Both in-situ and H-8 PAR values outside this time range tended to close to 0, and made the ratio unexplainable. After that, solar zenith angle with 10 minute frequency everyday was computed:

cos Z = sin ∅∙ sin δ+ cos ∅∙ cos δ∙ cos h ,

∅=latitude, δ=solar declination, h=solar hour angle

(3-1)

δ= sin^-1(0.39785∙ sin ((

t0=12-LC-ET, LC=longitude correction, ET=equation of time (3-4) LC=(longitude-Standard longitude)

In addition, the performance of Himawari-8 was checked before further analysis. In fog-free condition, Himawari-8 PAR value should be close to in-situ PAR value while in the foggy condition, two values would be distinct as a result of sunlight blocking by fog. The relationship of Himawari-8 PAR and in-situ PAR during fog-free and foggy condition were analyzed respectively. Their root mean square error (RMSE) and rRMSE were presented in results section. The RMSE and rRMSE are defined as follows:

𝑅𝑀𝑆𝐸 = √∑^𝑁_𝑖=1(𝑦_𝑖 − 𝑦̂ )_𝑖 ²

3.4. Validation data

There were two time lapse cameras. The first time lapse camera (Brinno TLC200 PRO) was installed at 14.5K station since 2018 for model development and validation, and the second camera (Bushnell Trophy Cam) was mounted at MC station since mid-March 2020 for additional validation. The detailed validated process would be provided in section 3.5.2. The time lapse cameras were set to collect image every 10 minutes during day time only (6 a.m. to 6 p.m.), and can be played in Brinno Video Player software with timestamp at each image bottom. We repeatedly changed SD card and batteries of the camera, and checked the machine condition monthly. To decide if fog occur or not, we visually classified each image into two categories, which was true for fog occur and false for the fog-free condition (Figure 3.2). The standard was that if cloud bottom touch the mountain top of the images, then set as fog event. With this approach, we can get a binary ground truth time series data about fog events for 14.5K and MC stations respectively.

Figure 3.2 Examples of (a) fog-free situation and (b) foggy moment. Following our classified standard (a) would be set to false while (b) would be true.