Examining El Niño–Southern Oscillation Effects in the Subtropical Zone to Forecast Long-Distance Total Rainfall from Typhoons: A Case Study in Taiwan
1. Introduction
Taiwan, an island with an area of 36 000 km2, lies in the main track of western North Pacific (WNP) typhoons (Fig. 1). The typhoon season is a distinctive climatic characteristic of Taiwan. Several typhoons make landfall in Taiwan every year, usually during late summer and early autumn, although winter typhoons also sometimes occur (Fan 2011). An average of 3.5 typhoons pass near or over Taiwan annually.
Geographical location of Taiwan and the study site, with historical typhoon tracks.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
El Niño–Southern Oscillation (ENSO) is the main cause of climate variability at seasonal and interannual time scales. Climate state changes, usually characterized by a shift in means, can cause formerly rare events that follow these mean changes to occur more frequently and with increasing variability (Salinger 1994). ENSO is characterized by an interannual cooling (La Niña) and warming (El Niño) of the eastern equatorial Pacific (Collins et al. 2010); however, since 1976, El Niño episodes of the Southern Oscillation have increased in frequency and have become more extreme (Salinger et al. 2000). Although ENSO originates in the tropical Pacific, it has a global effect on weather and climate. Several scholars have studied the influences of ENSO on tropical cyclone (TC; synonymous with typhoon in this paper) activities in the WNP, including formation, track, and intensity (Camargo and Sobel 2005; Camargo et al. 2007; Chan 2000; Elsner and Liu 2003; Teng et al. 2014). For example, Lander (1994) found that TCs are more likely to form east of approximately 160°E during El Niño events, when sea surface temperatures (SSTs) in the central and eastern equatorial Pacific are higher than normal. Elsewhere, Chan (1985) and Teng et al. (2014) have determined that typhoons tend to form farther to the east in the WNP during El Niño years, and Wang and Chan (2002) indicated that the average maximum intensity of TCs is higher and lower in El Niño and La Niña years, respectively, than in normal years.
Other research has demonstrated that the extreme precipitation of typhoons, which may be affected by ENSO, leads to numerous casualties and considerable economic loss. For example, Typhoon Soudelor (7–9 August 2015) generated 1400 mm of rainfall in northeastern Taiwan, resulting in 12 fatalities and more than $100 million (U.S. dollars) in damage. Consequently, developing a model that includes the effects of ENSO is crucial for predicting total rainfall in real time prior to a typhoon’s arrival in Taiwan, as well as saving both human lives and resources.
In Taiwan, the Central Weather Bureau (CWB) issues “Typhoon Warning over Ocean” (TWO) alerts every 3 h when a typhoon is expected to affect Taiwan within 24 h. Subsequently, “Typhoon Warning over Ocean and Land” (TWOL) alerts are issued every hour when a typhoon is expected to land or affect the region within 18 h. Within those final 18 h, the CWB also nowcasts the total precipitation. Because of high climatological uncertainty, the rainfall prediction provided by the CWB is a range; however, these ranges are not always accurate. At present, the CWB employs an ensemble forecasting method that is dynamic and flow dependent to quantify and communicate forecast uncertainty. However, CWB total rainfall nowcasts often have high prediction errors. These errors are primarily attributable to the island’s Central Mountain Range (CMR). The CMR is 340 km long and 80 km wide, with an average height of 2500 m. As a typhoon approaches Taiwan, the CMR influences its track and intensity: specifically, the topography increases the rainfall amount significantly by lifting moist air over the windward side of the mountains. Hong et al. (2015) found that significant mesoscale variations caused by orographic effects—including track deflection, secondary low development, intensity changes, and mesoscale changes in pressure, wind, and precipitation—are common and contribute to the challenge of providing accurate quantitative typhoon precipitation forecasts in Taiwan. For example, during the approach of Typhoon Soudelor in 2015, the CWB issued a rainfall nowcast ranging from 400 to 600 mm over the Hualien County plain (eastern Taiwan); by contrast, the actual total rainfall was 199 mm. Total rainfall nowcasting is also nonexistent when a typhoon is several days away from Taiwan. However, long-distance total rainfall nowcasting is essential for enabling the residents of Taiwan to make appropriate flood control and disaster prevention preparations.
Improvements in both computer capability and algorithm efficiency have made real-time typhoon rainfall prediction models increasingly useful for water resource management (Wu et al. 2010; Wu and Chau 2013). Antolik (2000), Madsen et al. (2014), Maier and Dandy (2000), Michaelides et al. (2009), and Scofield and Kuligowski (2003) have conducted in-depth reviews of rainfall prediction models, covering an extensive range of research on this topic. In Taiwan, numerous studies have also successfully addressed rainfall prediction and identification problems for typhoons affecting Taiwan in real time (Chang et al. 2013; Hong et al. 2015; Kuo et al. 2016; Wang et al. 2016; Wei 2013; Wei et al. 2015). For example, Kuo et al. (2011) built a dynamic time series model for long-term extreme rainfall to investigate common trends in annual maximum precipitation. Wei and Roan (2012) addressed the rainfall retrieval problem for quantitative precipitation forecasting over land during typhoons. More recently, Lin and Jhong (2015) developed typhoon rainfall forecasting models that yield 1–6-h advanced hourly rainfall predictions for southern Taiwan. In addition, Lo et al. (2015) developed an artificial neural network–based model for forecasting precipitation in eastern Taiwan, using an automatic calibration approach to adjust the training parameters.
The purpose of this study was to develop a model that can predict the total rainfall during a typhoon when the storm is still several days away from Taiwan. The concept used in this paper is inspired by the previous work of Wei (2014), who predicted the hourly rainfall during a typhoon invasion in Taiwan. Wei (2014) used meteorological and radar reflectivity data to simulate the operational forecasting of real-time hourly rainfall for developing a rainfall forecasting model for typhoons. When a typhoon makes landfall over Taiwan, no extra energy is added to the system from the open sea and the typhoon’s intensity thus weakens. Therefore, we did not consider the ENSO state in our formulation of an hourly rainfall prediction model. However, a typhoon originates from a low pressure system that develops over tropical or subtropical waters; thus, the climatic variability in the tropical waters should be taken into consideration. The ENSO indices, such as the Southern Oscillation index (SOI) and Niño-3.4, can be used to identify the degree of the climate variability according to variability records for past consecutive months in the eastern and central equatorial Pacific region. Accordingly, in this paper we added the ENSO indices—SOI and Niño-3.4—in our formulation of the total rainfall prediction model in order to identify the ENSO state (El Niño and La Niña). When a typhoon forms, the typical climatic variability in that specific month in the tropical or subtropical waters can be determined in advance. Thus, the added climatic information such as high or low SST, which might be affecting the typhoon intensity, can be employed to increase prediction efficiency.
