Machine learning reveals divergent eco-efficiency paths in China’s two great river basins

China’s Yellow River Basin and Yangtze River Economic Belt together account for roughly seventy percent of the country’s population and economic output, yet a new study shows that these two watersheds have traveled strikingly different roads in their pursuit of green growth. Researchers at Henan University of Science and Technology measured eco-efficiency, an indicator that balances resource inputs against environmental costs, across nineteen provincial-level units from 2010 to 2022. Their central finding is a paradox of adjustment: once favorable external conditions were statistically stripped away, the Yellow River Basin’s true eco-efficiency fell significantly while the Yangtze Economic Belt’s rose slightly, even though the northern basin still held the overall advantage. The work, published in Frontiers in Environmental Science, is the first to bring a Geographically Weighted Random Forest model into eco-efficiency research, combining spatial econometrics with machine learning to expose nonlinear thresholds and spatially shifting drivers that traditional linear models cannot see.

The team built a unified cross-basin framework around a three-stage super-efficiency Slack-Based Measure model. In the first stage, an input-oriented SE-SBM model incorporating undesirable outputs such as industrial wastewater, sulfur dioxide, and smoke and dust emissions estimated each province’s efficiency. Unlike radial DEA approaches, the slack-based formulation accounts for input redundancy and pollution as an inevitable cost of economic activity, with the three pollutants weighted objectively through the entropy weight method into a comprehensive pollution index. Inputs spanned labor, capital, urban built-up land, energy consumption, and, innovatively, general public budget expenditures on agriculture, forestry, and water conservancy as a policy input reflecting basin governance. Desirable outputs included real GDP and local fiscal revenue, capturing both economic scale and development quality, with all monetary indicators deflated to 2010 prices.

The second stage is where the measurement gains its rigor. Input slack variables from the first stage were regressed through a Stochastic Frontier Approach against a Technology-Economy-Institution framework of environmental controls: R&D investment intensity, per capita GDP, and the share of government expenditure in regional GDP. The SFA decomposes each slack into managerial inefficiency, random disturbance, and external environmental factors. The likelihood ratio tests for all five input slack variables passed at the one percent significance level, and gamma values close to one confirmed that managerial inefficiency overwhelmingly dominated the mixed error term. This meant the raw first-stage scores were distorted: the Yellow River Basin’s initial average of 1.225 was partly an artifact of favorable conditions, while the Yangtze belt’s 1.025 was suppressed by unfavorable ones.

After adjusting inputs to a common external environment and eliminating random disturbances, the third stage re-ran the SE-SBM model. The adjusted averages settled at 1.153 for the Yellow River Basin and 1.029 for the Yangtze Economic Belt. The spatial fingerprints diverged sharply. The Yellow River Basin followed a stepwise gradient of downstream greater than midstream greater than upstream, with Shandong retaining the top position at 1.704 even after adjustment and Gansu at the bottom near 0.895. The Yangtze belt displayed a concave pattern with higher levels at both ends and a pronounced collapse in the middle: Shanghai led at 2.063 while Hubei sank to 0.565, and nearly half of its provinces never reached the efficiency frontier. Shanghai and Inner Mongolia saw their scores decline after adjustment, revealing efficiency supported by economic foundation or resource endowment, whereas Gansu and Sichuan improved, suggesting external constraints had previously masked their real management performance.

Because static scores cannot capture momentum, the researchers applied the Global Malmquist-Luenberger index, decomposing it into technical efficiency change and technical progress. The Yellow River Basin averaged 1.027, an annual improvement of 2.7 percent, against 1.017, or 1.7 percent, for the Yangtze belt, and its curve was steadier. Technical progress was the primary contributor in both basins, but the decomposition exposed a fragile, single-engine growth pattern. In several provinces, including Inner Mongolia in the north and Zhejiang, Hubei, Sichuan, and Yunnan in the Yangtze belt, the efficiency change component fell below one, meaning technology introduction masked declining management efficiency. Midstream regions of both basins posted the highest GML values, a catch-up effect driven mainly by green technology adoption rather than improved resource allocation, with Gansu, Shaanxi, Shanxi, and Yunnan achieving average annual growth of 6.8, 4.5, 4.2, and 3.4 percent respectively.

