CNSW 1.0: Prefectural Reconstruction of China’s Surface Water Resources Using Machine Learning Methods

Validation of machine learning models

Among these models, the RF model demonstrates the lowest RMSE at 53.87 mm, outperforming the BR, GBM, and SVR models, which have RMSE values of 54.23 mm, 77.59 mm, and 91.88 mm, respectively (Table 2). RF also achieves the highest R² of 0.98, followed closely by BR, GBM, and SVR with R² values of 0.98, 0.97, and 0.95. BR shows the lowest PBIAS at 6.79%, indicating the best performance in terms of bias, while RF, SVR, and GBM follow with PBIAS values of 6.85%, 11.27%, and 11.45%, respectively. Conversely, the PR and BLR models exhibit poorer performance, with PBIAS values of 26.56% and 24.39%, respectively. The DTR model performs the worst, with an RMSE of 178.08 mm, R² of 0.82, and PBIAS of 28.6% (Supplementary Figure 3). Overall, RF and BR models demonstrate the best performance in the training dataset, with GBM and SVR also performing well, while PR, BLR, and DTR show inferior results.

In evaluating the performance of the test datasets, SVR exhibits the best results, with an RMSE of 93.07 mm, an R² of 0.95, and a PBIAS of 14.87%. The RF model also performs commendably, with an RMSE of 98.77 mm, an R² of 0.94, and a PBIAS of 15.13%. The BR and GBM models also demonstrate strong performance, with BR achieving an RMSE of 102.82 mm, an R² of 0.93, and a PBIAS of 15.72%, while GBM shows an RMSE of 99.27 mm, an R² of 0.94, and a PBIAS of 15.74%. Density scatter plots (Supplementary Figure 4) illustrate that the predictions of SVR, RF, GBM, and BR closely align with the y = x line, indicating superior performance. Conversely, the DTR model continues to underperform, with an RMSE of 167.50 mm, an R² of 0.82, and a PBIAS of 29.08%. PR and BLR also show suboptimal performance.

In the training dataset (Fig. 2), the RF and BR models significantly outperform the other models. In the test dataset, SVR, RF, BR, and GBM models all derive superior results compared to the other models, with SVR slightly surpassing RF, BR, and GBM. Overall, SVR achieves the best performance in the test dataset, while RF, BR, and GBM also perform well. A comprehensive analysis of both training and test datasets highlights that BR, GBM, RF, and SVR exhibit the strongest overall performance, whereas BLR, PR, and DTR show relatively poor performance.

Time series

CNSW 1.0 derived 16 datasets at the prefecture-level. In addition to the 14 datasets derived from various machine learning models, this study also calculates the average and median of these model predictions, resulting in two additional datasets of prefecture-level surface water resources for China. We analyzed the temporal trends in surface water runoff depth (Fig. 3a), the multi-year average spatial distribution (Fig. 3b), and the spatial evolution trends (Fig. 3c) of surface water resources in China from 2000 to 2020 across all 16 datasets.

Temporal analysis (Fig. 3a) reveals that, except for the dataset simulated by the MLP model, surface water resources in China exhibited an upward trend from 2000 to 2020, though this trend is not statistically significant. Notably, years such as 2002, 2006, 2010, 2012, 2015, 2016, 2019, and 2020 experienced relatively high surface water resources, while years like 2004, 2007, 2009, and 2011 were characterized by lower levels. According to the China Water resources Bulletin, the total surface water resources in China increased at an average annual rate of 1.77 mm from 2000 to 2020. Among the simulations, GAM, GPR, KNN, RF, and SVR closely matched the observed values, whereas the MLP model exhibited a trend contrary to the actual surface water resources trend, indicating suboptimal performance.

Spatial distribution

Figure 3b illustrates the multi-year average spatial distribution of surface water resources in China as simulated by each model. While all models consistently identify the abundant water regions in SE, their results diverge in Northwest, Northeast, and Southwest. The areas with the least surface water resources are found in NC and NW, particularly within endorheic basins and non-monsoon regions. In contrast, SE displays the highest surface water availability which were largely attributed to its elevated precipitation levels.