In the paper, we designed two models (scenarios)—one including the ENSO indices and one without the indices—and compared their prediction efficiency. The forecasting dates begin when a tropical depression is formed and terminate when a typhoon makes landfall. The Hualien Weather Station (Fig. 1) in Taiwan was selected as the experimental site. Two decision tree algorithms, namely, the C4.5 and random forests (RF), which are based on data mining techniques, were adopted for calculations. The results of the scenarios were subsequently compared with the values predicted by the CWB.
Data mining is the process of discovering knowledge from a large amount of data and requires no prior domain knowledge (Sun et al. 2007). The process involves numerous algorithms, including classification, clustering, association, prediction, and time series analysis. In particular, decision trees have become widespread for classification tasks because of their ease of use, interpretability, and ability to deal with covariates measured at different levels (Lariviere and den Poel 2005). The C4.5 algorithm was introduced by Quinlan (1993) and uses a divide-and-conquer approach to grow decision trees. Conversely, the RF algorithm developed by Breiman (2001) is an ensemble of classification trees; the theorems of this algorithm are presented in section 3b. Both the C4.5 and RF decision trees have been utilized to address numerous civil and oceanic engineering and atmospheric science problems. For example, the C4.5 has been adopted to study water reservoir control (Bessler et al. 2003; Wei and Hsu 2008, 2009), wave height prediction (Mahjoobi and Etemad-Shahidi 2008), and cloud-ceiling forecasting (Bankert and Hadjimichael 2007), whereas the RF has been adopted to study land-cover classification (Gislason et al. 2006; Schneider 2012), remote sensing (Belgiu and Dragut 2016), and groundwater aquifers (Baudron et al. 2013).
Because of the uncertainty associated with precise precipitation amounts, prediction values are provided using a range scale, similar to the system used by the CWB. To compare the range-scale-based forecasts, decision tree algorithms were used to create a categorical-type target for rainfall ranges with discretized intervals. The results generated by the categorical-type target were then compared with those generated by a continuous-type target.
The remainder of this paper is organized as follows. Section 2 describes the study area and the concept of real-time forecast processes. Section 3 describes the proposed total rainfall prediction model and model evaluations. Section 4 outlines the model with consideration of the ENSO effect and details several typhoon simulations. Finally, section 5 offers some conclusions.
2. Application
a. Study area and typhoons
Figure 1 shows that typhoons frequently make landfall in Taiwan, bringing strong wind gusts and downpours to the island. The Hualien Weather Station (23°58′37″N, 121°36′18″E), located on the Hualien County plain in eastern Taiwan, was selected as the study site. In total, 30 typhoon events (tracks shown in Fig. 1) that affected Taiwan with torrential rain and strong winds between 2001 and 2015 were analyzed. A list of the dates prior to typhoon landfall in Taiwan and the total rainfall amounts during the typhoons that affected the Hualien Weather Station are presented in Table 1.
Typhoons affecting the Hualien Weather Station and total rainfall at this station.
b. Data sources
The SSTs, satellite brightness temperatures (TBs), and the total rainfall amounts from each examined typhoon were included in the data. Records for the typhoon dates during 2001–15 were also collected.
1) Typhoon climatological characteristics
Typhoon climatological characteristics are produced by the Joint Typhoon Warning Center (JTWC) in Hawaii. This information is intended for U.S. government agencies, but it is also accessible to the general public (Chu et al. 2002). The JTWC maintains an archive of typhoon track data, which are commonly referred as the “best tracks.” Each best track file contains typhoon center locations (latitude and longitude), maximum sustained wind speed (km h−1), sea level pressure (SLP) at the typhoon center (mb), pressure of the last closed isobar (mb), radius of the last closed isobar (km), and radius of maximal winds (km) at 6-h intervals.
2) Sea surface temperatures
The SST data were collected from the Advanced Very High Resolution Radiometer (AVHRR) of the National Oceanic and Atmospheric Administration (NOAA). SSTs were measured in degrees Celsius and had a spatial resolution of 100 km. The AVHRR is a radiation-detection imager that can be used for determining cloud cover and surface temperature remotely. The latest version (AVHRR/3), launched by NOAA-15 in May 1998, has six channels that collect different bands of radiation wavelengths.
3) Satellite brightness temperatures
The microwave data collected from Special Sensor Microwave Imager/Sounder (SSMIS) satellite instruments are TB records. The SSMIS is part of the instrument suite attached to the Defense Meteorological Satellite Program series of satellites (Raytheon 2000). It is a seven-channel, four-frequency, linearly polarized passive microwave radiometric system that measures atmospheric, ocean, and terrestrial microwave TBs through four vertically polarized channels (19V, 22V, 37V, 85V) and three horizontally polarized channels (19H, 37H, 85H).
4) Total rainfall during typhoons
Rainfall data from the Hualien Weather Station, measured by automatic meteorological gauges, were also collected, along with the complete hourly data available from the CWB. In addition, the nowcasts of total typhoon rainfall on the Hualien Weather Station prior to typhoon landfall were gathered.
Following data collection, all covariates were assembled in the typhoon information dataset {A} (including attributes A1−A8 in Table 2), satellite microwave dataset {B} (including attributes B1−B8 in Table 2), and typhoon rainfall dataset {R}. In total, the data are composed of 750 records measured in 6-h intervals (i.e., the forecasting horizon is 6 h). Table 2 also shows the mean, minimum, and maximum values obtained from the dataset {A, B, R}. Notably, the scattering index (SI) attribute (B8) was derived from the TB data (see section 2e for details).
Statistical attributes of all datasets.
c. ENSO indices
ENSO is the strongest interannual climate mode and is inherently atmosphere–ocean coupled (Neelin et al. 1998). Various metrics can be used to characterize El Niño, such as SST, SLP, surface winds, surface temperature, and outgoing longwave radiation. For the present study, the SOI and Niño-3.4 were selected as the indices to monitor ENSO.
1) Southern Oscillation index
The SOI is the normalized pressure difference between Tahiti and Darwin, Northern Territory, Australia. Several slight differences are noticeable in the SOI values calculated at various centers; for this study, we obtained SOI data from the Global Climate Observing System (GCOS), which uses a method introduced by Ropelewski and Jones (1987). The SOI is a time series used to characterize large-scale SLP patterns in the tropical Pacific, and is computed using monthly mean SLP anomalies at Tahiti (T) and Darwin (D). Trenberth (1984) noted that SOI values are derived using normalization factors based on annual means and that they maximize the signal-to-noise ratio. The SOI was selected as an indicator because it is the oldest indicator of the ENSO state, with station records that begin in the late 1800s (Barnston 2015).
