Challenges for Science and Governance, Berlin: Springer, Heikkila, T. Molnar, K. Pohl, B. Bach H.
Table of contents
Connect with us Never miss a thing! Sign up for our newsletter and stay up-to-date Send E-newsletter monthly. Because accurate precipitation data can improve runoff estimation, for comparison we used precipitation data sets from two different sources to calculate runoff. This data set takes advantage of multiple sources from gauges, satellites, and atmospheric reanalysis models. Average annual precipitation is calculated using data for — In this study, 28 independent variables concerning topography, climate, land cover, and soil fraction were used to predict Rc.
The data set used in model training and validation only covered data only within China, whereas this study aimed to apply the model to the larger CTRB riparian countries. Because countries outside China may have different climate regimes, terrains, land covers, and soils, we evaluate the representativeness of the data from within China. A higher OVL indicates a higher concordance between the two distributions. The average OVL was more than 0. Most of the independent variables had high OVLs. Among the climate and terrain factors, the aridity index, wind speed, and standard deviation of slopes had high OVLs greater than 0.
The aridity index, potential evapotranspiration, precipitation, average temperature, and slope had similar distributions in samples from China and the CTRB riparian countries, whereas the elevation had quite different ranges and distributions. We used k fold cross validation is used to estimate the model generalization performance. In this study, k was All the samples for model development were divided into tenfold, and each fold was then used once as the validation set, with the rest as the training set.
Comparatively, the machine learning models have straighter and narrower clouds tightly surrounding their diagonal lines, indicating that the Rc estimates were more accurate. In addition to their varying accuracy, the models provided variable importance information that measured the importance of independent variables. In LM, standardized regression coefficients are used to measure variable importance. Precipitation and elevation were positively correlated with Rc, and temperature was negatively correlated.
After removing collinearity, the rankings of variable importance varied greatly. The GB and RF models showed similar variable importance. However, there was a great difference in ranking between LM and the machine learning models. One reason lies in the different meanings of variable importance between LM and the machine learning models. In RF and GB, variable importance was determined by calculating the relative influence of each variable, which was based on the variable selection counts during splitting and the squared error over all trees improvement to the model as a result of each split.
On the other hand, the standardized regression coefficients in LM were highly affected by collinearity between the predictors. To examine the relations between Rc and its explanatory variables, we analyzed the partial dependence of Rc on the major explanatory variables in RF. The average Rc tended to increase when higher aridity index, precipitation, standard deviation of elevation, or standard deviation of slope was used. There was no obvious trend between Rc and elevation or solar radiation, indicating that elevation and solar radiation were not good explanatory variables in RF.
After the models were cross validated, they could be used to estimate Rc values for all the CTRB riparian countries using the covariates of this region. The Rc maps estimated from machine learning models show very similar spatial patterns, but the spatial heterogeneity of Rc from LM is less notable in some regions. By multiplying gridded precipitation products, we could convert Rc maps to runoff depth maps, which represent the volume of runoff discharged from a specific area expressed as an equivalent depth over that area.
For all models, runoff depth presented similar spatial patterns that correspond well with the runoff coefficient. Another way is to conduct a quantitative assessment of the results from these models. Water volume or water resources of all CTRB countries was calculated by multiplying runoff depth by the area of the corresponding country. Clearly, the machine learning models generally outperformed the traditional LM.
Interestingly, the model estimates derived from MSWEP precipitation generally outperformed those derived from WorldClim, even though WorldClim precipitation was used in model training. The best coefficient of determination R 2 was 0. The lowest RMSE was Recent estimates indicate that Afghanistan has Rc was found to be greater than 0. Overall, the results of the present study are consistent with those of other studies. The performances of the three machine learning models used in this study were similar, with the stacking model being the best by a small margin.
All the results indicate that machine learning is highly effective and eminently suitable for Rc prediction and water resources estimation in ungauged basins. Several studies have applied machine learning techniques to various disciplines and have indicated the effectiveness of machine learning for prediction e. The effectiveness of machine learning has also been proved by many previous studies.
