Temporal and spatial variations of convection, clouds and precipitation over the Tibetan Plateau from recent satellite observations. Part II: Precipitation climatology derived from global precipitation measurement mission

This sequence of papers examines spatio‐temporal variations of precipitation over the Tibetan Plateau (TP) based on satellite observations. Here in Part 2, spatial patterns of seasonal and diurnal variations of precipitation have been examined based on the Global Precipitation Measurement Mission (GPM) and three additional satellite products. The results show a spatial dipole pattern of two distinct seasonalities: The central TP is marked by strong July peaks and exhibits rainfall contributions of the monsoon season (May–September) of more than 70%, whereas northwestern and southern regions of the plateau exhibit significantly smaller amplitudes in the annual cycle. In some southern regions which are characterized by very high summer mean precipitation and more extreme rain rates, winter months (October–April) contribute significantly to the total annual mean precipitation. In addition, there are larger differences in seasonal curves along a west‐to‐east axis, than along a north‐to‐south axis. The spatial patterns of diurnal precipitation over the TP are more complex compared to seasonality and point to multiple components, which construct the regional differences. These show also a seasonal dependence and are characterized by a stronger afternoon to early evening peak (17:00 LST time, 11:00 UTC) and weaker nighttime peak (23:00 LST, 17:00 UTC) during the monsoon season and over the plateau compared to its surroundings. Furthermore, it was shown that convective precipitation during the monsoon season contribute only up to 30% to the total precipitation, whereas more than 70% is produced by the 90th percentile of daily rain rates. An important characteristic of summer precipitation is hence that a significant part of the extreme precipitation is non‐convective. This paper reveals new features of spatial patterns in seasonal and diurnal precipitation and highlights the importance of non‐monsoonal components for seasonal precipitation variations.

This sequence of papers examines spatio-temporal variations of precipitation over the Tibetan Plateau (TP) based on satellite observations.Here in Part 2, spatial patterns of seasonal and diurnal variations of precipitation have been examined based on the Global Precipitation Measurement Mission (GPM) and three additional satellite products.The results show a spatial dipole pattern of two distinct seasonalities: The central TP is marked by strong July peaks and exhibits rainfall contributions of the monsoon season (May -September) of more than 70 %, whereas northwestern and southern regions of the plateau exhibit significantly smaller amplitudes in the annual cycle.In some southern regions which are characterized by very high summer mean precipitation and more extreme rain rates, winter months (October -April) contribute significantly to the total annual mean precipitation.In addition, there are larger differences in seasonal curves along a west-to-east axis, than along a northto-south axis.The spatial patterns of diurnal precipitation over the TP are more complex compared to seasonality and point to multiple components, which construct the regional differences.These show also a seasonal dependence and are characterized by a stronger afternoon to early evening peak (17:00 LST time, 11:00 UTC) and weaker nighttime peak (23:00 LST, 17:00 UTC) during the monsoon season and over the plateau compared to its surroundings.Furthermore, it was shown that convective precipitation during the monsoon season contribute only up to 30 % to the total precipitation, whereas more than 70 % is produced by the 90th percentile of daily rain rates.An important characteristic of summer precipitation is hence that a significant part of the extreme precipitation is non-convective.This paper reveals new features of spatial patterns in seasonal and diurnal precipitation and highlights the importance of non-monsoonal systems (MCS), which can reach horizontal dimensions larger than 300 km (Maddox, 1980;Feng et al., 2014).Because these originate partly from the TP and are partly transported by large-scale circulation systems towards the plateau (Maddox, 1980), MCS exhibit both seasonal (Hu et al., 2016) and regional variations (Hu et al., 2017).The linkages between precipitation, large-scale circulation, topography and local convection are current research questions which need to be explored.
In order to investigate convection dynamics and its importance for precipitation over the TP, it is first important to create a robust climatology of prevailing cloud and precipitation patterns, which reveal general variations of convection in the region.Since cloud microphysics and cloud vertical structure are directly linked to precipitation formation, the purpose of this paper sequence is to simultaneously cast some light on spatio-temporal variations of cloud properties and precipitation and to create a climatology from satellite missions which complement each other in terms of temporal resolution and retrieved parameters.Whereas Part 1 (Kukulies et al., 2019) focused on differences of cloud microand macrophysical properties in three subregions of the TP, which are affected differently by large-scale atmospheric circulation and moisture transport (westerly-dominated domain, transition zone, monsoon-dominated domain), this paper aims to complement the findings by presenting the spatio-temporal variations of precipitation with a focus on regional differences of seasonality and diurnal variations.The reason for analyzing spatial patterns of precipitation at these temporal scales is to elucidate multiple components of precipitation variations.In order to get a better understanding of how these components interact with each other, we aim to identify regional and seasonal dependencies of precipitation characteristics, such as diurnal maximum, convective and extreme precipitation.
Moreover, we focus on satellite retrievals from the Global Precipitation Measurement Mission (GPM), since data products such as GPM IMERG are available at a high spatial and temporal resolution and combined microwave and infrared observations.To put this observations in a perspective of other high resolution multi-sensor satellite products of precipitation, the climatologies are also compared to CMORPH, TRMM (3B42) and MSWEP.
The paper is structured as follows: Section 2 briefly presents the satellite data sets and methods which were used to identify spatial patterns of temporal variations in precipitation.In Section 3, seasonal variations in the different datasets and spatial patterns of seasonality are presented.Section 4 focuses on diurnal precipitation variations and its regional/seasonal dependencies and Section 5 examines additional features of seasonal precipitation, namely extreme and convective precipitation.In Section 6, the findings from this part of our paper sequence are summarized and