2) Niño-3.4 sea surface temperature anomaly
Numerous indices, such as Niño-1+2, Niño-3, Niño-3.4, and Niño-4, are used to monitor the tropical Pacific, all of which are based on the sea surface temperature anomaly (SSTA) averaged across a given region (Rasmusson and Carpenter 1982; Trenberth and Stepaniak 2001). The Niño-3.4 index, defined as 5°N–5°S, 120°–170°W, uses a 5-month running mean. El Niño or La Niña events are defined when the Niño-3.4 SSTs exceed ±0.4°C for a period of 6 months or more. At present, the Niño-3.4 anomalies may be thought of as representing the average equatorial SSTs across the Pacific approximately from the date line to the South American coast, and Niño-3.4 is therefore identified as the most ENSO-representative index (Barnston et al. 1997). Thus, we employed Niño-3.4 SSTA as an attribute for scenario modeling.
Both the SOI and Niño-3.4 indices were obtained from the GCOS Working Group on Surface Pressure. The GCOS is a long-term user-driven operational system capable of providing the comprehensive observations required for monitoring climate systems and detecting climate change and its causes. Notably, dataset {S} is preprocessed according to the corresponding SOI and Niño-3.4 monthly data collected from each typhoon. For example, if a typhoon occurs in July, then the 6-h records in dataset {S} are included as the July data found in the SOI and Niño-3.4 indices.
d. Concept of the proposed real-time forecast processes
Figure 2 conceptualizes the real-time forecasting process that uses the proposed long-distance total rainfall forecast (LTRF) model when typhoons begin developing and may approach Taiwan in the future. The forecast horizons start with the formation of a tropical depression and end when the typhoon makes landfall in Taiwan or dissipates. The procedure is performed as follows:
Step 1: Initiate the process when a tropical depression forms and obtain the predicted tracks from the JTWC.
Step 2: Determine whether the typhoon might affect Taiwan according to the predicted tracks at period t. If the track of the typhoon passes over Taiwan, then obtain the typhoon climatological characteristics and continue to step 3; otherwise, go to step 9.
Step 3: Obtain various sensor data, including the SSTs, TBs, and ENSO indices. The ENSO indices are used later in step 5 for the climate scenario with ENSO effects.
Step 4: Derive the SI (section 2e) from the TB data.
Step 5: Select the climate scenarios. Scenario 1 (using the {A, B, R} dataset for input–output patterns) examines climatological conditions without ENSO effects, whereas scenario 2 (using the {A, B, S, R} dataset) examines climatological conditions with ENSO effects. The new attribute, dataset {S}, includes the ENSO indices of SOI and Niño-3.4 SSTA.
Step 6: Run the proposed LTRF model (section 3).
Step 7: Derive the total precipitation forecasts
Step 8: Record the forecast values
Step 9: Update the time period t = t + 6.
Step 10: Identify whether the typhoon makes landfall in Taiwan or whether it dissipates according to the typhoon tracks issued by the JTWC. If it makes landfall in Taiwan or dissipates, then terminate the prediction process; otherwise, return to step 1.
Flowchart of the real-time processes of the LTRF at 6-h intervals.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
e. Scattering index
Grody (1991) indicated that a SI greater than 10 K is an accurate global indicator of rain.
3. LTRF model development
a. Procedure for LTRF model development
The following steps comprise the LTRF model:
Step 1: Refine the {A, B, S, R} datasets from the various sensors. All records are preprocessed on a 6-h scale prior to modeling. The target field is the total rainfall during typhoons.
Step 2: Preprocess the raw data for classification trees (C4.5 and RF) and regression trees (RF). To create categorical-type target labels, the dataset {R} is partitioned into several intervals using C4.5 and RF algorithms (section 3b).
Step 3: Classify the datasets into training and testing subsets (section 3b). The training subset is used to build model structures and parameters, followed by the testing subset that evaluates the performance of the model to verify its generalizability.
Step 4: Design the climate scenarios. Scenario 1 uses {A, B, R} and scenario 2 uses {A, B, S, R}.
Step 5: Construct the scenario models using the training subsets.
Step 6: Run the constructed scenario models using the testing subsets and compute their performance metrics.
Step 7: Evaluate the results and obtain the optimal scenarios for model representations.
b. Algorithms
The theories of the two tree-based algorithms are described in the following subsections.
1) C4.5
An attribute and split are selected to minimize the entropy. In C4.5, given k, the splitting criterion used is GainRatio(k) = Gain(k)/SplitInformation(k) (Apté and Weiss 1997). This ratio expresses the proportion of information generated by a split that is helpful for developing the classification, and it may be thought of as a normalized information gain or entropy measure for the test. A test that maximizes this ratio is selected, as long as the numerator (the information gain) is larger than the average gain across all tests. Notably, the numerator in this ratio is the standard information entropy difference achieved at k. A more complete discussion of C4.5 is provided by Quinlan (1993).
2) Random forests
The design of RF was influenced by the idea of random subspace selection described by Ho (1995) and the work of Amit and Geman (1997), who introduced the idea of searching over a random subset of the available decisions when splitting a node. The RF learner is typically grown using classification and regression tree algorithms (Breiman et al. 1984), in which binary splits recursively partition the tree into homogeneous or near-homogeneous terminal nodes (Chen and Ishwaran 2012). RF obtains various trees on the basis of a single training set and then introduces randomness into the tree construction (Scornet 2016). Notably, the procedure stands out for its ability to deal with complex nonlinear relationships between variables while also minimizing problems with overfitting (He et al. 2016).
RF trains each individual tree on bootstrap resamples (M samples) of the total dataset. For each split on the node of a tree, RF uses a random selection of a subset of Ntry predictors (covariates), rather than the total explanatory variables (N), to grow each tree. Thus, M decision trees are fitted and the final result is decided by average or majority voting (He et al. 2016).
c. Data preprocessing and division
In the present study, we partitioned the precipitation scale into several intervals to build a classification tree. Specifically, rainfall was classified into 25-mm groups and the quotient was rounded up to an integer, which was then preceded by the word class. For example, if the rainfall amount was 139 mm, the target was classified as “class 6.” We also evaluated other interval partitioning cases, comprising 25, 50, 75, and 100 mm, to determine the suitable partition size, and denoted these intervals as Int-25, Int-50, Int-75, and Int-100, respectively. The continuous-type target field case with no interval was denoted as Int-0 in the subsequent analysis.
