1. Introduction
Daily rainfall is a critical variable used to characterize hydrological dynamics in tropical catchments. In Southeast Asia, many operational and productive micro-watersheds are relatively small (<16 km²) [
1,
2,
3], but are utilized for water supply, irrigation, and hydroelectric power. Despite their small size, these micro-watersheds exhibit high spatiotemporal variability in rainfall patterns [
4,
5], including extreme rainfall events that release large volumes of water within a short duration [
6,
7]. Rainfall intensity and distribution vary at short distances in micro-watersheds (~1 km). Additionally, coastal tropical watersheds are influenced by convective processes that determine cloud formation, as well as interactions between seasonal winds from the Asian Monsoon and local topography and geomorphology.
Most micro-watersheds in Southeast Asia have rainfall data conflicts, which are characterized by one or more of the following problems: lack of rain gauges, sparse rain gauge coverage, missing rain gauge data, inefficient data sharing policies, or ineffective data management [
8]. These conflicts often cause ineffective watershed management, which can lead in turn to water resource, flood, water-borne disaster, and biodiversity management failures. Satellite precipitation data can be used as a support for ungauged sub-catchments or those with data conflicts; however, their use is constrained by spatial resolution and sub-catchment size limitations. The coarse grid size of satellite data is a major constraint, whereby satellite precipitation data are unable to effectively represent spatial rainfall distribution at the sub-catchment scale [
9,
10,
11].
The most advanced satellite data product of the Global Precipitation Mission (GPM) is available at a half-hourly scale with an optimal resolution of 0.1° (~100 km² grid at the Equator) [
12]; however, it cannot represent the rainfall patterns of small tropical sub-catchments, known as micro-watersheds [
13]. Despite the recent development of many spatial downscaling techniques, most of these techniques are designed for monthly-scale [
8,
14,
15,
16] or larger measurements (seasonal, annual) [
17,
18]. Thus, most downscaling approaches are limited to monthly rainfall downscaling and cannot be applied to daily-scale rainfall data. These spatial downscaling techniques are based on empirical relationships with surrogate environmental variables as downscaling proxies; they have been developed for non-humid tropical climates in terms of orography, solstice equinox season, and vegetation [
19,
20] and are therefore not appropriate for the humid tropics. Although one technique has been developed for the humid tropics, its selected proxy variable is less meaningful for coastal and maritime regions [
21].
Therefore, an appropriate approach is needed for downscaling satellite precipitation data to produce improved daily rainfall data for humid tropical environments, independent of rain gauges. Daily downscaling approaches by Ryo [
22] that use in-sync calibration using rain gauge streamflow or merging with in situ rain gauge data by Chen et al. [
23] and Lopez et al. [
24] have produced reliable results. Unfortunately, both of these methods require in situ rain gauge measurements, and many micro-watershed catchments have no or very limited in situ data. Therefore, the challenge in spatially downscaling daily-scale satellite precipitation data without in situ rain gauge measurement comprises the identification of environmental parameters that strongly characterize rainfall patterns at the hourly local scale, for use as proxies. Most torrential rainfall in the humid tropics of Southeast Asia is produced by rain events lasting 2–4 h. At this scale, local heterogeneity rainfall patterns are highly influenced by seasonal monsoon wind flow, relative humidity, topography, and proximity to the sea [
25,
26,
27,
28].
Integration of satellite rainfall estimates with influential local-scale environmental factors, in the context of higher spatial (<0.1°) and temporal resolution (1–2 h), would allow the appropriate characterization of localized rainfall. At present, global digital elevation models (DEMs) are available at a resolution of 15 m, thus allowing the accurate modelling of orographic effects if coupled with wind vector data on a reasonable time scale. These types of wind data are generally obtained through direct observation, interpolation, or model estimation when limited data are available. It is essential to select a suitable orographic model for humid tropical climate conditions and coastal environments that also matches the resolution of the satellite gridded data. The modelling period should be specific to the monsoon season, when orography is the strongest influential factor. Experiments are required to evaluate the model. A successful experiment would greatly aid in filling the current gap in rainfall dynamics modelling in the humid tropics at the micro-watershed level.