Accepted Article
This article is protected by copyright.All rights reserved.
discussed in the context of large-scale atmospheric circulation and convective systems.We also discuss retrieval uncertainties and inherent technical limitations and give recommendations for future studies of precipitation in the TP region based on satellite observations.The key findings and conclusions are summarized in Section 7.

| Satellite data
The Global Precipitation Measurement (GPM) Core Observatory (Hou et al., 2014) consists of a passive and an active sensor which measure hydrometeors in near-real time.It orbits the globe between 65 • S to 65 • N within 93 minutes (on average 16 times per day) and lies in a near-equatorial orbit at an altitude of 407 km.It has thus a revisit time of 1-2 hours and is able to capture diurnal variations of different precipitation types.The passive satellite sensor GPM Microwave Imager (GMI) has an increased number of microwave channels and frequency range compared to its predecessor TMI of the Tropical Rainfall Measurement Mission (TRMM).In addition, the Dual-Frequency Precipitation Radar (DPR) of the GPM Core Observatory operates at two frequencies (Ka-band at 35.33 GHz and Ku-band at 13.6 GHz) and allows for the record of vertical hydrometeor profiles with a large range of particle sizes, since it provides additional information about the droplet size distributions of snow and rain.It has also a larger capacity of detecting snow and light rainfall and is thus more suitable for high latitude and altitude environments in comparison to single-frequency precipitation radars (Kulie and Bennartz, 2009).
The different data products of GPM which have been used in this study are listed in Table 1.GPM IMERG provides a uniform and gauge-calibrated dataset including observations from more than 20 precipitation sensors of TRMM/GPM and comprehensive data bases of microwave brightness temperatures as well as vertical hydrometeor profiles from various passive and active satellite sensors (Huffman et al., 2015b).Due to the improved algorithm for high latitude and altitude regions and the high temporal resolution, GPM IMERG appears to be most promising among the GPM data products.The IMERG algorithm merges precipitation estimates from microwave measurements with microwavecalibrated infrared (IR) observations and rain gauges, in order to produce a global gridded product at a high spatial and temporal resolution.We use here version 05 of the IMERG "final run", which incorporates monthly gauge data and is in

Accepted Article
This article is protected by copyright.All rights reserved.
particular produced for scientific purposes (Huffman et al., 2015b).In addition, we used the daily gridded dataset GPM 3GPROF, which is based on DPR and provides convective vs. stratiform precipitation fractions.
For Mainland China, different satellite rainfall products have been evaluated (Zhou et al., 2008;Tang et al., 2016), whereby GPM IMERG reproduces probability density functions and diurnal cycles of precipitation which are closest to station measurements (Tang et al., 2016).Moreover, Xu et al. (2017a) showed that GPM IMERG reduces the wellknown wet bias in precipitation observations over the TP and outperforms the Tropical Rainfall Measurement Mission (TRMM) for this specific region.However, because GPM data is only available from 2014 and covers hence a short time period, GPM IMERG was compared to three other global high-resolution precipitation products, which also merge measurements from multiple satellite sensors, but use different algorithms for the final estimates (Table 1): TRMM (3B42), CMORPH (CRT) and MSWEP (Multi-Source Weighted-Ensemble Precipitation) .All used datasets were chosen at the highest spatial and temporal resolutions available, in order to get an accurated representation of diurnal and regional variations.
While TRMM and CMORPH have been used in several studies of precipitation over the TP (Gao and Liu, 2013;Maussion et al., 2014;Tang et al., 2016;Xu et al., 2017b), the recently created global dataset MSWEP is used for the first time for the TP region.In this datasets, optimal precipitation estimates are created based on multiple satellite observations combined with both ground-based measurements and reanalysis data, whereby the different data sources are merged based on weight maps.For example, in a region like the TP, where gauge observations are sparse, more weight is given to satellite observations and reanalysis data.It should be noted, that the three datasets used are not completely independent: Even though these are constructed from different retrieval algorithms and data assimilation systems, they also share information: MSWEP includes, for instance, TRMM and CMORPH in the final precipitation estimate.

| Precipitation variations in the study domain
Figure 1a displays the DEM of the TP domain.In this study, the TP domain is constraint to all locations (pixels) within this domain above 3000 m ASL.It is also the domain chosen for the regionally averaged metrics describing precipitation variations, in order to focus on the high elevation region of the TP and to exclude regions south of the plateau, because

Accepted Article
This article is protected by copyright.All rights reserved.In Part 1 of this paper sequence (Kukulies et al., 2019), variations of cloud characteristics based on cloud profiles were compared between three subregions of the TP, which are affected differently by large-scale atmospheric circulation.Here, we chose to examine the spatial patterns of seasonality, diurnal variations and extreme/convective precipitation over the entire TP instead of regions with predefined boundaries, since the high-resolution gridded data allows (in contrast to the cloud profiles) for a more accurate identification of regional differences.
In order to detect new regional patterns than the known north-south differences (Maussion et al., 2014;Yao et al., 2013), we performed a principal component analysis (PCA) on both seasonal and diurnal variations of precipitation over the TP.To identify the spatial patterns of temporal variations rather than actual variations in precipitation intensity, precipitation rates were first normalized by calculating the monthly (hourly) contributions in % to total annual (daily) precipitation.The input feature space for the PCA, was thus the normalized seasonal (diurnal) vector for each grid cell in the domain displayed in Figure 1a (n grid cells x 12 monthly/24 hourly precipitation contributions).The corresponding