Next, the k-fold cross-validation approach was employed to evaluate the two tree-based algorithms. Cross validation is a technique to evaluate predictive models by partitioning the original sample into a training set (to train the model) and a test set (to evaluate the model). Of the k subsamples in this study, a single subsample was retained as the validation data for model testing, and the remaining subsamples were used as training data. This study adopted tenfold cross validation because it is one of the most commonly used approaches.
d. Parameter setup
For the classification trees, the classified outputs of class labels obtained using C4.5 were transformed into their original units (mm) of rainfall to compare the prediction errors of the four partitioning interval choices (Int-25, Int-50, Int-75, and Int-100). For example, if the predicted rainfall was categorized into class 5 using the original partition interval of 25 mm, then the rainfall value was restored to (5–0.5) × 25 mm = 112.5 mm. Here, the combined {A, B, R} dataset was used for modeling. As shown in Fig. 3a, the minimum number of instances per leaf for C4.5 was set as 1 for all four partitioning intervals.
Calibration of model parameters: (a) minimum number of instances per leaf for C4.5 and (b) number of trees for RF.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
For the RF algorithm, the two vital parameters are Ntry and M. The Ntry determines the variation among different decision trees, and the M influences the extent of overfitting (Liaw and Wiener 2002). Typically, Ntry = log2 (N + 1). Higher values of M are expected to yield more accurate performance but require more computational resources. This study employed a standard approach—namely, the out-of-bag (OOB) error or estimate—to determine the optimal M values. As described by Zhang et al. (2016), each tree is constructed using the bootstrap sample. The OOB datasets are used to produce an unbiased estimate of the prediction accuracy. Each OOB sample is provided as input into the constructed trees to generate a test set for classification. Finally, m is considered the class that receives the most votes each time sample M shows the OOB error. The proportion of instances when m is not the true class of M averaged over all samples is the OOB estimate.
Subsequently, the generalizability of the RF model was investigated by OOB errors. Figure 3b shows that no significant decrease in OOB errors was found after 90 individual decision trees of RF were constructed, which means that adding more trees is unnecessary; therefore, the number of decision trees was set to 90 for these partitioning intervals.
e. Analysis and evaluation
As noted in step 4 (section 3a), the {A, B, R} dataset is first used for modeling scenario 1. We adopted three combinations, comprising the subsets {A, R}, {B, R}, and {A, B, R}, which were then validated for pattern evaluations. The model parameters of the {A, B, R} combination were calibrated in a previous section, and the model parameters of the {A, R} and {B, R} combinations were trained using the same method. Figures 4a,b reveal the accuracy of the intervals obtained using the three combinations of subsets with the C4.5 and RF models, respectively. However, because numerical targets were employed in the case of Int-0, continuous-type outputs were generated, which prevents the accuracy of these results from being demonstrated in the figure.
C4.5 and RF model results for the data combinations of {A,R}, {B,R}, and {A, B, R} regarding (a),(b) accuracy; (c),(d) MAE; and (e),(f) RMSE.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
The classification accuracy can be defined as follows: accuracy (%) = correct trials/total simulation trials. This ratio expresses the number of predicted runs that match the desired target, divided by the total number of simulation trials. Figure 4a shows that increasing the interval sizes for the nominal target resulted in higher accuracies for the C4.5 and RF models. This occurs because fewer class labels can yield superior classification results within the limited domain ranges of the target.
Figures 4c,e illustrate that the optimum value for C4.5 occurs at the partitioning interval of Int-25 using the subset {A, B, R}, where MAE = 28.7 mm and RMSE = 77.3 mm. By contrast, Figs. 4d,f show that the optimum value for RF occurs at the continuous target of Int-0 using the subset {A, B, R}, where MAE = 20.1 mm and RMSE = 59.4 mm. Finally, Fig. 5 depicts scatterplots that compare the target and predicted results for the partitioning intervals of Int-0, Int-25, Int-50, Int-75, and Int-100 obtained using the {A, B, R} subset. On the basis of these results, the data combination of {A, B, R} was selected for subsequent analysis.
Scatterplot of the observations and predictions of scenario 1 with different partitioning intervals: (a) Int-0, (b) Int-25, (c) Int-50, (d) Int-75, and (e) Int-100.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
4. LTRF model with consideration of ENSO effects
In this section, we develop an advanced LTRF model that also examines the effects of ENSO (step 4 in section 3a).
a. Climate scenarios and evaluations
In this section, we evaluate the prediction models for the two climate scenarios. The scenario 1 model was built using the {A, B, R} data combination described in section 3c; by contrast, the scenario 2 model employed the {A, B, S, R} data combination, which is built in this section. We also adopted the same tenfold cross-validation approach for model evaluation and selected the same partitioning intervals as those for scenario 1. Figure 6a shows accuracy plots with various intervals for the two climatic scenarios. Specifically, the accuracies of scenario 2 ranged from 94.4% to 97.0% and from 97.5% to 98.5% for the C4.5 and RF algorithms, respectively, and they increase as the partitioning interval sizes increase. Figure 6b depicts the MAE results of the two scenario models using C4.5 and RF. Notably, when C4.5 was used, optimal MAE values of 28.7 and 14.9 mm were achieved for the partitioning interval of Int-25 for scenarios 1 and 2, respectively. Conversely, when RF was used, optimal MAE values of 20.1 and 10.0 mm occurred for the continuous target of Int-0 for scenarios 1 and 2, respectively. Overall, scenario 2 had lower MAEs than scenario 1 did, regardless of whether C4.5 or RF was used. Figure 6c depicts the RMSE results for these scenario models. Similar to the MAE results, scenario 2 had lower optimal RMSEs (32.9 and 23.7 mm for C4.5 and RF, respectively) than scenario 1. A scatterplot of the observations and predictions from scenario 2 with different partitioning intervals obtained using the subset {A, B, S, R} is presented in Fig. 7; these results revealed that scenario 2, with the {A, B, S, R} data combination, can provide the most accurate predictions.
C4.5 and RF model results for scenario 1 of {A, B, R} and scenario 2 of {A,B,S,R}: (a) accuracy, (b) MAE, and (c) RMSE.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
Scatterplot of the observations and predictions for scenario 2 with different partitioning intervals: (a) Int-0, (b) Int-25, (c) Int-50, (d) Int-75, and (e) Int-100.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
According to the definitions of both improvement rates, the higher the IMAE and IRMSE values are, the more accurate the performance of the predicted outcomes is. Both performance metrics can range from 0% to 100%, and the improvement rates computed according to the results of all the scenario models in the present study are listed in Table 3. We also calculated the average performance metrics of various partitioning intervals for specific scenarios and models. By examining the IMAE of scenario 2, we determined that the average metric resulting from the use of RF (88.9%) was higher than that from the use of C4.5 (74.7%); a similar outcome was found for the IRMSE. Furthermore, we found that the optimal model (i.e., RF) was the same for scenarios 1 and 2; thus, the RF was suitable for scenarios 1 and 2.
