ResponsetoReviewerComments:回答审稿人的意见

Response to Reviewer Comments

We thank both the reviewers for their thoughtful/useful comments and suggestions. Their

comments have improved the manuscript effectively. We have included almost all of

their suggestions and below we present a point-by-point response to their comments.

Reviewer A

General Comments

1. Comment on assumptions of linear regression, using a linear regression as opposed to

other nonlinear models like artificial neural network, nonlinear regression etc..?

We have checked the distribution of the predictors, and we can report that they are all

Normally distributed (figure not shown), so is the Thailand summer monsoon rainfall.

Thus, the key assumption of Normal distribution for Linear Regression is satisfied.

Neural network and nonlinear regression models require large sample sizes. While the

sample size in this research is relatively small for LOCFIT it does not suffer to the same

extent as other nonlinear models. Furthermore, LOCFIT, being “local” in nature has the

capability to capture any feature (linear or nonlinear) present in the data.

We found strong linear correlation between the summer rainfall and its predictors (Table

1). Hence, Linear Regression model was used as a benchmark – besides, it is one of the

most popular methods in practice.

2. Why CCA type models were not considered as a benchmark..

We thank the reviewer for pointing the two references on CCA, which we have included

in the narrative.

CCA type models are better suited for predicting a dependent field (i.e. rainfall at several

stations) from field(s) of independent variables (e.g., Tropical SST, SLP etc.). In this

paper we are predicting a single time series (i.e. the Thailand summer rainfall index)

hence regression based models, such as the ones used here are apt.

3. Issue of non-stationarity….

We agree, that if the relationship between the Thailand summer rainfall and ENSO and

other Indo-Pacific predictors changes in time then new predictors have to be identified.

As shown, this relationship is seen only in the post-1980 period hence the forecasting

models have some success in this period.

4. Comment on the

We appreciate the reviewers point about multiple sources of uncertainty. This is beyond

the scope of this research. In the approaches proposed here, model uncertainty is captured.

If the predictors capture the physical relationship with the rainfall then the system

uncertainty too will be captured. While the ensembles are generated using some form of

Monte Carlo, but they are ensembles, nonetheless. We wish to clarify that the statistical

models used in this work for ensemble prediction should be distinguished from the

ensemble techniques adopted using general circulation models (GCMs).

5. Predictor – rainfall relationship…

As can be seen from (Figures 1,2) and Table 1 the large-scale climate (i.e, tropical ocean-atmospheric variables) and Thailand summer rainfall show relationship only in the post-1980 period. This epochal behavior of the relationship is explained by shifts in the ENSO

features explained in detail in our paper (Singhrattna, et al., 2004). Because we devoted

that paper entirely in explaining the decadal/inter-decadal variability of Thailand rainfall,

we focused this paper purely on developing tools for forecasting the Thailand summer

rainfall.

Minor Comments

Modify the title to “Seasonal forecasting of Thailand Summer Monsoon Rainfall”

We like the suggestion and have modified the title accordingly.

Provide Key Words

Key words have been provided at the end of the abstract

Table 1: How does the change in correlation between SOI and summer rainfall from -0.44 for MJJ SOI to 0.45 JJA SOI affect the role of SOI as a predictor.

It is a typo and we apologize for the same. The SOI and rainfall correlation for

MJJ is +0.44 and not -0.44 (as shown in the Table). We have corrected this.

Furthermore, SOI did not enter into the final set of predictors so in that sense it

did not impact the forecasts presented.

Instead of providing a website for IOD, which some readers may not be familiar with,

why not provide the basic information such as equation and the data type and domain

used to compute IOD? On the other hand, is IOD that useful as a predictor, given Figure

1 shows that the correlation between MAM IOD and ASO rainfall decreased

monotonically since 1960s?

IOD index is computed as SST anomaly difference between Eastern and Western

tropical Indian Ocean. The details of the dataset, regions, the physical ignificance,

etc. are described in detail in the Saji et al. (1999) paper, which we have referred.

Our aim here, as a first step, was to compute the correlation between Thailand

summer rainfall and all the standard tropical Indo-Pacific indices. Furthermore, as

the reviewer noted, the IOD index was not a useful predictor in the final set of

predictors that were selected. In fact, the SST index that was used as a predictor

covers part of the IOD region.

We have added a couple of sentences on the IOD at the end of Section 2.

For LOCFIT, what order of polynomial equations was used in the seasonal prediction of

the summer rainfall of Thailand? Why not represent a polynomial equation and state

what orders were mostly used?

We used only local „linear‟ polynomials. We have mentioned this at the end of

Model Evaluation Section. Typically, local linear or quadratic works best – of

course, the polynomial order can also be selected using the GCV criteria. In this

research, given the small sample size we fixed the order of the polynomial to be 1

(i.e. linear) – but the neighborhood size (alpha) was obtained objectively using the

GCV criteria. The equation for the GCV criteria is now given. The “local” aspect

of the method is what provides the rich capability to capture any arbitrary

functional form exhibited by the data.