Accepted Article
This article is protected by copyright.All rights reserved.
output are the eigenvector strengths for each grid cell, which indicate the deviation from the mean conditions of the dataset and can be used to reconstruct the first and second principal components of the seasonal and diurnal vectors.This is done by adding the eigenvector strengths to the per feature mean (seasonal/diurnal vector) and multiply it by the principal axis in feature space for the highest (first principal component) and second highest (second principal component) explained variance.This allows for the identification of regions or clusters of locations with distinct principal components for seasonal variations of precipitation (e.g.due to different large-scale atmospheric circulation impacts).

| SEASONALITY
The monthly contributions to the total annual mean precipitation for the four satellite data products averaged over the TP domain are displayed in Figure 2. It can be seen that the seasonal curves exhibit the same shape and similar monthly fractions in all data sets.This shows the robustness of the facts that the regional precipitation over the TP has a distinct peak with the highest contributions from June -August and that the monsoon season from May -September contributes to more than 70 % to the total annual precipitation.GPM IMERG exhibits the highest contribution during the monsoon season (80 %) despite the fact that mean rain rates are lower than in the other datasets (Fig. 3).
The average precipitation rates (mm/day) during the monsoon season May -September show also similar spatial patterns in the four datasets (Fig. 3).There is a distinct northwest-southeast diagonale with the highest mean precipitation rates occurring in the southeastern TP.Outside the 3000 m boundary, the most extreme precipitation occurs south of the Himalayas.In addition, there is a region in the northwest outside of the plateau, which is marked by higher precipitation rates compared to the surroundings, which is clearly visible in all datasets (Fig. 3).In general, MSWEP has the latter two features most pronounced and the highest precipitation rates over the plateau.This might be due to the incorporation of reanalysis data, which has a known wet bias for the high altitude region of the TP (Gao et al., 2015;Lin et al., 2018).In addition, MSWEP is the only dataset that has daily rain rates during May -September exceeding 10 mm/day within the 3000 m boundary of the plateau (Fig. 3).
Even though it has long been recognized that the monsoon season dominates the temporal variations over the plateau (which is also clearly visible in Figure 2), the regional differences of variations in the seasonal cycle are not

Accepted Article
This article is protected by copyright.All rights reserved.
Contribution to total annual precip (%) GPM 2014-2017TRMM 2014-2017MSWEP 2014-2016CMORPH 2014-2017 F I G U R E 2 Monthly contributions to total annual mean precipitation (%) for GPM, CMORPH, TRMM and MSWEP.The seasonal curve shows the average over the TP domain for the years 2014 -2017 of GPM, CMORPH and TRMM and 2014 -2016 for MSWEP.
established.Figure 4 shows that some of the mountain ranges stretching from the northwest to southwest receive a significant part of the precipitation between October and April, when mid-latitude westerlies rather than the monsoon circulation dominates the atmospheric water vapor transport (Yao et al., 2013;Liu et al., 2016).This region, where the contribution of summer precipitation between May and September only contributes with less than 30 % to the total annual precipitation, accords very well among TRMM, MSWEP and GPM (Fig. 4).In the CMORPH data, the pattern looks slightly different and the lower summer precipitation contributions are rather constraint to the northeast of the northwestern mountain ranges.To get clearer picture of the regions within the TP in which the seasonal cycle of precipitation is controlled by other factors than the monsoon circulation, we examined the first two principal components of the normalized seasonality (Fig. 5, 6).
Figure 5 shows the sign and strengths of the first principal component (PC1), which explains 52.8 % of the variance in the GPM data.In other words, the colors indicate how much the seasonal curves in each grid cell deviate from the

Accepted Article
This article is protected by copyright.All rights reserved.regionally averaged seasonal precipitation.The red regions are characterized by a distinct summer precipitation peak, whereas the blue regions have significant monthly contributions to annual precipitation outside of the monsoon season.
In order to demonstrate the main features of seasonal precipitation in the two opposing regions, the seasonal curves at two locations (A and B) with extreme eigenvector strength values are shown in comparison to the part of the curves, which are explained by PC1 at the respective location (Fig. 5).Locations A and B are chosen from the 90th percentile of positive (blue) and negative (red) eigenvector strength values in the plateau region.The black line in Figure 5 shows the regionally averaged curve over the entire domain (TP + surroundings), to visualize how much the seasonal curves at the example locations A and B (dashed lines) can deviate from regionally averaged seasonal curve.The solid lines display the two contrasting features of the PC1 for blue (example point A) and red (example point B) regions: The PC1 for regions with strong negative eigenvector strength is marked by a distinct summer peak and the PC1 of regions with strong positive eigenvector strength values is marked by a flatter seasonal curve.The points reveal, hence, an interesting feature of the southern parts of the TP: The "blue" pattern (which represents a higher contribution during the winter

Accepted Article
This article is protected by copyright.All rights reserved.months compared to the red regions) expands along the southern mountain ranges towards the southeastern edges of the TP.This means that the seasonal variation in some parts of the southeastern strongly resembles the northwest outside of the plateau.
Comparing Figure 5 and 3, it becomes clear that the blue regions are marked by relatively high precipitation rates during summer.Hence, the flatter seasonal curves do not indicate the absence or smaller impact of the monsoon circulation, but rather the additional influence of other large-scale atmospheric circulation systems.Knowing that midlatitude westerlies arrive from the northwest (here characterized by strong positive eigenvector strengths) and transport moisture to the plateau between October and April, the dipole pattern in Figure 5 suggests the far-reaching impact of westerlies.
While the PC1 explains the regional differences in seasonality between the major parts of the central plateau compared to the surrounding region and edges of the plateau, the second principal component (PC2), which explains 21.8 % of the variance in the data, clearly divides the plateau into a western and an eastern region (Fig. 6).The blue

