Mapping global hotspots and trends of water quality (1992–2010): a data driven approach

We use water quality data from GEMStat which is a globally harmonized database on freshwater quality developed by United Nations Environmental Programme - Global Water Quality (UNEP-GEMS), maintained by the International Centre for Water Resources and Global Change and hosted by the Federal Institute of Hydrology in Koblenz. Raw data for the six water quality indicators are mapped in supplementary information 1. As many observations are not correctly encoded in GEMstat, we clean the raw data to exclude outliers, including observations flagged as 'suspect' by GEMStat. Furthermore, we removed observations with abnormal values regarding the properties of the pollutant (ex: pH > 14, Tw = 0 °C in the tropical band) ⁶ , and excluded observations that are abnormal regarding the distribution of all observations in a given station and were considered as outliers (i.e. we excluded values x_i,t when in station i, $\left| {{x_{i,t}}} \right|$ >[x⁻ _i,t + 2.5 σ_x ], where x⁻ _i,t represents the average value of all x_i,t in station i at tim t and σ_x represents the standard deviations of all x_i,t in station i. The R² of the final models are not sensitive to the exclusion of these outliers while the root mean square error (RMSE) of predictions decreases. Finally, some countries are overrepresented in our sample (Brazil for DO, TP and Tw; New Zealand for EC; South Africa for NOxN and pH). To limit spatial bias in the results, we randomly sampled observations from the country with the highest number of observations and limit the number of observations to that of the second most represented country.

We constructed a data set of possible drivers to train and predict the model (supplementary information table 1). Data come from 14 sources and include sanitation related variables, gross domestic product (GDP) per capita [23], population [24], urbanization rate, fertilizer use [25], croplands extent, livestock, precipitation and temperature [26], runoff [27], elevation, distance to shore, soil composition (soil pH and EC) [28], and river flows [29]. Squares, cubes, and interactions of variables were constructed to provide additional flexibility to the model. In total, our database contains 66 possible drivers (including squares, cubes and interactive terms).

Model estimation, fitting, and prediction were done with the ranger and caret libraries in R 4.1 [30–32]. Model training for each water quality indicator was done as follows. First, covariates with a near zero variance were excluded. Second, we randomly split the sample into ten folds and use cross-validation (CV) techniques—meaning that a given observation is used only to train or predict the model, but not for both. Third, we modeled water quality as a function of its drivers. We let the algorithm identify which variables to include for making accurate predictions. We estimate for each water quality indicator 1000 trees. Fourth, we explored which drivers of water quality were selected by the model with variable importance plots. We checked that these drivers are coherent with the literature using partial dependence plots [33]. Fifth, the final model is used to predict global values for all available grid cells (around 60 000), at a monthly time scale between 1992 and 2010. For each pollutant, we statistically analyzed trends in predicted pollution using modified Mann–Kendall trend test to account for auto-correlation [34]. Mann–Kendall tests were performed on a random sample of 3000 grid cells for each pollutant.

We chose RF for its general prediction performance compared to other regression techniques such as linear, partial least squares or support vector regressions. However, RF can present drawbacks, such as its sensitivity to time and/or spatial autocorrelations [35]. Such dependencies may lead to over-optimistic predictions. We test the accuracy of the predictions using station-blocs CV and wa ter basin leave one-out CV (LOOCV). In the station-bloc CV, instead of randomly allocating water quality observations to folds, we attributed monitoring stations to ten different folds and trained the model using CV. In the basin LOOCV all observations from a given basin were successively excluded from the training procedure and only used for testing. This was done to simulate the absence of a large geographical area. Finally, we conducted a temporal validation using an annual LOOCV approach (all observations of a given year are sequentially excluded from the training to be used for testing).

The validity and spatial transferability of RF predictions relies on the similarity that exists between the values of the predictors in the training and prediction samples. The spatial imbalances in our training dataset means that our model could not be able to predict trustworthy values for some parts of the world. This is for example the case for Sub-Saharan Africa for which we have extensive observations from South-Africa, and more limited observations from Ghana, Lesotho, Mali, Senegal, Sudan and Tanzania. Recent advances allow us to determine area of applicability (AOA) and dissimilarity index (DI) [36]. AOA is defined as the area, for which the CV error of the model applies. It is based on DI, a metric based on the minimum distance to the training data in the predictor space. We determine the AOA for each water quality indicator [37].