In this study, we hypothesized that wind and orographic variables could be used as surrogate factors to transform coarse rainfall estimates to a daily-scale product, because the surrogate data had higher spatial and temporal resolution. We primarily utilized the high-resolution DEM as the downscaling input, coupled with hourly local wind estimates. We made the following assumptions: (1) the downscaling period occurred during the monsoon season when wind and orographic factors had considerable effects on the rainfall pattern, and (2) the selected watershed was primarily influenced by coastal effects and proximity to the sea. These downscaling assumptions are suitable for humid tropical micro-catchments, which experience substantial wind and orography effects at hourly temporal scales.
The proposed algorithm used proxy environmental variables as downscale modelling factors, instead of rain gauge data, such that it is suitable for ungauged catchment usage. The main objectives of this study were to use high-resolution hourly local orography effects and GPM rainfall data to determine the actual amount of precipitation that reached the ground at a daily scale, and to evaluate the accuracy and reliability of the downscaled product in comparison with rain gauge data. The Mass–Dempsey coastal wind model was modified for use with the uniform gridded pixel data used in this study. The quality of the downscaled rainfall products was then evaluated based on three criteria: (i) quantification of actual rainfall amount on the ground, (ii) representation of spatial rainfall distribution, and (iii) visualization of effective spatial rainfall patterns for micro-watersheds.
3. Results
This section is divided into four sub-sections, where each section represents a different element of spatio-temporal rainfall data performance, including (1) spatial rainfall pattern and distribution, (2) temporal rainfall data, (3) quantitative rainfall measurement, and (4) effective rainfall visualization at micro-watersheds scale.
3.1. Spatial Rainfall Pattern Representation Assessment
The bias ratio between the satellite rainfall and rain gauge data was used as an indicator to assess the goodness of fit of the rainfall spatial pattern (
Table 3). The downscaled data better represented actual rainfall spatial patterns, compared with the raw GPM data. However, bias ratio reduction was slightly affected by elevation. In comparison with raw rainfall data, the average bias ratio at high-elevation stations was reduced from 61% to 26%, whereas bias ratio at low-elevation stations was reduced from 64% to 35%. Thus, downscaled rainfall data generally tended to underestimate actual rainfall.
Downscaled satellite rainfall data occasionally overestimated rain gauge during heavy rainfall months (October–December). However, such instances were exceptionally rare and occurred at only one station (Klinik Bidan Kuala Abang). This overestimation may have been due to the close proximity of this station to the sea; raw GPM rainfall data pixels may have been mistakenly identified as land, instead of sea.
3.2. Temporal Rainfall Representation Assessment
Time-series analyses showed an improvement in seasonal trends of downscaled GPM rainfall (
Figure 8), with an increase in the correlation value from 18% to 22%. Greater bias was observed during the wet season, when downscaled GPM rainfall data consistently underestimated actual rainfall. Correlation was stronger at wetter stations (mainly at higher elevations) than at other stations; however, it was generally slightly lower (5%–8%) during the wet season, presumably due to greater error propagation during high-intensity rainfall.
3.3. Quantitative Rainfall Error Assessment
On average, downscaled GPM rainfall estimates showed smaller error values than the raw GPM rainfall estimates, with some variation among seasons and stations (
Table 3). The quantitative accuracy of the downscaled rainfall estimates was better in the wet season than in the normal season; respective improvements of 34% and 25% were observed during November and December, compared with May and June. During the wet season, the error decreased from 41 to 27 mm/d, while the error during the normal season decreased from 16 to 12 mm/d. Error reduction capacity was higher at high-elevation stations (Rumah Pam Paya, Bandar Al-Muktafi Billah Shah and Kg. Surau), especially during the wet season.