Table 4: why the non-exceedance probabilities for 1987 were all 0%

This means that all the ensembles from the methods, especially LOCFIT and

Linear regression are well to the right of the observed (i.e. all the ensemble

members exceed the lower threshold) . This means that the non-exceedance

probability is zero. Note that these are forecasts issued on April 1st and hence,

likely to be of lesser skill, as can be seen in Figure 5b.

Some color plots shown in Figure 2 are too small to be readable. Enlarge the plots. In

contrast, Figure3 can be reduced.

We have re-generated all the figures eliminating the above mentioned

shortcomings.

Figures 6& 7: Labels should be provided to the pdfs plotted. The authors explained that

700mm (90th percentile) is chosen to represent wet conditions. In this arbitrarily chosen,

given that it is only a 10-year return period flood? I presume the light dotted curve

represents the climatology pdf in Page 17? What is a climatological pdf? Please explain.

As mentioned above, we have re-generated the figures. The figure captions

explain the figures better. Now, it is the dashed line which represents the

climatological PDF and the solid line is that from the ensembles. The dotted line

is the actually observed value. Climatology PDF is one that is computed on all of

the historical data. We have clarified this in above mentioned section.

Most equations should be re-typed

We have re-typed the equations and made the symbols consistent, throughout the

paper.

There are typos appearing randomly in the paper.

We have checked for typos/grammatical errors carefully and have eliminated all

of them.

Reviewer B

General Comments

1. Labeling throughout paper needs to be consistent (sometimes “LOCFIT” sometimes

“Normal K-NN”)

This was the case, especially in the figures. We have now made this ,

referring only to LOCFIT

2. Much information is presented doubly ( in tables and figures) – there is potential to

reduce somewhat here. Also, as detailed below, the authors can fold SST and SAT into a

single predictor that will be better to apply than the two they currently show.

We fully agree with the reviewer‟s suggestion and as a result we have removed Tables 2

and 3 since the information provided here is also available through Figures 4 and 5,

respectively. However, we retained Table 1 and Figure 1. Table 1 shows the correlation

between all the indices and Thailand summer rainfall for all the seasons (including the

summer season). While in Figure 1, we only show moving window correlation of four

indices for just one season.

We agree that SST and SAT index can be folded into a single predictor. In fact, in the

final set of predictors in the forecast models only SST is included – this is due to the fact

that both these indices have significant information in common.

Specific Comments

1. Mid p. 3: The authors refer to the lack of literature regarding specifically the monsoon

over Thailand, ……. Currently GAME is only mentioned as a source of some of the data

sets used in the study.

We do recognize that the Thailand monsoon is part of a larger Austral-Asian monsoon

system. However, the variability of Thailand summer rainfall is unique. Besides, the

predictability and the large body of understanding of the Austral-Asian monsoon system

are not of much help if it cannot be specifically used to forecast the Thailand rainfall. In

our paper we demonstrate for the first time the potential for predicting Thailand summer

rainfall.

We are thankful to the GEWEX/GAME effort for the data and have mentioned the same

in the acknowledgments. We are aware of the GEWEX/GAME efforts to forecast flows

in the Chao Phraya basin of Thailand. However, all of these efforts involve (a) short term

flow forecast (i.e. days to weeks) and, (b) using watershed models. None of the efforts, to

our knowledge (looking at the publications on the GAME website) have focused on

forecasting seasonal Thailand rainfall or streamflows. We do refer to two key papers (Jha

et al., 1997 and 1998) on the hydrologic predictions in the Chao Phraya basin.

2. Sec 2, data set 1: There should be a map of the locations of the rainfall stations used in

the statistical regressions.

Map showing the location of all the stations was provided in our Singhrattna et al. (2004)

paper and also in Singhrattna (2003). We do agree with the reviewer‟s suggestion. So as

not to increase the number of figures we have provided the latitude and longitude of the

three stations used to obtain the Thailand summer rainfall and temperature (SAT) index.

3. Data set 2: Were monthly means used?

Yes

4. Sec 2: a limited list of data sets are given, without discussion of why these were chosen

over others (e.g., why not use GPCP or CRU gridded rainfall?). What determined the

choice of these data?

Since we had observed station data, we feel it is likely to be better than GPCP or CRU

which are gridded data. Furthermore, the observed rainfall is highly correlated to GPCP

data (over 0.7 in the Chao Phraya region), as we showed in Singhrattna et al. (2004,

Figure 2) and Singhrattna (2003). Thus, the results in our paper will be insensitive to the

above choice of the data sets.

5. Sec 3: Kanae et al. (200; J. Hydromet.)…….., that other sources of trends such as land

use change or global warming may make this a non-stationary process, and thus degrade

their linear statistical relationships?

We thank the reviewer for the references, some of which we were not aware of. We have

included the two relevant references at the end section 3.1. As the reviewer mentions, all

the studies in the references mentioned are in the general South Asian region but not

necessarily over Thailand and are from limited modeling studies. We do agree that land

cover changes can degrade linear relationships between the Thailand summer rainfall and

ocean-atmospheric features. But there isn‟t enough land-cover related data to quantify

this effect. We are working on a just funded grant to precisely investigate this issue.