Kernel density estimation and the Dagum Gini coefficient then traced how the distributions evolved. The Yellow River Basin’s curves showed rising peaks and narrowing width, with its gravity center shifting from low-value to high-value concentration, a dual optimization of spatial balance and efficiency improvement moving toward convergence. The Yangtze belt instead settled into a prominent main peak with a long right tail, and long displayed double or even multi-peak structures in which high-efficiency cores coexisted with solidified low-efficiency sinks, a polarization the authors liken to a Matthew effect of the efficient growing more efficient. The Gini decomposition delivered the decisive structural verdict: inter-regional disparities contributed over seventy percent of the Yellow River Basin’s imbalance in every year except 2019, and between forty-two and sixty-one percent in the Yangtze belt, making gaps between upstream, midstream, and downstream regions the primary root of spatial inequality in both watersheds.

The methodological centerpiece is the Geographically Weighted Random Forest, which builds a local random forest at each provincial location using neighboring samples weighted by a Gaussian kernel of great-circle distances. Bandwidth selection employed leave-one-province-out spatial cross-validation across thirteen candidate bandwidths from 150 to 1,500 kilometers; the 200-kilometer fixed bandwidth achieved the lowest cross-validated RMSE of 0.3095 and MAE of 0.2133. Each of the nineteen local forests contained 500 regression trees trained on bootstrap samples, with two candidate variables per split and a minimum leaf size of five. Under a common benchmark, GWRF held a coefficient of determination at or above 0.94 and outperformed OLS, Spatial Regression Tree, Random Forest, and GWR models, confirming its comparative strength in capturing nonlinear, spatially heterogeneous relationships in high-dimensional panel data.

Partial dependence plots extracted from the model revealed thresholds that average-effect models would blur. Urbanization boosted eco-efficiency once the rate exceeded roughly 0.2 in the Yellow River Basin, but the Yangtze belt needed to surpass 0.4 before its urban infrastructure fully leveraged agglomeration advantages. Environmental regulation showed a turning point near 0.2 in both basins, with predicted efficiency rising rapidly with regulatory intensity before the gains became gradual. Industrial structure and financial development traced U-shaped trajectories, declining first as provinces absorbed the adjustment costs of eliminating backward capacity or expanding finance, then recovering beyond turning regions near 0.5 for industrial structure in the Yangtze belt and 0.2 for financial development in the Yellow River Basin. Openness carried divergent attributes: the Yangtze belt successfully imported advanced foreign environmental technologies during its highly open period, while the Yellow River Basin saw efficiency decline during a relatively higher openness phase, consistent with absorbing some high-pollution foreign enterprises.

Local feature-importance maps for 2010, 2014, 2018, and 2022 exposed two distinct spatial architectures. The Yellow River Basin showed a regulation-oriented constraint pattern: educational development held high local importance in upstream provinces such as Inner Mongolia, Gansu, and Qinghai, while environmental regulation dominated the midstream and downstream heavy-industry regions, where strict policy constraints are needed to break through carbon lock-in. The Yangtze Economic Belt presented a market-oriented radiation pattern, with openness and financial development most important in downstream coastal areas like Shanghai and Jiangsu, and urbanization most important in the midstream, where urban cluster agglomeration reduced pollution control costs. Notably, the high-importance areas of industrial structure did not stay fixed but rotated spatially among midstream provinces, indicating that the focus of China’s green industrial transformation is shifting from coastal pioneers toward inland hinterlands. The authors caution that these results are model-based associations rather than causal effects, given the observational panel design, and that prefecture-scale data and emerging factors such as digital transformation and carbon trading remain outside the current framework.

Subject of Research: Comparison and spatial heterogeneity of eco-efficiency in two major river basins of China: nonlinear mechanism analysis based on spatial machine learning

Article Title: Comparison and spatial heterogeneity of eco-efficiency in two major river basins of China: nonlinear mechanism analysis based on spatial machine learning

Article References: Liu, P., Dong, D., Zhao, F., Meng, J., & Deng, G. (2026). Comparison and spatial heterogeneity of eco-efficiency in two major river basins of China: nonlinear mechanism analysis based on spatial machine learning. Frontiers in Environmental Science, 14, Article 1937303. https://doi.org/10.3389/fenvs.2026.1937303

Image Credits: AI Generated

DOI: 10.3389/fenvs.2026.1937303

Keywords: Comparison, spatial, heterogeneity, eco-efficiency, major, river, basins, China, nonlinear, mechanism, analysis, based