Improvement rate of the scenario models using tenfold cross validation.
b. Testing typhoons
To validate the proposed real-time LTRF processes, we simulated the processes using historical typhoons. First, the collected typhoon data were reclassified into training and testing sets and the LTRF prediction models were rebuilt. Typhoon events that occurred between 2001 and 2013 were used for training, and those that occurred between 2014 and 2015 (specifically, Typhoons Matmo and Fung-Wong in 2014 and Soudelor and Dujuan in 2015) were used for testing. Figure 8 depicts the historical tracks of the testing typhoons. In addition, Table 4 lists the typhoon intensity categories, times at which TWO and TWOL alerts were issued by the CWB, and total rainfall at the Hualien Weather Station. In the table, the Saffir–Simpson hurricane wind scale (Simpson 1974) is used to define typhoon intensity.
Typhoon tracks of the four testing typhoons: (a) 2014 Matmo, (b) 2014 Fung-Wong, (c) 2015 Soudelor, and (d) 2015 Dujuan.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
Information on the testing typhoons.
c. Simulation results and evaluations
Figures 9, 10 illustrate the simulation results of scenarios 1 and 2 for the 2014 testing typhoons, respectively, and Figs. 11, 12 provide the simulation results of scenarios 1 and 2 for the 2015 typhoons, respectively. In the figures, the y axis represents the total precipitation amounts (observed and predicted). The observed values are constant because they are the sum of the amounts of rain during periods when specific typhoons landed in Taiwan. Furthermore, the x axis is referred to the 6-hourly time series. Lower time series values (earlier periods) at the x axis mean a typhoon is far away from Taiwan, and higher values (later periods) mean that the typhoon is approaching. In all four figures, the five panels show the simulation results using partitioning intervals Int-0, Int-25, Int-50, Int-75, and Int-100, respectively; the light green sections represent the CWB-reported range of total rainfall. Moreover, because the CWB nowcasts reported a range, they were replaced by the mean of the maximal and minimal values (i.e., the solid green lines). This allows us to compare the model simulations, CWB nowcasts, and observations in these figures.
Simulation results of scenario 1 using Typhoons Matmo and Fung-Wong. CWB-reported range of total typhoon rainfall (light green).
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
Simulation results of scenario 2 using Typhoons Matmo and Fung-Wong. CWB-reported range of total typhoon rainfall (light green).
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
Simulation results of scenario 1 using Typhoons Soudelor and Dujuan. CWB-reported range of total typhoon rainfall (light green).
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
Simulation results of scenario 2 using Typhoons Soudelor and Dujuan. CWB-reported range of total typhoon rainfall (light green).
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
By comparing the observations (blue lines) with the simulations (red triangles and black hollow points, which refer to the C4.5 and RF model predictions, respectively), we determined that some simulations overestimated or underestimated the observations, whereas others closely aligned with the observations. By comparing the CWB nowcasts with the simulations, we found that the model predictions were closer to the observations than were the CWB nowcasts. Moreover, we found that the model predictions of scenario 2 were more accurate than those of scenario 1.
We used the oceanic Niño index (ONI) to identify the ENSO episodes during which the four testing typhoons occurred. The ONI is the de facto standard that NOAA uses to identify El Niño (warm) and La Niña (cool) events in the tropical Pacific (Rojas et al. 2014), and it is composed of a 3-month running mean of the SST anomalies for the Niño-3.4 region. Events are defined as five consecutive overlapping 3-month periods at or above the 0.5°C anomaly for El Niño events, or at or below the −0.5°C anomaly for La Niña events. According to the ONI, Typhoons Soudelor and Dujuan developed during El Niño events, whereas Typhoons Matmo and Fung-Wong developed under normal conditions (i.e., between the 0.5° and −0.5°C anomalies).
Figures 13, 14 display the MAE and RMSE performance metrics, respectively. Each figure also contains the metrics of both scenarios for the C4.5 and RF models under normal conditions (i.e., Typhoons Matmo and Fung-Wong; Figs. 13a,b, 14a,b) and during El Niño events (i.e., Typhoons Soudelor and Dujuan; Figs. 13c,d, 14c,d). Notably, scenario 2 using RF with a continuous target of Int-0 produced the optimal MAE and RMSE results for normal and El Niño events. Table 5 presents the average improvement rates of various partitioning intervals computed according to their individual typhoon results. Specifically, we found that the average MAE and RMSE metrics in scenario 2 using RF were higher than in scenario 2 using C4.5 and in scenario 1 using C4.5 and RF in each testing typhoon; thus, scenario 2 yielded more accurate total rainfall estimates than scenario 1 did.
MAE metrics for (a) Typhoon Matmo, (b) Typhoon Fung-Wong, (c) Typhoon Soudelor, and (d) Typhoon Dujuan.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
RMSE metrics for (a) Typhoon Matmo, (b) Typhoon Fung-Wong, (c) Typhoon Soudelor, and (d) Typhoon Dujuan.
Citation: Journal of Atmospheric and Oceanic Technology 34, 10; 10.1175/JTECH-D-16-0216.1
Average improvement rate of the testing typhoons.
Here, we observe that the rainfall prediction of the model with ENSO indices is superior to its counterpart that lacks ENSO information during normal events (i.e., Matmo and Fung-Wong). This could be because the ENSO indices contain values that are higher (strong El Niño), slightly higher (weak El Niño), moderate (neutral event), slightly lower (weak La Niña), and lower (strong La Niña). When the decision tree algorithms are applied, the tree-based model could be built according to this additional information. Therefore, these tree-based models can be used to identify the intensity of ENSO and make informed decisions regarding rainfall. In addition, although the typhoons in the simulation do not contain La Niña events, we expect that the models that use scenario 2 will yield estimates more accurate than those of scenario 1. Thus, we suggest that the proposed real-time LTRF model process can effectively forecast total typhoon rainfall when the imported data include information on the state of ENSO.
5. Conclusions
Typhoons can greatly affect Taiwan, producing torrential rain and strong winds. This paper introduced the use of novel LTRF models to provide real-time total precipitation predictions several days prior to typhoon landfall, which would facilitate appropriate flood control measures and advanced disaster prevention. The LTRF models were formulated using scenario 1, which did not consider ENSO effects, and scenario 2, which did consider them. The scenario models were constructed using the tree-based algorithms C4.5 and RF. Using information that is currently employed in operational forecasting, the typhoon paths issued by the JTWC were collected to determine whether a typhoon might affect Taiwan. The Hualien Weather Station, located in eastern Taiwan, was selected as the study site. Typhoon records for 2001–15 were collected and the forecasting horizon was set at 6 h. To validate the use of the proposed real-time LTRF process, typhoon events that occurred during 2014–15 were simulated. Finally, the model simulations, observations, and CWB nowcasts were compared.
