Statistical Evaluation and Trend Analysis of ANN Based Satellite Products (PERSIANN) for the Kelani River Basin, Sri Lanka

Satellite-based precipitation products, (SbPPs) have piqued the interest of a number of researchers as a reliable replacement for observed rainfall data which often have limited time spans and missing days. e SbPPs possess certain uncertainties, thus, they cannot be directly used without comparing against observed rainfall data prior to use. e Kelani river basin is Sri Lanka’s fourth longest river and the main source of water for almost 5 million people. erefore, this research study aims to identify the potential of using SbPPs as a dierent method to measure rain besides using a rain gauge. Furthermore, the aim of the work is to examine the trends in precipitation products in the Kelani river basin. ree SbPPs, precipitation estimation using remotely sensed information using articial neural networks (PERSIANN), PERSIANN-cloud classication system (CCS), and PERSIANNclimate data record (CDR) and ground observed rain gauge daily rainfall data at nine locations were used for the analysis. Four continuous evaluation indices, namely, root mean square error (RMSE), (percent bias) PBias, correlation coecient (CC), and Nash‒Sutclie eciency (NSE) were used to determine the accuracy by comparing against observed rainfall data. Four categorical indices including probability of detection (POD), false alarm ratio (FAR), critical success index (CSI), and proportional constant (PC) were used to evaluate the rainfall detection capability of SbPPs. Mann‒Kendall test and Sen’s slope estimator were used to identifying whether a trend was present while themagnitudes of these were calculated by Sen’s slope. PERSIANN-CDR performed well by showing better performance in both POD and CSI.When compared to observed rainfall data, the PERSIANN product had the lowest RMSE value, while all products indicated underestimations. e CC and NSE of all three products with observed rainfall data were also low. Mixed results were obtained for the trend analysis as well. e overall results showed that all three products are not a better choice for the chosen study area.


Introduction
Rainfall is a crucial component in the hydrological cycle that plays a vital role in water-related activities such as agriculture, hydro-power generation, water resource management, climatology, meteorology, and disaster mitigation [1]. A research study done by Watson and Challinor [2] stated that rainfall is sensitive to the accuracy of models used to simulate crop growth, especially in areas that experience limited rainfall. is marks the importance of precipitation data for hydrological models. Accurate data is important to calibrate and validate hydrological models which will be used for future forecasting for di erent applications [3]. erefore, selecting accurate rainfall data as the input is of utmost importance [4].
ere are three main types of techniques used in the estimation of rainfall measurements. ey are ground-based rain gauge observations, weather radar observations, and satellite data sets [5]. Observed rain gauge data are thought to be the most accurate and dependable source of rainfall data. However, collecting continuous rain gauge data are a di cult task in many parts of the world due to administrative and technical constraints [6]. Due to complex topographical features, installation and maintaining ground-based rain gauges have been limited [7]. e distribution of rain gauges is limited due to high installation and maintenance costs mainly in developing countries [8]. erefore, missing data are a huge concern due to these limitations in rain gauge data, [4]. Due to the short length of rainfall records in these rain gauge stations, accurate hydrological modeling with the use of these data will lead to uncertain outcomes [6].
As a solution for the abovementioned limitations and shortcomings identified in rain gauge data, multisatellite high-resolution precipitation products such as tropical rainfall measuring mission (TRMM) multisatellite precipitation analysis (TMPA), precipitation estimation from remotely sensed information using artificial neutral network (PERSIANN) system, multisatellite precipitation analysis, and multisatellite rainfall estimate with climate prediction center (CPM) and morphing technique (CMORPH) and weather radar observations are widely in use around the world [9][10][11].
SbPPs provide higher spatiotemporal data at no cost [12]. However, these SbPPs are only used after a careful investigation in the desired study area. Since it was discovered that these SbPPs have certain uncertainties, such as accuracy that is affected by topographical features of the study area and precipitation mechanism due to seasonal and regional climate conditions, such accuracy evaluations of the products with respect to rain gauge data are done for each area of concern [7,9,10], which cannot be ignored if we plan to use them in hydrological applications [6]. A study done by et al., [13] in the Ganzi river basin of the Tibetan plateau to evaluate the impact of satellite data sets to be used in hydrological modeling for that area used CMORPH-CRT, PERSIANN-CDR, 3B42RT, and 3B42 satellite data sets against observed rainfall data using HEC-HMS model to find out that TRMM-3B42RT and CMORPH-CRT show good performance in the respective area and they also suggested that TRMM-3B42RT is a better choice overall for hydrological models in the Ganzi river basin of Tibetan plateau.
On the other hand, PERSIANN products have been widely applied to many countries. Sorooshian et al. [14] have tested the PERSIANN precipitation products to tropical rainfalls in the United States. eir findings concluded the necessity of further improvement to the SbPPs in the direction of spatial resolutions and accuracy. In addition, Ashouri et al. [15] stated the high applicability of PERSIANN products in various hydrological scenarios. Hurricane Katrina (in 2005) and Sydney flood in Australia (in 1986) were two of the verified cases of higher accuracies from PERSIANN precipitation products. However, Hong et al. [16] stated the overall PERSIANN products are in good agreement with northwestern Mexico rain gauges. Nevertheless, in some of the cases, the PERSIANN underestimated the precipitations. Furthermore, Miao et al. [17] have also ensured the performance of PERSIANN products in eastern China. erefore, PERSIANN precipitation products can be applied to many other areas after investigating their accuracy over localized regions. Sri Lanka is a country in which the economy mainly depends on agricultural activities. erefore, the presence of accurate rainfall data for future drought and flood forecasting is essential [18][19][20][21]. Even though Sri Lanka has dense rainfall measuring stations (more than 500 for the total area of 65610 km 2 ), some of the stations are scattered in urbanized areas. e remote areas are not uniformly covered by rain gauges. In addition, some of the rain gauges are not properly functioning and regularly maintained due to their remote locations and other logistic issues. erefore, there can be some disputes in the measured rainfalls even though most of the researchers treated them to be highly accurate.
us, it is important to understand the behavior of SbPPs as an alternative input variable for hydrologic and hydraulic studies. However, the spatio-temporal variation of rainfall is highly dynamic from catchment to the catchment. erefore, the acceptability of a particular SbPP has to be investigated in a localized context.
Despite the fact that several studies have been conducted in many places around the world, no such evaluation of SbPPs using observed rainfall data has been conducted for the Kelani river basin in Sri Lanka, which is the country's fourth longest river [22,23]. is research study aims to identify the appropriateness of using PERSIANN SbPPs as an alternative to observe rain gauge measurements and also to analyze the trends in both SbPP and rain gauge data in the Kelani river basin, Sri Lanka. As per the authors' best knowledge, this is the first-ever study to investigate the acceptability of any SbPPs against the ground-measured rainfall in the Kelani river basin.