6. Table 1: why is there a big sign change in the MJJ relationship with SOI?

Reviewer A too pointed this out. This is a typo and we have fixed it. The SOI and rainfall

correlation for MJJ is +0.44 and not -0.44 (as shown in the Table).

7. Fig 1: Interesting – how does this compare with the relationships found in other

decadal ENSO-monsoon studies (e.g., Miyakoda et al., 2003; J. Meteor. Soc Japan)?

It is interesting and there are some similarities – we have included this reference.

8. Sec 3.2: Over the subtropics and tropics, essentially SAT=SST (similarity in Figs 2a

and 2c). Also, the SAT from the NCEP reanalysis is dubious over land…. This would

reduce SST and SAT to simply “surface temperature” and reduce the number of figures.

We agree with the reviewer that SAT from the NCEP reanalysis can be dubious over land

and consequently, we have removed Figure 2(a). In fact, we used the observed land

temperatures from the three stations as the SAT predictor index – as such Figure 2(a) is

redundant. We also agree that CRU and CAMS could be used to better investigate the

land temperature relationship – especially, over the larger Eurasian region.

9. Sec 3.3 There is an inconsistency here. The authors show that the relationships to

monsoon rainfall are not constant over long periods, … Is it really a viable operational

prediction methodology?

We submit (and hopefully have demonstrated) that the proposed approaches, especially

the nonparametric methods will serve as an effective tool for Thailand rainfall. We agree

with the reviewer about the non-stationary aspect of the predictor – rainfall relationship.

This is something that we are seeing in other parts of the globe and will have to contend

with. To guard against this, we suggest checking the predictor-rainfall relationship

periodically and if relationships have weakened and new predictors will have to be

identified.

Understanding the decadal variability of the rainfall and the seasonal prediction are two

clear and separate goals that could be related or may be not. The former, we address in

great detail in Singhrattna et al., (2004) and Singhrattna (2003).

10. Why are the errors (for ensembling) not generated and added separately for each

predictor, but instead added to the mean estimate?

We are not clear about the reviewer‟s question. We assume the reviewer meant to say „for

each method‟. If so, then the errors generated for a given model are „specific‟ to that

model and are a result of the error formulation in that model – hence, errors from one

model cannot be added to the mean forecast of another.

11. Sec 4.2, 1st para: All this description is very elementary – is it necessary for such a

paper?

We feel that the 1st para provides continuity with the Linear regression discussed in the

preceding section. Furthermore, it is a short para and does not add to the length of the

paper.

12. Sec 5, para 1: Cross validation ensures the inability to forecast extremes. In this

sense, it unduly penalizes the method.

Yes. But this is the best way to estimate the predictive capability of the methods. We

would also add that cross-validation does not lead to inability in forecasting extremes –

because the methods (LOCFIT and Linear regression) fit polynomials and hence, can

extrapolate. In fact, the ability to forecast extremes will depend largely on the ability of

the predictors to provide useful information on the extremes.

13. p. 16: Two sections numbered “5”. Exactly how many cases in the training set (state

earlier than p. 17)?

We have corrected the section numbering (Reviewer A too mentioned this).

Since we have a very small sample size (i.e., 22 values, 1980 – 2001) we evaluated the

model skills in a cross-validated mode. In that, a value is dropped and the model fit to the

rest of the data and the dropped value is predicted. So, all the skills shown in the paper

are cross-validated skills. This we describe in section 5 (“model evaluation”).

14. Table 2: Please report the significances for r.

We think the reviewer meant Table 1 where we show the correlation ( r ) between

Thailand summer monsoon rainfall and the large-scale climate indices. The 95%

significant level is +/- 0.41. We have mentioned this in the Table caption.

15. How can the skills of linear regression and LOCFIT be compared? On what basis are

they called “similar”?

The skill measures used (correlation, LLH and RPSS) are model independent. These

measures capture the ability of the model to capture various distributional properties.

Hence, it is valid to compare the models on these measures.

We mention that the linear regression and LOCFIT exhibit “similar” skills from Figure 4

where the skills from the two models are generally close for the most part. However, for

the extreme years (Figure 5) the nonparametric models do much better.

16. Fig 5 & Table 3: There is a big disparity in skill between wet and dry years. Dry year

skill is poor, suggesting the cause is not reflected in SST or SLP, but something else.

Please discuss this disparity.

There is reduction in skill in the dry years relative to wet years – but not by a large

magnitude as the reviewer suggests. The disparity could be due to differences in ENSO

flavor or nonlinearity in the ENSO teleconnection. We are not sure at this point.

17. Fig 6 & 7: It is very difficult to read the dotted line and labels.

We have re-generated the figures clearing up this difficulty. The dashed curve is the

climatological PDF, the solid line is that from the ensembles and the dotted line is the

observed value.

18. Table 4 and mid p. 18: In each set of 3 extremes, one forecast is a bust. The 1-out-3

failure rate is the kind that can jeopardize governments!

We agree. Further investigation is required to sort this out. Our paper offers a first step in

this direction.