Its disadvantage with respect to traditional models has to do with understanding and interpreting the model. The machine learning models not only predicted Rc with high accuracy but also provided information on variable importance. This is also consistent with previous studies Beck et al. The three most important factors were aridity index, precipitation, and standard deviation of slope, representing the controlling role of climate and topographical factors in the water cycle.
These factors were followed by elevation, solar radiation, and standard deviation of elevation. Gudmundsson and Seneviratne found that inclusion of land parameters soil texture and slope in their study did not improve runoff estimates.
- Crimson Eve.
- Water Resources Assessment of China's Transboundary River Basins Using a Machine Learning Approach.
- Recent advances on elliptic and parabolic issues: Proc. of Swiss-Japanese seminar?
- Mental Maps.
- Susanne Schmeier | IHE Delft Institute for Water Education.
However, that inference was probably incorrect because standard deviation of slope was not considered in their study. The variables of land cover and soils were generally ranked lower than others, which also agrees well with Merz et al. Similarly, Oudin et al. That is possibly due to spatial scale effects. From plot scale to catchment scale, the effects of land cover and soils tend to decrease Cerdan et al. According to Rogger et al. The present study used a spatial resolution of 0. The fact that higher estimation accuracy for runoff and water resources was achieved from models using MSWEP precipitation e.
However, precipitation is one of the most important features mentioned in previous sections, which seems paradoxical at first glance. The following considerations may clarify this. As mentioned, precipitation plays an important role in runoff and water resources estimation. Three of the precipitation products were introduced in section 3. Using remote sensing precipitation data to estimate runoff is promising. Hong et al. The performance of TRMM in our study was also not satisfactory and requires improvement. In this study, the machine learning models were trained and validated using a data set from within China and then applied to all the CTRB riparian countries.
This may have risked extrapolation in prediction. Although China is a large country with an area of 9.
To evaluate the reliability of extrapolation reliability, we did a representative analysis and a k fold cross validation. Overall, the data set from within China had good representativeness for model development and application. Although there was a marginal difference in the probability density distribution of elevation, the influence was limited due to its less important role compared with those of the aridity index, precipitation, and standard deviation of slope. Hydroclimatic conditions, however, have changed greatly worldwide due to climate change and human activity during the past century.
Climate factors have a great impact on water resources. The climate and topographical factors are the most important variables in predicting the runoff coefficient, indicating their great potential effects on hydrological processes. The runoff coefficient tends to increase when precipitation or the aridity index increases, which is in turn exaggerated in runoff and water resources.
For a large area, the topographical factors are normally considered stationary. However, for a small area, the landscape may shift greatly because of geological and human activities. There are three main justifications for ignoring the effect of nonstationarity. First, the focus of this study was the multiyear averaged runoff and runoff coefficient, so the averaged effect of multiyear may have weakened the effect of nonstationarity. Second, because nonstationarity has a very large spatial variability, it is difficult to treat and express in this first tentative study, which we plan to repeat later to more precisely estimate runoff.
These maps have greater spatial resolution and accuracy than those of other products. This study set the stage for using machine learning techniques in transboundary water resources assessment.
Progress on Transboundary Water Cooperation: Global baseline for SDG indicator 6.5.2
Before the maps and results are used, some points should be clarified. Moreover, the results for water resources in the CTRB should be carefully used because the basin boundaries may differ from those in other literature. Although the results of using machine learning to assess China's transboundary water resources assessment are satisfactory and are mostly in agreement with previous work, there are some limitations, and there is room for future improvement.
The following are some subjects for future study. Several studies have done this kind of research Carlisle et al. Water resources assessment in a changing climate.