The analysis reveals that regions experiencing a decline in surface water resources are concentrated in four key areas: the NW inland river basins, NC and EC, the Southwestern river basins, and the Southeastern coastal region (Fig. 3c). Conversely, the most substantial increases in surface water resources are observed in the border areas of Jiangxi, Anhui, Zhejiang, and Jiangsu provinces. Additionally, Qinghai Province, Sichuan Province, Guizhou Province, and the three northeastern provinces exhibit notable increases in surface water resources. The Ili Kazakh Autonomous Prefecture in Xinjiang also shows significant growth in surface water availability.

Quality evaluation and comparison

This study assesses the deviation of each dataset in CNSW 1.0 from the total surface water resources reported by the China Water Resources Bulletin using the formula for annual totals. A PBIAS value closer to zero indicates a smaller discrepancy between the dataset and the Bulletin’s reported (Tables 3, 4). Among the datasets, simulations by BLR, BR, ENR, GLM, LR, and SVR closely matched the observed totals in the early years. However, in recent years, these simulations generally show lower values than those observed. Specifically, BLR’s simulated totals were accurate up to 2012 but fell below observed values post-2013. BR’s simulations were higher than observed totals before 2007 but were lower in most subsequent years, except for 2015. ENR showed similar discrepancies. GLM results were higher than observed values before 2006 (except 2003) and slightly lower thereafter (except 2008 and 2012). LR exhibited discrepancies comparable to GLM. SVR data were slightly higher than observed values until 2015 (except 2003) but fell below observed totals from 2016 onward.

Across all datasets, simulations of total surface water resources in China by DTR, PR, RF, GAM, GBM, GPR, KNN, Median, and Average generally yielded lower values than the observed totals. Specifically, DTR consistently underestimated the total surface water resources, with an average discrepancy of approximately 9%. Simulated totals from PR and RF were also slightly lower than observed values. GAM’s simulations were below observed totals in all years except 2000 and 2002. GBM simulations were lower in all years except 2000–2003. GPR’s simulations were lower than observed values in all years except 2000. The discrepancies for KNN and Median were similar to those of GPR. The Average dataset’s simulations closely aligned with observed values before 2006 but showed a slight decrease afterward. Notably, only the MLP consistently overestimated total surface water resources across almost all years, with a significant overestimation of 38% in 2000. Additionally, while observed surface water resources increased from 2014 to 2015, MLP simulations indicated a decrease, and in 2018, MLP showed an increasing trend contrary to the observed decline.

In addition to assessing discrepancies in simulated total surface water resources across China, we also evaluated the accuracy of CNSW 1.0 at the provincial and regional scales using metrics of R² and PBIAS (Fig. 4 and Supplementary Figures 5, 6). Nationally, eight datasets—GAM, GBM, GPR, KNN, PR, RF, and the average and median values—demonstrated exceptional simulation accuracy (R² > 0.95), with RF achieving the highest accuracy (R² = 0.98). However, PBIAS analysis revealed varying stability. BR showed the lowest absolute national bias (−1.65%), indicating excellent stability, whereas RF, despite high accuracy, exhibited a negative bias (−8.14%). In contrast, MLP had both poor accuracy (R² = 0.55) and severe overestimation (PBIAS = 11.55%), highlighting significant deficiencies. At the regional scale, substantial spatial variability was observed. CC showed near-perfect accuracy (R² = 0.99) and negligible bias (PBIAS = 0.09%) due to complete data coverage. SC also displayed high accuracy (R² > 0.95) and minimal bias (PBIAS < 0.55%). EC had generally excellent accuracy (R² > 0.95), with RF (R² > 0.95, PBIAS = 0.79%) and SVR (PBIAS = 0.21%) performing optimally. WS exhibited accuracy discrepancies, with GAM, PR, and RF showing good performance (R² = 0.90–0.95), yet significant underestimations occurred (RF: PBIAS = −22.07%). WN faced greater challenges due to sparse data; RF achieved highest accuracy (R² = 0.88) and relatively lower bias (PBIAS = 5.51%), whereas MLP and GAM showed extreme deviations (PBIAS: 64.56%, 50.79%). In NC, all models had moderate accuracy (R² < 0.90), with severe MLP overestimation (PBIAS = 206.42%) contrasting with BLR’s best regional accuracy (R² = 0.880). NC showed RF as the top-performing model (R² = 0.837, PBIAS = 3.73%), whereas ENR and MLP obviously underperformed (R² < 0.40). Several provinces, including Beijing, Chongqing, and Shanghai, showed perfect accuracy (R² = 1.00, PBIAS = 0%). RF consistently demonstrated robust accuracy and bias control across diverse provinces: Ningxia (R² = 0.97, PBIAS = −0.64%), Anhui (R² = 0.91, PBIAS = −0.43%), and Yunnan (R² = 0.95). Conversely, severe datasets deficiencies were evident in Heilongjiang and Shanxi provinces, with MLP showing obvious overestimations (Heilongjiang: PBIAS = 367.27%, Shanxi: PBIAS = 145.80%) and negligible accuracy (R² ≈ 0).