The simulation results showed that the proposed LTRF model that included ENSO effects can accurately forecast total typhoon rainfall when typhoons are still several days away from Taiwan. In other words, ENSO indices can improve prediction outcomes. We suggest that this is because ENSO is the most significant interannual signal associated with air–ocean interaction in the tropical Pacific; as the ENSO signal changes, the SST, thermocline, and atmospheric circulation patterns in the tropics likewise change (Teng et al. 2014). Therefore, the TC weather systems, including precipitation, are also influenced by ENSO. As indicated by Wang and Chan (2002), the average maximum intensity of TCs is higher and lower in El Niño and La Niña years, respectively, than in normal years. This implies the total potential precipitation received during a typhoon could depend on ENSO events. Thus, the SOI and Niño-3.4 ENSO indices can be used to provide advanced information regarding typhoons.
We also compared the C4.5 and RF models through simulation, with RF yielding more reliable results than C4.5. As has been noted by Breiman (2001) and Banfield et al. (2007), the primary advantage of RF, which uses bagging, is its ability to test the accuracy of an ensemble without removing data from the training set (as is done with a validation set). Oshiro et al. (2012) have also indicated that the RF algorithm is useful for classification, regression, and other tasks, because it constructs numerous decision trees and classifies new instances using a majority-vote method. According to our experiment, RF is more effective than C4.5 for predicting total typhoon rainfall.
The CWB of Taiwan issues TWOL alerts and total typhoon precipitation nowcasts 18 h prior to typhoon landfall. Comparing the CWB nowcasts with the simulations of decision trees, we found that the model predictions were closer to the observations than were the CWB nowcasts. This suggests that the LTRF model is more useful than the CWB ensemble for forecasting when a typhoon might land or affect a region within 18 h. As shown in the experimental results, the CWB total rainfall nowcasts can have high prediction errors. Therefore, we suggest that the results derived from the proposed LTRF model can provide an initial value for the operational ensemble forecasting model when the CWB issues real-time hourly nowcasts during approaching typhoons.
Acknowledgments
The support of the Ministry of Science and Technology in Taiwan (MOST104-2111-M-019-001) is greatly appreciated. Data were provided by NOAA, the JTWC, the CWB of Taiwan, and the GCOS Working Group on Surface Pressure. Wallace Academic Editing is acknowledged for editing this manuscript.
REFERENCES
Amit, Y., and D. Geman, 1997: Shape quantization and recognition with randomized trees. Neural Comput., 9, 1545–1588, doi:10.1162/neco.1997.9.7.1545.
Antolik, M. S., 2000: An overview of the National Weather Service’s centralized statistical quantitative precipitation forecasts. J. Hydrol., 239, 306–337, doi:10.1016/S0022-1694(00)00361-9.
Apté, C., and S. Weiss, 1997: Data mining with decision trees and decision rules. Future Gener. Comput. Syst., 13, 197–210, doi:10.1016/S0167-739X(97)00021-6.
Banfield, R. E., L. O. Hall, K. W. Bowyer, and W. P. Kegelmeyer, 2007: A comparison of decision tree ensemble creation techniques. IEEE Trans. Pattern Anal. Mach. Intell., 29, 173–180, doi:10.1109/TPAMI.2007.250609.
Bankert, R. L., and M. Hadjimichael, 2007: Data mining numerical model output for single-station cloud-ceiling forecast algorithms. Wea. Forecasting, 22, 1123–1131, doi:10.1175/WAF1035.1.
Barnston, A. G., 2015: Why are there so many ENSO indexes, instead of just one? Accessed 13 February 2017, https://www.climate.gov/news-features/blogs/enso/why-are-there-so-many-enso-indexes-instead-just-one.
Barnston, A. G., M. Chelliah, and S. B. Goldenberg, 1997: Documentation of a highly ENSO-related SST region in the equatorial Pacific. Atmos.–Ocean, 35, 367–383, doi:10.1080/07055900.1997.9649597.
Baudron, P., F. Alonso-Sarria, J. L. García-Aróstegui, F. Cánovas-García, D. Martínez-Vicente, and J. Moreno-Brotóns, 2013: Identifying the origin of groundwater samples in a multi-layer aquifer system with Random Forest classification. J. Hydrol., 499, 303–315, doi:10.1016/j.jhydrol.2013.07.009.
Belgiu, M., and L. Dragut, 2016: Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens., 114, 24–31, doi:10.1016/j.isprsjprs.2016.01.011.
Bessler, F. T., D. A. Savic, and G. A. Walters, 2003: Water reservoir control with data mining. J. Water Resour. Plann. Manage., 129, 26–34, doi:10.1061/(ASCE)0733-9496(2003)129:1(26).
Breiman, L., 2001: Random forests. Mach. Learn., 45, 5–32, doi:10.1023/A:1010933404324.
Breiman, L., J. Friedman, R. A. Olshen, and C. J. Stone, 1984: Classification and Regression Trees. Chapman & Hall, 368 pp.
Camargo, S. J., and A. H. Sobel, 2005: Western North Pacific tropical cyclone intensity and ENSO. J. Climate, 18, 2996–3006, doi:10.1175/JCLI3457.1.
Camargo, S. J., K. A. Emanuel, and A. H. Sobel, 2007: Use of a genesis potential index to diagnose ENSO effects on tropical cyclone genesis. J. Climate, 20, 4819–4834, doi:10.1175/JCLI4282.1.
Chan, J. C. L., 1985: Tropical cyclone activity in the northwest Pacific in relation to the El Niño/Southern Oscillation phenomenon. Mon. Wea. Rev., 113, 599–606, doi:10.1175/1520-0493(1985)113<0599:TCAITN>2.0.CO;2.
Chan, J. C. L., 2000: Tropical cyclone activity over the western North Pacific associated with El Niño and La Niña events. J. Climate, 13, 2960–2972, doi:10.1175/1520-0442(2000)013<2960:TCAOTW>2.0.CO;2.
Chang, C. P., Y. T. Yang, and H. C. Kuo, 2013: Large increasing trend of tropical cyclone rainfall in Taiwan and the roles of terrain. J. Climate, 26, 4138–4147, doi:10.1175/JCLI-D-12-00463.1.