Sustainable Human Development
Previous sections indicate that climate factors are important in water resources estimation. This study did not consider environmental change and the nonstationarity of runoff, but in the past decades, the climate has changed substantially due to human activities. Inevitably, climate change will have an impact on runoff generation. Thus, more studies should be made to consider the impact of nonstationarity hydrologic conditions on water resources assessment. More accurate and worldwide information.
Journal list menu
Although high accuracy could be achieved in predicting runoff coefficients using machine learning, inaccurate precipitation data limit the accuracy of runoff and water resources estimation. Thus, more accurate information is required. Also, only data from China were used for model development, raising the question of generalizability.
Collecting worldwide runoff information is recommended. We thank all the organizations, institutes, governments, and universities for providing the valuable data sets. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries other than missing content should be directed to the corresponding author for the article. Volume 55 , Issue 1. If you do not receive an email within 10 minutes, your email address may not be registered, and you may need to create a new Wiley Online Library account.
If the address matches an existing account you will receive an email with instructions to retrieve your username.
- Transboundary waters | International Decade for Action 'Water for Life' !
- 5. Transboundary Waters;
- Logic of Internationalism: Coercion and Accommodation.
- Transboundary waters.
- Bibliographic Information.
- Professional Portrait Posing: Techniques and Images from Master Photographers (Pro Photo Workshop);
- RIVER BASINS COMPONENT!
Open access. Water Resources Research Volume 55, Issue 1. Research Article Free Access. Shaofeng Jia Corresponding Author E-mail address: jiasf igsnrr. Jia, E-mail address: jiasf igsnrr. Tools Request permission Export citation Add to favorites Track citation. Share Give access Share full text access. Share full text access. Please review our Terms and Conditions of Use and check box below to share full-text version of article. Abstract A comprehensive and reliable assessment of the water resources in China's transboundary river basins is vital for water resources management and peaceful development.
Figure 1 Open in figure viewer PowerPoint. Figure 2 Open in figure viewer PowerPoint. Figure 3 Open in figure viewer PowerPoint. Surface plot of random forest RF grid search results. Runoff coefficient Rc is in the range of 0 to 1. Figure 4 Open in figure viewer PowerPoint. Figure 5 Open in figure viewer PowerPoint. Pearson correlation matrix of independent variables for clarity, land use and soils are not shown here.
Figure 6 Open in figure viewer PowerPoint. Ranges and distributions of six major predictors in China and in China's transboundary river basins riparian countries. All variables are reported in their native units. Rc is in the range of 0 to 1. Figure 7 Open in figure viewer PowerPoint. The x axis is the surveyed runoff coefficient Rc from the Ministry of Water Resources of China MWR, , and the y axis is the predicted Rc of cross validation from the models.
The solid line in each figure represents Figure 8 Open in figure viewer PowerPoint. Relative variable importance in a linear model, b linear model collinearity removed , c gradient boosting, and d random forest. For the linear model, standardized regression coefficients were used as the variable importance measurements yellow bars denote negative coefficients. For gradient boosting and random forest, the variable importance values are scaled to 0—1. Figure 9 Open in figure viewer PowerPoint. Partial dependence plots of runoff coefficients of six major explanatory variables in random forest.
Figure 10 Open in figure viewer PowerPoint. Runoff coefficient estimates for China's transboundary river riparian countries from a linear model LM , b gradient boosting GB , c random forest RF , and d stacking model ES. Figure 11 Open in figure viewer PowerPoint. WC is WorldClim. Bold numbers in the table denote the best result for each criterion. Figure 12 Open in figure viewer PowerPoint. WR is water resources.
Note that the scales on both axes are logarithmic. Amur China River basin This study km 3 Other sources km 3 Amur Irrawaddy Salween Mekong Pearl Yarlung Tsangpo Brahmaptra Ganges Indus Red River Ob Tarim Figure 13 Open in figure viewer PowerPoint. Surveyed runoff depth is from MWR The solid line in each figure represents , and dashed lines represent the fitted linear regression. Abbaspour, K.