Considering both R² and PBIAS, the RF model emerged as the most optimal choice across national, regional, and provincial scales. While the BR model exhibited the lowest absolute bias at the national level, its performance declined significantly in regional applications, particularly in East and Southwest China. In contrast, RF effectively balanced high predictive accuracy with controlled bias, consistently maintaining R² values above 0.80 and achieving notably low biases even in regions with complex hydrological conditions. This comprehensive evaluation highlights RF’s robustness and adaptability, making it the preferred model for reliable surface water resource simulations across diverse hydrological contexts.

Intermodel comparison

This study compares the CNSW 1.0 dataset with four other datasets—CNRD v1.0, GRUN, ISIMIP2a, and ISIMIP3a—across four dimensions: simulation accuracy, total surface water resources discrepancies in China, prefecture-level discrepancies, and spatial distribution from 2000 to 2010. The end years for the datasets are 2018 for CNRD v1.0, 2014 for GRUN, 2020 for ISIMIP2a, and 2019 for ISIMIP3a. For consistency, spatial distribution analysis is restricted to the 2000–2010 period, while other comparisons utilize the full available time span. ISIMIP2a and ISIMIP3a averages are based on 18 and 14 qualifying models, respectively.

The comparative analysis of CNSW 1.0 datasets (RF, Average, Median) against alternative runoff datasets (CNRD v1.0, GRUN, ISIMIP2a, and ISIMIP3a) reveals significant differences in both R² and PBIAS at national, regional, and provincial scales. At the national scale (Fig. 5a,b), ISIMIP2a shows the highest simulation accuracy (R² = 0.98), equal to CNSW 1.0’s RF model (R² = 0.98). However, ISIMIP2a demonstrates substantial bias (PBIAS = −10%) compared to CNSW 1.0 datasets, which maintain biases within 10%, highlighting their stability. Conversely, GRUN significantly underestimates water resources nationally (PBIAS = −31.7%), while CNRD v1.0 and ISIMIP3a display substantial overestimations exceeding 10%. Regionally, CNSW 1.0’s RF consistently exhibits superior overall performance. In CC, RF achieves near-perfect accuracy (R² ≈ 1.0, PBIAS ≈ 0%), in stark contrast to the substantial biases observed in CNRD v1.0 (44.37%) and GRUN (−56.1%). Similarly, in EC, RF maintains high precision (R² > 0.95, PBIAS < 1%), whereas CNRD v1.0 and GRUN exhibit considerable deviations. In NC, RF outperforms alternative datasets with superior accuracy (R² = 0.84, PBIAS = 3.73%), significantly reducing the overestimation seen in CNRD v1.0, ISIMIP2a, and ISIMIP3a. For EN, RF achieves comparable accuracy (R² ~0.88, PBIAS = 0.66%) to ISIMIP datasets (PBIAS = −0.24% and 13.64% for ISIMIP2a and ISIMIP3a), while GRUN markedly underestimates regional water resources (−57.52%). In Northwest China, comparative datasets systematically overestimate surface water resources (CNRD v1.0: 59.23%, GRUN: 47.44%, ISIMIP3a: 31.08%), whereas ISIMIP2a underestimates (−18.58%). In contrast, RF maintains robust performance with controlled bias (R² = 0.88, PBIAS = 5.51%). In humid South China, RF demonstrates exceptional accuracy (R² > 0.95, PBIAS ≈ −0.51%), distinctly outperforming comparative datasets such as GRUN and CNRD v1.0. The WS region presents challenges for all models, with RF exhibiting a systematic underestimation, yet remaining comparable to GRUN and performing better than ISIMIP2a. These findings underscore RF’s overall superiority in maintaining high accuracy and minimizing bias across diverse hydrological conditions.