Accepted Article
This article is protected by copyright.All rights reserved.and red regions differ mainly due to a shift from a PC2 precipitation peak in July (blue) to a PC2 precipitation peak in August (red).To exemplify how seasonal curves in dark-blue and dark-red colored regions can look like, we show again the seasonal curves and their corresponding PC2 at two locations A and B, which represent extreme positive (blue) and negative (red) eigenvector strength values.
The PC1 of the annual cycle revealed the feature of the seasonal cycle that explained most of the variance.This feature could, hence, be attributed to the large-scale atmospheric circulation, which has the most important impact on the shape of the seasonal curve.The physical interpretation of the second principal component is less straight-forward.
However, we can see a distinct east-west separation and the earlier summer precipitation peak in the west compared to the east could indicate the impact of the East Asian summer monsoon, which typically has a later onset than the Indian

Accepted Article
This article is protected by copyright.All rights reserved.
summer monsoon arriving from the Bay of Bengal (Yihui and Chan, 2005).Another possible mechanism for the shifted summer precipitation peak is the formation of MCS in the eastern TP due to a east-west heating contrast (Sugimoto and Ueno, 2010), since MCS also can contribute to large portions of summer precipitation (Hu et al., 2016).Summarizing the results from the PCA, the first two principal components of the seasonal cycle together explain ca.75 % of the variance in the data and show clear dipole patterns even within the higher elevations of the TP (e.g.differences along the southern mountain range).The two distinct spatial clusters revealed by the first principal component are regions with a clear summer peak contrasted by regions with more moderate seasonal variations.
Another important finding, which expands previous knowledge about variations in seasonal precipitation over the TP, is

Accepted Article
This article is protected by copyright.All rights reserved.
that the west-to-east differences which are shown by both the first and the second principal components seem to be more dominant compared to the previously recognized north-to-south differences.This indicate more complex patterns of the impact of large-scale atmospheric circulation on seasonal precipitation than simple north-to-south regimes.

Accepted Article
This article is protected by copyright.All rights reserved.
(May -September and October -April) based on 30 min data for GPM and CMORPH (Fig. 7a, 7b) and 3 hourly data for TRMM and MSWEP (Fig. 7c, 7d).In order to highlight the role of elevation for the diurnal cycle, the dashed lines display the diurnal cycles for the TP (27 -45 • N, 70 -105 • E) when the lower elevated surroundings are included (TP + surrounding).All datasets show smaller amplitudes in the diurnal curves when lower elevated surrounding regions are included and larger precipitation contributions during nighttime hours.CMORPH exhibits earlier diurnal precipitation peaks compared to the other datasets, whereby the diurnal curves for the two seasons show also remarkable differences (Fig. 7b).While GPM, MSWEP and TRMM show a clear evening peak during both seasons (around 17:00 local solar time, 11:00 UTC), CMORPH shows a noon and afternoon peak from October to April and an afternoon as well as an evening peak (15:00 and 20:00 local solar time, 09:00 and 14:00 UTC) from May to September.All datasets agree, however, on the clear valley around 12:00 local solar time (04:00 UTC), except for CMORPH during the dry season (Fig. 7b).
While the regionally averaged curves in Figure 7 do not reveal any significant seasonal differences (except for CMORPH), the more detailed spatial patterns of diurnal maximum precipitation show clear differences between the wet and dry season (Figure 8).During the wet season (May -September), the western plateau is dominated by maximum precipitation in the afternoon to early evening (15:00 -18:00 local solar time), while the eastern plateau exhibits earlier afternoon precipitation maxima between 12:00 and 15:00 local solar time.Maximum precipitation in the time slots 15:00 -18:00 and 21:00 -00:00 local solar time occur at many locations of the plateau during this season, as was depicted in Figure 7a.Lower elevated terrains, e.g. at the leeside of the Himalayas, are marked by maximum precipitation during nighttime in the wet season (Fig. 8).During the dry season (Oct -Apr) the spatial distribution of diurnal precipitation maxima shows a generally more patchy pattern, so that it is difficult to distinguish any larger regions which is marked by the same maxima.Only the northwestern plateau stands out by shifting the early eveningevening precipitation peak to a peak in the early morning with maximum precipitation between 06:00 and 09:00 local solar time (Fig. 8).
The differences between the wet and dry season in Figure 8) show clearly that spatial patterns of diurnal variations are seasonally dependent, even if the regionally averaged diurnal curves of the two seasons look similar (Fig. 7a).This indicates that seasonal, diurnal and regional precipitation patterns are not strictly independent from each other due to interactions of precipitation mechanisms at different scales.For instance, the fact that we can distinguish regional patterns of maximum precipitation in the wet season demonstrates that the monsoon circulation is an important

Accepted Article
This article is protected by copyright.All rights reserved.To further investigate systematic regional differences of diurnal precipitation, we also performed a PCA on hourly contributions to the total daily precipitation amount (Fig. 9).The first principal component of diurnal cycles over the TP can only explain 34.4 % of the GPM data, which indicates that diurnal variations over the TP have in general many additional components.This underlines the importance of smaller scale precipitation processes for the diurnal cycle in complex mountain terrains.
The pronounced dipole pattern which is visualized by the red vs. blue regions, shows that the TP within the 3000 m boundary distinguishes from the lower elevated surrounding regions (Fig. 9).The two distinct regional clusters become