Study Area.
Sri Lanka is a country having 103 river systems that start from the central mountains and flow down to the coastal areas. Out of those 103 rivers, the Kelani river is the fourth-longest river which originates from the Sri Pada Mountain range (Adam's peak range), and passes through four districts of Nuwara-Eliya, Kegalle, Gampaha, and Colombo in the country and flows down to the Indian Ocean through the Colombo [23]. It is about 145 km long and the basin drains an area of over 2340 km 2 being the seventhlargest river basin in Sri Lanka. e Maha Oya and Attanagalu rivers mark the divide between the northeast and northwest boundaries of the watershed [22]. Because it is a river basin in the country's wet zone, the river's average annual discharge is around 4969 million cubic meters (MCM) [24]. During the southwest monsoon season, the Kelani river basin receives the most rainfall, with an average annual precipitation of 2400 mm [7].
e Kelani river is located in a very fertile and populated zone with the elevation varying from sea level to above 2000m above mean sea level (AMSL). e upper catchment is mountainous, whilst the lower catchment is nearly flat [22]. Due to these geographical orientations and the terrain of the basin, the Kelani river basin is very vulnerable to flooding [23]. is basin is of utmost importance to the country as this basin provides hydropower using Castlereagh, Laxapana, and Maussakelle reservoirs in the upper catchment area [23,25]. e elevation map and the rain gauge stations located in the Kelani river basin are demonstrated in Figure 1.

Rain Gauge Data.
e observed rain gauge data were obtained from Sri Lanka's department of meteorology, which is the country's o cial authority for meteorological data gathering. e analysis was based on daily data from nine rain gauge stations located throughout the catchment from 1989 to 2016.