The accuracy of downscaled rainfall data improved at high- and low-elevation stations (<80 m) by averages of 36% and 29%, respectively, in the wet season; these trends were not observed in the dry season. Greater error reduction was observed in the wet season than in the dry season. The downscaled GPM data always performed similarly to or better than the raw GPM data. No case where the downscaled GPM data performed worse than the raw GPM data was found.
3.4. Qualitative Rainfall Assessment
Downscaled GPM rainfall data portrayed better rainfall variation than the raw GPM data during both seasons (
Figure 9). Significant impact appeared at high elevated areas. The heavy rainfall in the high elevated areas that is caused by the interaction between the incoming atmospheric moisture from the northeast direction during the wet season is clearly depicted by the downscaled rainfall map. Hence, the downscaled GPM rainfall maps were also capable of depicting the likely pattern of the interpolated ground rainfall, with better localized rainfall pattern detail on the micro-catchments level.
4. Discussion
The proposed downscaling procedures were able to produce high-resolution daily rainfall data from GPM data. The downscaled rainfall data also had higher accuracy and qualitative ability compared to the raw GPM data. Without the use of rain gauge data, this downscaling technique is suitable for the data conflicting micro-watersheds. The outcome of the downscaling is also robust throughout any season despite relying on the regional wind data and elevation data as inputs. Hence, the downscaling technique is suitable for operational rainfall mapping, monitoring, and assessment. Although the ground precipitation radar can supply rainfall rate data with a fine spatiotemporal resolution, it is hindered by several limitations. Such limitations are poor data management and archiving, usage is limited for weather forecasting and difficult image data processing.
The downscaling technique could contribute to water resource management in the critical tropical micro-watersheds. Many water resources related micro-watersheds in Southeast Asia are highly vulnerable to drought and yet they suffer from a lack of rainfall data [
32,
33,
34]. In addition, such watersheds also have a high risk of flooding and other related disasters (e.g., landslide) due to heavy and extreme rainfall. Therefore, the innovative solution of high-resolution daily rainfall maps would fill a significant gap on rainfall data in the effective water management and disaster framework [
35].
There was a drawback that will need to be adhered to in the future works for the betterment of the algorithm. The first is regarding the slight underperformance at an elevation below 80 m. Because the downscaling algorithm incorporates an orography effect that strongly affects the undulated and higher elevation areas, the less elevated coastal areas could be strongly affected by other significant factors such as distance to the sea. A study by Hayward and Clarke [
36] clarified the relationship between the rainfall and distance to the sea in Sierra Leone. Yamanaka [
37] also highlighted this in his work on the equatorial coastal rainfall. Furthermore, the heat contrast differences between the land surface and sea temperature could play a role in influencing the rainfall distribution in tropical coastal areas [
38,
39]. More recent research highlighted the influence of coastline geomorphology on the rainfall distribution. This was indicated by Alfahmi et al. [
40] and Yamanaka [
41] in their work in Indonesia. Incorporating such variables into the downscaling procedures could be able to improve the output performance in the future.
The second limitation in our study is the use of interpolated regional wind data to represent the high-resolution of vertical velocity and condensation rate (0.02°). Since, in theory, vertical convection could happen at smaller region scales than 0.02°, our downscaling approach is therefore unable to represent such variation, and this might affect the accuracy of the results. Considering that capturing such high-resolution local wind data is a mounting challenge especially for a remote, forested, and mountainous watershed; the other plausible way to anticipate such limitation is through the innovation of various wind data downscaling methods, such as those introduced by Hasager [
42] and Bentamy et al. [
43]. However, obtaining both high-temporal and high-resolution (<0.25°) data is yet to be achieved. Nevertheless, with the accuracy that is obtained through the use of regional wind data, we believe that it was a worthwhile solution that could compromise the absolute absence of ground rain gauge data for many tropical micro-watersheds. A continuous improvisation is highly recommended and subjected to further research.