Accepted Article
This article is protected by copyright.All rights reserved.
even more apparent in Figure 10 which shows the histograms of the eigenvector values for PC1 of diurnal variations for TP + surroundings (Fig. 10a) compared to TP (Fig. 10b).While the entire domain including surrounding regions at lower level elevations exhibits a distinct bimodal distribution, the TP (at elevations above 3000 m ASL) is marked by a unimodal distribution for PC1 eigenvector values.This means that clear clusters of diurnal variations are only visible when the lower elevated surrounding regions are included, which again underlines the dependence of precipitation patterns on elevation and topography.
In Figure 9, we show the diurnal cycles and their corresponding first principle components of two example points with strong eigenvector strength values of opposite sign (same as Fig. 5 and 6).It can be seen that PC1 at location A manifests an afternoon-early evening peak (17:00 local solar time, 11:00 UTC), which is clearly higher than the average diurnal curve of the TP + surrounding domain.At point B, which is used as an extreme example for dark red colored regions, the diurnal curve deviates even more from the regional average.Here, the diurnal cycle is characterized by a late evening (06:00 local solar time, 00:00 UTC) and early morning peak (23:00 local solar time, 17:00 UTC).

| EXTREME PRECIPITATION
Having identified major spatio-temporal features of mean rain rates, it should be noted that these patterns do not necessarily coincide with patterns of heavy rain rates.For example, the southern regions of the TP are strongly bluecolored in Figure 5 and exhibit therefore a rather flat seasonal curve even though the region is marked by high mean precipitation rates (Fig. 3).That means that regions with similar mean rain rates during summer (Fig. 3) but seasonal curves with a pronounced summer peak (Fig. 5 ) experience more extreme rain rates during only a few months of the year (mostly between May and September).
Figure 11 shows the spatial distribution of the 90 t h percentile of daily summer precipitation.The figure reveals that over the TP, indeed, the patterns are quite similar to total mean summer precipitation (Fig. 3).Interestingly, the eastern parts of the southern mountain ranges (around 95 • E), which are marked by significant precipitation contributions outside of the monsoon season (Fig. 5), also are the regions which experience the most extreme rain rates during summer (Fig. 11).

Accepted Article
This article is protected by copyright.All rights reserved.Furthermore, the monthly total precipitation from daily rain rates above the regionally averaged 90 t h percentile to total monthly precipitation exhibits the same shape as monthly contributions to total annual mean (Fig. 12a).This underlines firstly that most of the extreme precipitation occurs between May and September and secondly that most of the summer precipitation is produced by extreme precipitation.The total amount of precipitation produced by 90th percentile rain rates during the summer months (May -Sep) stands for around 80 % of the total summer precipitation in all four datasets (Fig. 12b).

Accepted Article
This article is protected by copyright.All rights reserved.

| CONVECTIVE PRECIPITATION
Figure 13 shows the seasonal contributions of convective precipitation to the total precipitation based on DPR in the high resolution 3GPROF product for 12 months from 2014 to 2015.Whereas only a small part of precipitating

Accepted Article
This article is protected by copyright.All rights reserved.F I G U R E 1 2 Extreme precipitation in GPM (2014GPM ( -2017)), TRMM (2014TRMM ( -2017) ) and MSWEP (2014-2016) and CMORPH (2014-2017).pixels (0 to 10 % for most of the grid cells) are classified as convective between October and April, the contribution of convective precipitation varies between 10 and 40 % in the eastern TP between May and September.The annual cycle of convective precipitation follows the same curve as for total precipitation (Fig. 2) and shows a pronounced peak during the summer months with more than 80 % of convective precipitation occurring between May and September (Fig. 14).The seasonal difference indicates that convective precipitation is closely related to the monsoon circulation and is even more pronounced south of the TP, where the contributions of convective precipitation are between 60 and 80 %.This is understandable as development of convection also needs moisture which can be provided by the monsoon winds.The results verify that most of the convective precipitation occurs south of the Himalayas, but the north to south gradient suggested in (Maussion et al., 2014) is less visible and instead the west to east difference seem to be more pronounced in 3GPROF product.Further, this previous study showed that between 70 and 80 % of the precipitation is of convective type in the central TP and up to 100 % in the Himalayas (Maussion et al., 2014).The contribution of convective precipitation based on 3GPROF is hence significantly lower than the previous study (Fig. 13, 14), which suggests that the role of convection (in both the westerly and monsoon season) for precipitation needs further consideration.