Satellite Rainfall Products.
ree types of SbPPs from the PERSIANN family of products were used for the analysis. Precipitation data for all nine stations mentioned in Table 1 were obtained from these SbPPs. Table 2 contains general information about the SbPPs employed, and Figure 3 graphically depicts the spatial coverage of the nine recorded rainfall stations in each product's appropriate spatial resolution. e three SbPPs used for the study are as follows: (i) Precipitation estimation from remotely sensed information using arti cial neural networks (PERSIANN)  Applied Computational Intelligence and Soft Computing (ii) Precipitation estimation from remotely sensed information using arti cial neural networks-cloud classi cation system (PERSIANN-CCS) (iii) Precipitation estimation from remotely sensed information using arti cial neural networks-climate data record (PERSIANN-CDR) PERSIANN-CDR gets input from the infrared imagery from geostationary (GEO) satellites. It takes the inputs from GEO satellites available since 1979 under a spatial resolution of 10 km. PERSIANN algorithm was developed in 1997 using low Earth orbit (LEO) satellites and high-frequency samples from GEO satellites. e precipitation data can be

Methodology
e accuracy of SbPPs was statistically evaluated using continuous evaluation indicators and categorical indices. Due to the uneven ranges of data availability of three SbPPs (refer to Table 2) and ground-measured rainfall, a common period of 2003-2016 (14 years) was considered for the analysis. In addition, the SbPPs data were extracted directly to the coordinate of the ground-measured rain gauge station. erefore, the impact of spatial resolution of SbPPs to ground-measured is minimum. Furthermore, nonparametric tests were then used to determine the trends and magnitudes of trends observed in the SbPPs for the area of interest in this study.
Rainfall intensity class based on their rainfall thresholds, P are described as below [7,21,31]: (i) No/tiny rainfall -P < 1 mm (ii) Light rainfall -1 mm ≤ P < 2 mm (iii) Low moderate rainfall -2 mm ≤ P < 5 mm (iv) High moderate rainfall -5 mm ≤ P < 10 mm (v) Heavy rainfall -P ≥ 10 mm For the analysis, 1 mm/day threshold was used which falls under the "light rainfall" intensity class. e contingency table used for the analysis is also given in Table 3 [7,21].

Nonparametric Tests.
e monthly, annual, and seasonal trends of the two datasets were analyzed using the Mann- where; y j and y i are monthly precipitation values. e normalized test statistic, Z c signifies the standard normal distribution and the positive or the negative value of Z c indicates whether the trend is increasing or decreasing, respectively, [33].
where, n is the number of data points, t is the number of tied groups and t i is the number of datasets in the i th group. For a normal distribution with a mean of 0 and a standard deviation of 1, the probability density function, f(z), is given by, If Z is negative and the probability is greater than 0.95, the trend will be found to be decreasing at a 95% significance level. Similarly, if Z is positive and the probability is more than 0.95, the trend will be found to be increasing. If the probability is less than 0.95, it is assumed that there is no trend in the data [34].
Sen's slope is a linear slope estimator used for determining the magnitude of a trend series [35]. e slope of data sets was determined by using Equation (12) and, Sen's slope is given by Equation (13). Positive Q i values signify increasing trend patterns and, negative Q i values signify decreasing trend patterns.
where j > k, X corresponds to a data value at a j/k time and k � 1,2, . . ., N , n is odd, 1 2 d n/2 + d (n+2/2) , n is even. Figure 4 presents the visual rainfall variations for all stations over the years for observed and SbPPs. Only the annual rainfalls are shown in the figures due to the complexity of the plots. Even though the rainfall from one year to another year is not connected, lines were chosen to develop the plots to showcase the variation of rainfall. Variations can be observed from observed rainfalls to SbPPs and some of these variations are significant. However, the peaks and troughs of SbPPs match the observed rainfalls. In addition, the special features can be seen in Figures 4(e), 4(g), 4(h), and 4(i). e observed rainfall line is not continuous (abnormal in the years 2013 and 2015 to Figure 4(h)). e observed daily rainfalls were unavailable for these years (or for many days) due to instrumental issues and some other logistic issues (like flooding in the area). erefore, the objective of this research study is clearly stated. For PERSIANN-CCS, the lowest FAR seen was 0.33 in Maliboda, and the highest FAR was seen as 0.45 in Hanwella.  Comparatively, from the stationwise mean values, higher results for accurate rainfall detections (POD) were observed in PERSIANN-CDR (POD > 0.8). PERSIANN product showed high performance in having a lower fraction of false rainfall detections (FAR) compared to other products.

Categorical
PERSIANN-CCS and CDR both indicated good performance in correct rainfall predictions (CSI). Among the three products evaluated for detection and prediction capabilities, PERSIANN-CDR proved to be a better product by showing better performance in both POD and CSI. Similar results can also be seen in a study done in upper Nan river basin, Northern ailand by Gunathilake et al. [7]. Figure 5 represents the results obtained. PERSIANN-CDR showed the highest percentage bias as observed in the results. CC values for all products in all nine stations varied between 0.14-0.49. e stationwise mean values calculated for each product showed that the correlation between the two datasets with all three satellite products shows not much difference. e correlation is relatively less and falls in the moderate range [30]. However, all products showed positive correlations with the observed data.