R2 for random splits, basin-block, station-block cross validation, and temporal splits are synthetized in table 1 and illustrated in supplementary information figures 2–5. A high correlation was found between observed and predicted water quality. With standard random splits validation techniques, the model explains 81% of the observed variability in the testing sample for pH, 70% for EC, 79% for DO, 71% for NOxN and 94% for Tw. This performance compares well to, or exceeds, the performance of other recently published process-based models. For example, RMSE of predictions for water temperature is half as large as reported for global process-based water temperature models [3–6]. A lower model performance was found for TP, where it predicts 37% of the observed variability in the testing sample. The prediction power of the models decreases, without collapsing, when using spatially structured CV based on basins or stations. R2 for Tw decreases only from 94% to 87%. For DO and pH, R2 are also preserved, but at lower levels. Higher decreases are observed for EC, NOxN, and TP, particularly for basin-block CV. However, further tests attest that this loss of predictive power is driven by a handful of basins (e.g. Schelbe basin for EC). The model preserves predictive power but cautions need to be taken to interpret local values. When station-block CV is used, more predictive power is maintained compared to basin-block CV. Finally, we follow previous assessments in process-based models by splitting water quality observations into three classes (good, medium or bad, thresholds displayed in supplementary information table 2). Our model accurately predicts the class of water quality and outperforms the process-based models in this task. As an illustration, for salinity, accuracy increases from 80% using a process-based model (appendix B of [19]) to 96% in the data-driven approach described in this paper. Our predictions also compares well to other recently developed machine learning models for nutrient pollutants [21]. The difference in accuracy between the different pollutants is likely to be driven by several factors, including the difference in quality of input data for both the pollutants and drivers, the type of drivers included in the model, and the higher inherent complexity to predict certain type of pollutants than others. This higher difficulty to predict some pollutants more than others, particularly with spatially-structured validation, translates into the number of drivers needed in the model to predict a pollutant accurately (see supplementary figure 7). For Tw, only two drivers have a variable importance larger than 50 (the square and cube of monthly air temperature, supplementary figure 7). On the other hand for nutrient pollutants (NOxN and TP), all 10 drivers included in the model have a variable importance larger than 50 (supplementary figure 7). This highlights the higher complexity to predict nutrient pollution than Tw. For DO, pH, and EC, 2, 5, and 6 drivers have a variable importance larger than 50.

Monthly time series data from 1992 to 2010 are generated for the six water quality indicators. Figure 2 shows the predicted average value of water quality between 2000 and 2010. Figure 3 displays the predicted change in annual average water quality between 1992 and 2010. Table 2 synthetizes the main results of the trend analyses. Water quality exhibited a significant change between 1992 and 2010 in 52% of the grid cells on average. For each pollutant, the percentage of grids where water quality deteriorated (30%) is greater than the percentage of grids where water quality improved (22%).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Supplementary information 7 highlights that a combination of hydro-climatic and socio-economic variables best predict all pollutants. Supplementary information 9 maps DI. The results indicate that for EC, Temp, TP and to a lesser extent pH, we can confidently extrapolate predicted water quality in continents like Sub-Saharan Africa despite the limited input water quality data. This is because the model can complement local African data by data from other continents, possibly at different times (e.g.: Latin America or South Asia in the 1990s shared important similarities with large parts of Sub-Saharan Africa in later decades). For DO and NOxN, the procedure indicates that uncertainties exist in some areas to predict water quality, including in Sub-Saharan Africa.

Unsafe levels of water quality are widely found in most parts of the world, driven by both climate and anthropic pressures. Low-, middle- and high-income countries all face unsafe levels but for different types of pollutants. Our model uncovers water quality hotspots in data scarce regions. Low levels of DO—a sign of unsafe water when levels are below 5–6.5 mg l⁻¹—are widely predicted in large parts of Sub-Saharan Africa, Latin America, and South and Southeast Asia (figure 2). Along with hydro-climatic variables, the lack of access to basic sanitation is a key covariate associated with low levels of DO (supplementary information 7). The infrastructure gap that prevails in most low- and middle- income countries explains these low values of DO. In places where the infrastructure gap has widened because of high population growth and low investment in sanitation, DO has decreased during the study period. This is, for example, the case for coastal parts of China, India, Nepal, or in the northeast region of Brazil.

The concentration of NOxN in water is the highest in densely populated areas with intensive economic activities. England, Belgium, Germany, and some parts of France are the predicted global hotspots of nitrate-nitrite, notably because of intensive animal farming (poultry and pig) and agricultural activities. The challenge of NOxN in most high-income countries has persisted during the period studied and has worsened in fast-growing economies, such as in South Asia, East Asia (e.g. eastern China) and parts of Mexico (figure 3). In these fast-growing areas, intensive animal farming, combined with high population density, excessive fertilizer use, and infrastructure gaps contribute to high nutrient pollution levels (supplementary information 7). A certain degree of caution should be taken when interpreting data from some parts of East Asia, because of the dissimilarities between the training and testing samples for NOxN.

High levels of salinity, as reflected by EC, are driven by geological conditions, drier climates, and the use of fertilizers, which is in correspondence with the overview salinity drivers identified in various river basins across the world [38]. Thus, Australia, Mexico, the Southern USA and Central Asia are salinity hotspots because of their drier climates (supplementary information 7). Over the study period 1992–2010, EC is predicted to have increased the most in India. Turning to Tw and pH, we find that soil composition and air temperature are, respectively, strong determinants of observed levels (supplementary information 6). Large parts of the world have experienced increases in water temperature greater than 1 °C in less than 20 years because of climate variations and change. Such increases in water temperature can have detrimental effects on aquatic life [39–41].