Accepted Article
This article is protected by copyright.All rights reserved.

| Precipitation and convection
Referring back to the key findings of Part 1 of this paper sequence (Kukulies et al., 2019), the GPM-derived precipitation climatology confirms that stratiform cloud systems and stratiform precipitation occur frequently over the plateau.This was shown by the higher frequency of stratiform and ice cloud layer types and the radar reflectivity profiles in Part 1 and is here visible through relatively low contribution of convective precipitation all year round (Fig. 13).Compared

Accepted Article
This article is protected by copyright.All rights reserved.to insights about convective precipitation from the downscaled WRF product High Asia Reanalysis (HAR) (Maussion et al., 2014), the contribution of convective precipitation estimated by GPM is significantly lower (Maussion et al., 2014).However, it should also be noted that convection can occur at different scales and that the satellite convective precipitation fraction relies primarily on the horizontal and vertical structure of the radar signal (Kummerow et al., 2001), whereas the convective precipitation in HAR data refers to the precipitation fraction produced by the cumulus parametrization scheme.
Since a significant part of precipitation-producing MCS occurs during the monsoon season (Maddox, 1980;Hu et al., 2016Hu et al., , 2017)), this suggests that the convective precipitation is probably partly associated with these MCS, but that the accompanied stratiform precipitation in MCS stands for much larger portion of precipitation.This is also reinforced by the fact that a large part of extreme precipitation during summer seems to be non-convective (Sections 5 and 6).
Further, the nighttime precipitation peaks south of the Himalayas suggest clearly that MCS can significantly effect

Accepted Article
This article is protected by copyright.All rights reserved.
the diurnal cycle of precipitation, since such systems produce usually most of the precipitation at night as they mature (Li et al., 2008;Xu and Zipser, 2011).It is not trivial to separate effects of topography on the incoming monsoon flow from heating-driven MCS south of the Himalayas, which both represent mechanisms that could affect diurnal precipitation.
However, MCS are an example for how seasonally dependent weather systems (Hu et al., 2016) also interact with local processes (such as topography-forced upslope winds) and thereby affect diurnal cycles of precipitation.The different spatial patterns of diurnal maximum precipitation in the different seasons (Fig. 7, 8) suggest that regional, seasonal and diurnal controls of precipitation interact with each other and are all important components of precipitation.
The GPM data indicates that precipitation occurrences between October and April can, despite their generally very low rain rates, significantly contribute to the total precipitation occurring along a stretch from northwest to the southern mountain ranges of the plateau.The analysis of the first two principal components of seasonal precipitation suggest that non-monsoonal contributions significantly affect precipitation along the southern and northwestern edges of the TP.In Part 1 of this paper sequence, it was found that the large-scale moisture transport by mid-latitude westerlies does not only control convection regimes in the northern TP domain but also in the central regions of the transition zone (Kukulies et al., 2019).The precipitation climatology in this study also shows clearly that seasonal precipitation variations are not constraint to the north-south regionalization.This is consistent with Curio and Scherer (2016) who investigated precipitation dynamics over the TP based on HAR and found that despite the regional and seasonal variations of precipitation-producing mechanisms, the westerly circulation affects precipitation over the entire plateau and is not only limited to the northern region.
The PCA of diurnal precipitation revealed that the diurnal cycle has multiple components and less pronounced spatial patterns (or weaker explained variance of the principal components) compared to seasonal precipitation.This emphasizes the complexity of diurnal cycles of precipitation, since it can be caused by interaction of processes at many different scales.For example, elevation and topography can affect precipitation at different scales from more local mountain-plain solenoids (Zhang et al., 2014) to large-scale elevation-dependence of precipitation (Li et al., 2017).
However, it should be noted that elevation-dependence is mainly controlled by two limiting factors: decreasing water vapor supply with elevation and increasing instability (Cuo and Zhang, 2017).Over the TP, the increasing instability has lead to a general increase in summer precipitation with elevation up to ca. 4000 m ASL, since enough water vapor can still be supplied in many regions (Cuo and Zhang, 2017).At higher elevations, precipitation amounts decrease due to an

Accepted Article
This article is protected by copyright.All rights reserved.
increasing lifting condensation level and simultaneously decreasing humidity (Cuo and Zhang, 2017).On a diurnal scale, elevation effects are more local and complex in the way that no larger-scale spatial patterns are dominant within the 3000 m boundary of the TP.Diurnal variations need hence to studied more detailed under the consideration of local topography.

| Satellite data
Inter-comparison studies have shown that satellite observations of precipitation over the TP can exhibit large differences (You et al., 2015;Wang et al., 2017).This pertains in particular for precipitation parameters such as precipitation phase, which are highly dependent on retrieval algorithms (Iguchi et al., 2010).Chen et al. ( 2016) quantified spatial scales of precipitation over the TP based on long-term satellite observations.They found that spatial coherences of precipitation show a strong dependence on directions, with the strongest correlations occurring along west-east.
Further, the representativeness of precipitation at any given point typically decreased exponentially with distance and exhibited spatial scales between 111 and 182 km, which is only a third of the spatial scale of annual surface temperatures (Chen et al., 2016).
Since the TP represents a high altitude environment with very high variations of elevations over small distances (Fig. 1b), generating distinct heating profiles and pressure fields, general assumptions of global satellite products like GPM (Elsaesser and Kummerow, 2015) do not accurately represent the specific atmospheric conditions over the TP.
We aim therefore to give a short discussion on technical issues and recommendations for future studies of the TP based on satellite observations of precipitation and convection.
Because of the low temperatures, an important capability of satellite sensors used for TP studies is the detection of snow and solid precipitation.Compared to the CloudSat sensor CPR, the GPM sensor DPR has a lower sensitivity for frozen hydrometeors and a noise reducing algorithm used for DPR might in some cases remove the radar signals from snow (Casella et al., 2017).Whereas DPR alone captures approximately 30 % of the total snow mass (Casella et al., 2017), the aim of the GPM IMERG product is to improve the robustness for light rain and solid precipitation in mid-latitude regions by using multiple sensors with complementary skills (Huffman et al., 2015a).However, the missing snow detection remains an issue despite improvements and could explain the very low precipitation amounts which