At the provincial scale (Fig. 6a,b), CNSW 1.0 consistently outperforms comparative datasets, with RF and median estimates maintaining biases within 1% in Guangdong, Guangxi, Hainan, Guizhou, and Jiangxi. In contrast, alternative datasets exhibit extreme deviations: CNRD v1.0 overestimates water resources in Hebei (276.13%) and Ningxia (365.32%), while GRUN underestimates in Heilongjiang (−51.77%) and Jilin (−62.36%). ISIMIP2a shows severe overestimations in North China (Hebei: 248.90%, Shanxi: 182.93%) and underestimations in the Southwest (Sichuan: −28.61%, Xinjiang: −59.46%). ISIMIP3a amplifies these biases, particularly in Hebei (422.99%) and Shanxi (346.62%), though with improved performance in Guangdong (−7.82%). Despite localized challenges, such as RF biases in Tibet (−51.97%), CNSW 1.0 exhibits superior stability across diverse hydrological conditions.

Overall, CNSW 1.0’s datasets effectively integrate observed data with interpolative methods, achieving robust accuracy (R² is generally above 0.90) and superior bias control at all scales. While comparative datasets occasionally match CNSW 1.0’s accuracy in specific regions, they often exhibit varied regional biases due to their reliance on purely predictive methods. This analysis underscores CNSW 1.0’s distinct advantage in accurately simulating China’s prefectural surface water resources across diverse hydrological conditions.

Figure 7 depicts the discrepancies in prefecture-level simulations between CNSW 1.0 and the four runoff datasets. CNSW 1.0 operates at a spatial resolution of 0.25°, whereas the other datasets use a coarser resolution of 0.5°. GRUN’s simulations consistently yield lower values than CNSW 1.0, exhibiting the lowest R² value. In contrast, CNRD v1.0 forecasts higher runoff values compared to CNSW 1.0. ISIMIP2a tends to overestimate runoff in prefectures with less than 400 mm of surface water resources and underestimate in those with more than 400 mm. ISIMIP3a shows a similar pattern but with a threshold of 800 mm, overestimating in prefectures with less than 800 mm and underestimating in those with more. The density scatter plots illustrating discrepancies for all models in ISIMIP2a and ISIMIP3a are provided in Supplementary Figures 9, 10.

Figure 8 illustrates the time series trends of measured versus simulated total surface water resources in China. CNRD v1.0 consistently overestimates values, with significant discrepancies in 2007 and 2013. GRUN tends to underestimate, with notable inconsistencies in 2001, 2007, and 2014. ISIMIP3a also shows overestimation, though less pronounced than CNRD v1.0, with discrepancies in 2007 and 2014. ISIMIP2a aligns most closely with the measured data, except for 2002, although its time series extends only to 2010, shorter than CNSW 1.0 and the other datasets. Overall, CNSW 1.0 demonstrates superior performance in terms of simulation accuracy, time series alignment, and total volume compared to the four other runoff datasets.

Among these, CNRD v1.0 most closely aligns with CNSW 1.0, particularly in North China (Fig. 9). GRUN displays notably different spatial distribution patterns, with generally lower surface water levels across all regions. ISIMIP2a accurately reflects the surface water distribution in southern China but shows significant deviations from CNSW 1.0 in northern regions. ISIMIP3a fails to capture the spatial variability of surface water resources in southern China and exhibits poor performance in the north. The spatial distribution characteristics for all years in ISIMIP2a and ISIMIP3a are further detailed in Supplementary Figures 11, 12.