Chen, X., and H. Ishwaran, 2012: Random forests for genomic data analysis. Genomics, 99, 323–329, doi:10.1016/j.ygeno.2012.04.003.
Chu, J. H., C. R. Sampson, A. S. Levine, and E. Fukada, 2002: The Joint Typhoon Warning Center Tropical Cyclone best-tracks, 1945-2000. U.S. Naval Research Laboratory Rep. NRL/MR/7540-02-16, 22 pp.
Collins, M., and Coauthors, 2010: The impact of global warming on the tropical Pacific Ocean and El Niño. Nat. Geosci., 3, 391–397, doi:10.1038/ngeo868.
Dougherty, J., R. Kohavi, and M. Sahami, 1995: Supervised and unsupervised discretizations of continuous features. Machine Learning: Proceedings of the Twelfth International Conference on Machine Learning, A. Prieditis and S. Russell, Eds., Morgan Kaufmann, 194–202, doi:10.1016/B978-1-55860-377-6.50032-3.
Elsner, J. B., and K. Liu, 2003: Examining the ENSO–typhoon hypothesis. Climate Res., 25, 43–54, doi:10.3354/cr025043.
Fan, C. W., 2011: Particles dynamics in a deep reservoir triggered by typhoons. J. Hydrol., 406, 82–87, doi:10.1016/j.jhydrol.2011.06.003.
Ferraro, R. R., and G. F. Marks, 1995: The development of SSM/I rain-rate retrieval algorithms using ground-based radar measurements. J. Atmos. Oceanic Technol., 12, 755–770, doi:10.1175/1520-0426(1995)012<0755:TDOSRR>2.0.CO;2.
Ferraro, R. R., F. Weng, N. C. Grody, and A. Basist, 1996: An eight-year (1987–1994) time series of rainfall, clouds, water vapor, snow cover, and sea ice derived from SSM/I measurements. Bull. Amer. Meteor. Soc., 77, 891–905, doi:10.1175/1520-0477(1996)077<0891:AEYTSO>2.0.CO;2.
Gislason, P. O., J. A. Benediktsson, and J. R. Sveinsson, 2006: Random Forests for land cover classification. Pattern Recognit. Lett., 27, 294–300, doi:10.1016/j.patrec.2005.08.011.
Grody, N. C., 1991: Classification of snow cover and precipitation using the Special Sensor Microwave Imager. J. Geophys. Res., 96, 7423–7435, doi:10.1029/91JD00045.
He, X., N. W. Chaney, M. Schleiss, and J. Sheffield, 2016: Spatial downscaling of precipitation using adaptable random forests. Water Resour. Res., 52, 8217–8237, doi:10.1002/2016WR019034.
Ho, T. K., 1995: Random decision forests. Proceedings of the Third International Conference on Document Analysis and Recognition, IEEE Computer Society Press, 278–282.
Hong, J. S., C. T. Fong, L. F. Hsiao, Y. C. Yu, and C. Y. Tzeng, 2015: Ensemble typhoon quantitative precipitation forecasts model in Taiwan. Wea. Forecasting, 30, 217–237, doi:10.1175/WAF-D-14-00037.1.
Kuo, Y. C., M. A. Lee, and M. M. Lu, 2016: Association of Taiwan’s October rainfall patterns with large-scale oceanic and atmospheric phenomena. Atmos. Res., 180, 200–210, doi:10.1016/j.atmosres.2016.05.012.
Kuo, Y. M., H. J. Chu, T. Y. Pan, and H. L. Yu, 2011: Investigating common trends of annual maximum rainfalls during heavy rainfall events in southern Taiwan. J. Hydrol., 409, 749–758, doi:10.1016/j.jhydrol.2011.09.015.
Lander, M. A., 1994: An exploratory analysis of the relationship between tropical storm formation in the western North Pacific and ENSO. Mon. Wea. Rev., 122, 636–651, doi:10.1175/1520-0493(1994)122<0636:AEAOTR>2.0.CO;2.
Lariviere, B., and D. V. den Poel, 2005: Predicting customer retention and profitability by using random forests and regression forests techniques. Expert Syst. Appl., 29, 472–484, doi:10.1016/j.eswa.2005.04.043.
Liaw, A., and M. Wiener, 2002: Classification and regression by randomForest. R News, No. 2 (3), R Foundation, Vienna, Austria, 18–22, http://CRAN.R-project.org/doc/Rnews/.
Lin, G. F., and B. C. Jhong, 2015: A real-time forecasting model for the spatial distribution of typhoon rainfall. J. Hydrol., 521, 302–313, doi:10.1016/j.jhydrol.2014.12.009.
Lo, D. C., C. C. Wei, and N. P. Tsai, 2015: Parameter automatic calibration approach for neural-network-based cyclonic precipitation forecast models. Water, 7, 3963–3977, doi:10.3390/w7073963.
Madsen, H., D. Lawrence, M. Lang, M. Martinkova, and T. R. Kjeldsen, 2014: Review of trend analysis and climate change projections of extreme precipitation and floods in Europe. J. Hydrol., 519, 3634–3650, doi:10.1016/j.jhydrol.2014.11.003.
Mahjoobi, J., and A. Etemad-Shahidi, 2008: An alternative approach for the prediction of significant wave heights based on classification and regression trees. Appl. Ocean Res., 30, 172–177, doi:10.1016/j.apor.2008.11.001.
Maier, H. R., and G. C. Dandy, 2000: Neural networks for the prediction and forecasting of water resources variables: A review of modeling issues and applications. Environ. Modell. Software, 15, 101–124, doi:10.1016/S1364-8152(99)00007-9.
Michaelides, S., V. Levizzani, E. Anagnostou, P. Bauer, T. Kasparis, and J. E. Lane, 2009: Precipitation: Measurement, remote sensing, climatology and modeling. Atmos. Res., 94, 512–533, doi:10.1016/j.atmosres.2009.08.017.
Mishra, A., R. M. Gairola, A. K. Varma, A. Sarkar, and V. K. Agarwal, 2009: Rainfall retrieval over Indian land and oceanic regions from SSM/I microwave data. Adv. Space Res., 44, 815–823, doi:10.1016/j.asr.2009.05.011.
Neelin, J. D., D. S. Battisti, A. C. Hirst, F. F. Jin, Y. Wakata, T. Yamagata, and S. E. Zebiak, 1998: ENSO theory. J. Geophys. Res., 103, 14 261–14 290, doi:10.1029/97JC03424.
Oshiro, T. M., P. S. Perez, and J. A. Baranauskas, 2012: How many trees in a random forest? Machine Learning and Data Mining in Pattern Recognition, Lecture Notes in Computer Science, Vol. 7376, Springer, 154–168, doi:10.1007/978-3-642-31537-4_13.