Accepted Article
This article is protected by copyright.All rights reserved.
have been observed during the winter season in this study.
Retrieval algorithms for precipitation from passive microwave sensors (e.g.GMI) are highly dependent on scattering from ice particles (Xu et al., 2017b), which generally is an advantage for the TP, because it has high occurrence frequencies of ice and mixed-phase clouds (Kukulies et al., 2019).However, the retrieval of precipitation type is more uncertain.Kirstetter et al. (2018) showed, for example, that GPM IMERG is able to capture the diurnal variability of precipitation at a regional scale, but highlight the concern that orographic and convective precipitation might be underestimated in high altitude regions.One reason for this misidentification has already been found in TRMM precipitation retrievals, where ground echo could be misidentified as melting layer of a stratiform cloud system, because the freezing level is close to the surface over the TP (Fu and Liu, 2007).It is therefore important to adequately address convective and stratiform precipitation over the TP in future studies, in order to determine whether the low contributions of convective precipitation in this study are due to technical errors or an indicator for other processes being more important for precipitation formation.
Compared to the Tropical Rainfall Measurement Mission (TRMM), GPM-retrieved precipitation rates tend to be closer to meteorological observations at ground-based stations of the TP (Xu et al., 2017b) and in contrast to the wet bias in regional climate models and reanalysis data for the TP (Su et al., 2013;Gao et al., 2015;Wang et al., 2017) the main limitation of GPM is its larger missing ratio, especially for snow.Both the diurnal variations and total precipitation associated with the westerly moisture transport to the plateau might therefore be higher as presented here.
It should also be noted that the published GPM data products are only available for a short time period, in particular for the high resolution data of GPROF (Table 1).In this study, it is assumed that the 3-years period from 2014 to 2017 can capture some of the general precipitation patterns (seasonal and diurnal variations) due to the high resolution.
GPM IMERG does not have the capacity yet to accurately represent inter-annual variability, but the IMERG retrieval algorithms is currently under development and aims to extend its time period back to the start of the TRMM era in 1998 (Huffman et al., 2015b).This prolonged time period will be object of more detailed future studies of precipitation over the TP.
Due to the inherent limitations of the satellite retrievals over the TP, we aim and recommend to improve retrieval algorithms for precipitation based on the ground radar stations over the TP (e.g. using ground-radar measurements as training data in machine learning techniques), in order to account for the specific conditions of the region and minimize

Accepted Article
This article is protected by copyright.All rights reserved.
technical uncertainties.Ma et al. (2018) developed and evaluated a multi-satellite dataset created for precipitation retrievals for the specific region of the TP, by including the two widely used satellite precipitation datasets CMORPH and PERSIANN.However, this dataset has not been compared to any ground measurements yet, which underlines the need for the incorporation of ground-based observations.To pursue further studies of convective processes in the TP region and make best use of modern satellite observations, we recommend the use of ground-radars located in the TP for validation and improvements of cloud and precipitation retrievals from satellite observations.We summarize therefore the most crucial aspects, which need to be addressed in future precipitation retrievals based on multi-satellites over the TP: (a) collocations to increase resolution and to reduce uncertainty, (b) accurate temperature profiles (e.g. for retrievals of cloud phase), (c) the reduction of ground clutter to improve the identification of stratiform and convective precipitation (d) and the calibration based on ground radars.

| CONCLUSIONS
This study investigated spatio-temporal variations of precipitation over the Tibetan Plateau (TP) based on the highresolution satellite dataset GPM IMERG as well as three additional multi-sensor satellite products.The monthly contributions of total and extreme precipitation as well as the spatial patterns of summer mean and extreme precipitation mostly agree between the datasets.Diurnal cycles show some differences, but a clear valley around noon was visible in all datasets.
We used GPM IMERG to perform principal component analyses on normalized diurnal and seasonal precipitation over the TP.
The three key features of this analysis are: 1.The first principal component of seasonal variations shows a clear spatial dipole pattern for two distinct seasonalities: Major parts of the central TP are clearly marked by a seasonal curve which peaks in July and exhibits contributions of the monsoon season (May -Sep) of more than 70 % to the total annual precipitation, whereas seasonal curves along the northwestern and southern edges of the plateau resemble those in northwestern regions

Accepted Article
This article is protected by copyright.All rights reserved.
outside of the plateau.Even in some southern regions which are characterized by very high mean summer precipitation and more extreme rain rates, winter months (Oct -Apr) contribute significantly to the total annual mean precipitation.Non-monsoonal components play, hence, an important even in some southern regions of the TP.

2.
The second principal component of seasonal precipitation reveals a west-east pattern, which emphasizes that large-scale differences in elevation affect seasonal precipitation.Since neither the first nor the second component shows a clear north-south pattern, we conclude that seasonal variations show larger regional differences on a west-to-east-axis.

3.
Compared to spatial variations of seasonality, the spatial patterns of diurnal precipitation over the TP are more complex and point to multiple components which construct the regional differences.However, the TP stands out compared to its surrounding lower elevated areas with a strong early evening peak (17:00 local time, 11:00 UTC) and a weaker nighttime peak (23:00 local time, 17:00 UTC).This latter night time peak is slightly more pronounced at elevations below 3000 m ASL, where diurnal variations are in general smaller compared to the TP.
To derive additional properties of dominant precipitation patterns, we also examined seasonal variations of convective and extreme precipitation.The results reveal that relative contributions of convective and extreme precipitation (90 t h percentile of daily precipitation) peak in July and follow the same seasonal curve as for total precipitation.However, convective precipitation over the TP contributes only up to 30 % to the total summer precipitation, whereas more than 70 % of the total summer precipitation amount is produced by the 90 t h percentile of daily rain rates.Hence, it could be concluded that a significant part of extreme precipitation are of non-convective type.
The two main questions, which have to be addressed in further precipitation studies on the TP are hence (a) the role of convection and (b) the role of the westerly circulation and other synoptic systems which affect the TP between October and April.Due to the inherent limitations of satellite observations over the TP, we also conclude that cloud and precipitation retrievals for the specific environment of the TP need to be validated and improved based on ground radar measurements, in order to create a more robust climatology and to reliably relate the latter to convective processes.