Continuous Evaluation Indices
Most of the stations produced negative NSE values in the PERSIANN-CCS product which indicates unacceptable performance. However, PERSIANN-CDR and PERSIANN obtained values in the range of 0 and 1 for NSE in most of the stations and thus proved acceptable levels of performance. None of the products showed optimal performance in NSE when compared with the stationwise mean values.
In continuous evaluation indices, the lowest RMSE value was observed in the PERSIANN product which indicated values lower than the stationwise mean value for six out of nine stations [36]. In PBias results, all three satellite data types show underestimation by indicating positive values. Compared to the other two satellite data sets, PERSIANN-CDR shows larger underestimations in all nine stations [36]. CC results for all three products showed values between 0-0.5 which cannot be counted as a significantly good range. NSE also showed values less than 0.5 in all stations for PERSIANN-CCS, PERSIANN-CDR, and PERSIANN which again is not closer to the optimal value of 1 [33]. Among the evaluated three products with continuous evaluation indices, PERSIANN-CDR showed better performance overall followed by the PERSIANN product and PERSIANN-CCS product. Figure 6 represents the results obtained.

MK Test and Sens Slope Estimator
Results. Trend analysis using MK test for monthly time scale was done to identify any similar significant positive or negative trends in observed data and satellite data. e trends observed were then quantified using Sen's slope estimator. e annual and seasonal trend analysis of the observed data in all nine stations yielded no significant trends. Similarly, no significant trends were observed in all three products under evaluation as well. However, in the monthly trend analysis, observed data showed a significant decreasing trend in Kitulgala station in the month of July with a magnitude of 7.5 mm/month. Increasing trends were observed in March at Deraniyagala and Avissawella stations. Another increasing trend was also observed at Chesterford station in August with a magnitude of 7.9 mm/ month. In PERSIANN-CCS products, significant increasing trends were observed in Kitulgala, Holombuwa, Deraniyagala, Chesterford, and Avissawella stations. In the PERSIANN-CDR product, no significant trends were observed in relation Applied Computational Intelligence and Soft Computing to the trends observed in the observed data. PERSIANN product produced significant trends in Glenn course and Hanwella showed 6.2 mm/month and 4.9 mm/month decreasing trends in January. Kitulgala, Deraniayagala and, Maliboda stations also showed decreasing trends in November with magnitudes of 6.4 mm/month, 6.4 mm/month and, 17.7 mm/month, respectively. PERSIANN-CCS was the only product that showed some relevance in the trends observed with observed data. Figure 7 represents the results obtained.

Conclusions
is study was carried out in the Kelani river basin, Sri Lanka, to determine the suitability of employing satellite-based precipitation products as an alternative to rain gauge observations and to analyze the trends of each dataset. ree SbPPs, namely, PERSIANN, PERSIANN-CCS, and PERSIANN-CDR were used in this research study. Nine rain gauge stations distributed in the Kelani river basin were used for the analysis. Both observed rainfall data and SbPP data were obtained between the years 1989-2016. Four continuous indices including RMSE, PBias, CC, and NSE, were used to check the accuracy of satellite-based precipitation products by comparing them against observed rainfall. Four categorical indices including POD, FAR, CSI, and PC, were used to evaluate the capability of detection of rainfall in satellite products on a daily scale. MK test was used to identify the trends, and Sen's slope estimator test was used to determine the magnitudes of trends between the two datasets. Results showed that PERSIANN-CDR performed well by showing better performance in both POD and CSI. e lowest RMSE value was observed in the PERSIANN product, and all products showed underestimations when compared with observed rainfall data. e correlation, CC, and NSE of all three products with observed data were also quite weak. Furthermore, from the trend analysis results, few trends were identified in both observed data and satellite data on monthly, annual and seasonal scales however the results were not significant. e results showed that all three products are not a better choice overall for the Kelani river basin, Sri Lanka. erefore, it is suggested to continue this study with more SbPPs that are readily available to find out the best SbPP alternative for observed rain gauge data in the Kelani river basin, Sri Lanka.

Data Availability
e climatic data used in this research study are available upon request for research purposes.

Conflicts of Interest
e authors declare that they have no conflicts of interest. Acknowledgments e authors are grateful to all SRP communities for making the precipitation data freely available to the international research community.