Quinlan, J. R., 1979: Discovering rules by induction from large collection of examples. D. Michie, Ed., Expert Systems in the Micro-Electronic Age, Edinburg University Press, 168–201.
Quinlan, J. R., 1993: C4.5: Programs for Machine Learning. Morgan Kaufmann, 302 pp.
Rasmusson, E. M., and T. H. Carpenter, 1982: Variations in tropical sea surface temperature and surface wind fields associated with the Southern Oscillation/El Niño. Mon. Wea. Rev., 110, 354–384, doi:10.1175/1520-0493(1982)110<0354:VITSST>2.0.CO;2.
Raytheon, 2000: SSM/I user’s interpretation guide. UG32268-900, Revision C, 96 pp.
Rojas, O., Y. Li, and R. Cumani, 2014: Understanding the drought impact of El Niño on the global agricultural areas: An assessment using FAO’s Agricultural Stress Index (ASI). Environment and Natural Resources Management Series 23, Food and Agriculture Organization of the United Nations, 42 pp.
Ropelewski, C. F., and P. D. Jones, 1987: An extension of the Tahiti–Darwin Southern Oscillation index. Mon. Wea. Rev., 115, 2161–2165, doi:10.1175/1520-0493(1987)115<2161:AEOTTS>2.0.CO;2.
Salinger, M. J., 1994: Climate variability, agriculture and forestry. AGM Tech. Note 196, WMO-802, 152 pp.
Salinger, M. J., C. J. Stigter, and H. P. Das, 2000: Agrometeorological adaptation strategies to increasing climate variability and climate change. Agric. For. Meteor., 103, 167–184, doi:10.1016/S0168-1923(00)00110-6.
Schneider, A., 2012: Monitoring land cover change in urban and peri-urban areas using dense time stacks of Landsat satellite data and a data mining approach. Remote Sens. Environ., 124, 689–704, doi:10.1016/j.rse.2012.06.006.
Scofield, R. A., and R. J. Kuligowski, 2003: Status and outlook of operational satellite precipitation algorithms for extreme-precipitation events. Wea. Forecasting, 18, 1037–1051, doi:10.1175/1520-0434(2003)018<1037:SAOOOS>2.0.CO;2.
Scornet, E., 2016: On the asymptotics of random forests. J. Multivar. Anal., 146, 72–83, doi:10.1016/j.jmva.2015.06.009.
Simpson, R. H., 1974: The hurricane disaster: Potential scale. Weatherwise, 27, 169–186, doi:10.1080/00431672.1974.9931702.
Sun, W., J. Chen, and J. Li, 2007: Decision tree and PCA-based fault diagnosis of rotating machinery. Mech. Syst. Signal Process., 21, 1300–1317, doi:10.1016/j.ymssp.2006.06.010.
Teng, H. F., C. S. Lee, and H. H. Hsu, 2014: Influence of ENSO on formation of tropical cloud clusters and their development into tropical cyclones in the western North Pacific. Geophys. Res. Lett., 41, 9120–9126, doi:10.1002/2014GL061823.
Trenberth, K. E., 1984: Signal versus noise in the Southern Oscillation. Mon. Wea. Rev., 112, 326–332, doi:10.1175/1520-0493(1984)112<0326:SVNITS>2.0.CO;2.
Trenberth, K. E., and D. P. Stepaniak, 2001: Indices of El Niño evolution. J. Climate, 14, 1697–1701, doi:10.1175/1520-0442(2001)014<1697:LIOENO>2.0.CO;2.
Wang, B., and J. C. L. Chan, 2002: How strong ENSO events affect tropical storm activity over the western North Pacific. J. Climate, 15, 1643–1658, doi:10.1175/1520-0442(2002)015<1643:HSEEAT>2.0.CO;2.
Wang, C. C., S. Y. Huang, S. H. Chen, C. S. Chang, and K. Tsuboki, 2016: Cloud-resolving typhoon rainfall ensemble forecasts for Taiwan with large domain and extended range through time-lagged approach. Wea. Forecasting, 31, 151–172, doi:10.1175/WAF-D-15-0045.1.
Wei, C. C., 2012: Discretized and continuous target fields for the reservoir release rules during floods. Water Resour. Manage., 26, 3457–3477, doi:10.1007/s11269-012-0085-2.
Wei, C. C., 2013: Improvement of typhoon precipitation forecast efficiency by coupling SSM/I microwave data with climatologic characteristics and precipitation. Wea. Forecasting, 28, 614–630, doi:10.1175/WAF-D-12-00089.1.
Wei, C. C., 2014: Simulation of operational typhoon rainfall nowcasting using radar reflectivity combined with meteorological data. J. Geophys. Res. Atmos., 119, 6578–6595, doi:10.1002/2014JD021488.
Wei, C. C., and N. S. Hsu, 2008: Derived operating rules for a reservoir operation system: Comparison of decision trees, neural decision trees and fuzzy decision trees. Water Resour. Res., 44, W02428, doi:10.1029/2006WR005792.
Wei, C. C., and N. S. Hsu, 2009: Optimal tree-based release rules for real-time flood control operations on a multipurpose multireservoir system. J. Hydrol., 365, 213–224, doi:10.1016/j.jhydrol.2008.11.038.
Wei, C. C., and J. Roan, 2012: Retrievals for the rainfall rate over land using Special Sensor Microwave/Imager data during tropical cyclones: Comparisons of scattering index, regression, and support vector regression. J. Hydrometeor., 13, 1567–1578, doi:10.1175/JHM-D-11-0118.1.
Wei, C. C., G. J. Y. You, L. Chen, C. C. Chou, and J. Roan, 2015: Diagnosing rain occurrences using passive microwave imagery: A comparative study on probabilistic graphical models and “black box” models. J. Atmos. Oceanic Technol., 32, 1729–1744, doi:10.1175/JTECH-D-14-00164.1.
Wu, C. L., and K. W. Chau, 2013: Prediction of rainfall time series using modular soft computing methods. Eng. Appl. Artif. Intell., 26, 997–1007, doi:10.1016/j.engappai.2012.05.023.
Wu, C. L., K. W. Chau, and C. Fan, 2010: Prediction of rainfall time series using modular artificial neural networks coupled with data preprocessing techniques. J. Hydrol., 389, 146–167, doi:10.1016/j.jhydrol.2010.05.040.
Zhang, T., D. Xia, H. Tang, X. Yang, and H. Li, 2016: Classification of steel samples by laser-induced breakdown spectroscopy and random forest. Chemom. Intell. Lab. Syst., 157, 196–201, doi:10.1016/j.chemolab.2016.07.001.