Accepted Article
This article is protected by copyright.All rights reserved.

Accepted Article
This article is protected by copyright.All rights reserved.
(a) DEM of the TP and surrounding region (b) Variation of elevation over the TP F I G U R E 1 Topography of the Tibetan Plateau (TP) based on 30 m resolution data from the Shuttle Radar Topography Mission (SRTM).
Figure (a) displays a DEM of the TP and surrounding region.Locations within the domain 27 -45 • N, 70 -105 • E and above 3000 m ASL are in this study defined as the TP. Figure (b) shows the standard deviation of elevation in every 3 km x 3 km grid.these are strongly affected by the Indian summer monsoon and would therefore dominate the signals for the entire domain (cf.sec.3).We refer to this domain (27 -45 • N, 70 -105 • E) as the "TP" when only elevations above 3000 m ASL are included and to "TP + surrounding" when the surrounding lower elevations within the domain are included.

Figure
Figure 1b highlights the strong topographic variation which can occur over small areas over the TP and especially in the northwest and along the steep slopes of the southern foothills (up to 1400 m standard deviation within 3 km grid cells).

F
I G U R E 4 Contributions of summer precipitation (May -Sep) to total annual mean precipitation (%) for MSWEP, CMORPH, TRMM and GPM.The seasonal curve shows the average over the TP domain for the years 2014 -2017 of GPM, CMORPH and TRMM and 2016 -2016 for MSWEP.

F
I G U R E 5 First principal component (PC1) of the normalized annual cycle over the TP, which explains 52.8 % of the variance in the GPM data.The colors red and blue indicate the sign and strength of the eigenvectors of PC1.The PC1 of variations in monthly contributions to total annual mean precipitation (2014 -2017) are shown for two example locations A and B (solid lines), which are marked by different PC1 of seasonality.In addition, the actual seasonal curves at the respective locations (dashed lines) are displayed in comparison to the regionally averaged seasonal curve for the TP domain (black line).

6
Second principal component (PC2) of the normalized annual cycle over the TP, which explains 21.8 % of the variance in the GPM data.The colors red and blue indicate the sign and strength of the eigenvectors of PC1.The PC1 of variations in monthly contributions to total annual mean precipitation (2014 -2017) are shown for two example locations A and B (solid lines), which are marked by different PC1 of seasonality.In addition, the actual seasonal curves at the respective locations (dashed lines) are displayed in comparison to the regionally averaged seasonal curve for the TP domain (black line).
Diurnal cycle of surface precipitation over the TP for a) GPM (2014 -2017), b) CMORPH (2014 -2017), c) TRMM (2014 -2017) and d) MSWEP (2014 -2016).The relative contributions to the total daily precipitation amount in % are based on 30 min data for GPM and CMORPH and 3 hourly data for TRMM and MSWEP.The dashed lines display the diurnal curves averaged over the entire domain (27 -45 • N, 70 -105 • E) including surrounding regions at lower level elevations.The curves are further divided into two seasons: May -September (red) and October -April (blue).

Figure 7
Figure 7 displays the normalized diurnal cycles of GPM, CMORPH, TRMM and MSWEP separated into two seasons systems, which form from intensive surface heating, also have a typical diurnal cycle(Li et al., 2008) and could thereby affect diurnal precipitation variations at a larger scale.Diurnal variations during the dry season are, by contrast, more controlled by local processes due to the absence or synoptic and meso-scale forcings.F I G U R E 8 Local solar time (UTC + 6) of diurnal maximum precipitation over the TP based on GPM IMERG (2014 -2017) for wet (May -Sep ) and dry (Oct -Apr) season.

F
I G U R E 9 First principal component (PC1) of the normalized diurnal cycle over the TP, which explains 34.4 % of the variance in the GPM data.The colors red and blue indicate the sign and strength of the eigenvectors of PC1.The PC1 of variations in hourly contributions to total daily mean precipitation (2014 -2017) are shown for two example locations A and B (solid lines), which are marked by different PC1 of diurnal variations.In addition, the actual seasonal curves at the respective locations (dashed lines) are displayed in comparison to the regionally averaged seasonal curve for the TP domain (black line).
Histograms of the first principal component of diurnal variations for (a) the TP and the lower elevated surrounding regions and (b) the TP domain.F I G U R E 1 1 90 t h percentile of daily rain rates during May -Sep for MSWEP, CMORPH, TRMM and GPM.The 90 t h percentile has been calculated for the years 2014 -2017 of GPM, CMORPH and TRMM and 2014 -2016 for MSWEP.
. to total precipitation May -Sep (%) (b) Contribution of total 90 t h percentile daily precipitation to total summer precipitation (May -September).

F
I G U R E 1 3 Spatial patterns of contribution of convective precipitation (%) based on GPM 3GPROF (2014 -2015) divided into two seasons (May -Sep vs. Oct -Apr).