A Distributed Machine Learning Model for Blue and Green Water Resources With Transferable Applications in Similar Climatic Zones

Abstract

Human activities profoundly impact the terrestrial water cycle and the spatiotemporal dynamics of blue and green water resources. Distributed hydrological models are essential for simulating the water resources within a basin. However, neither process-based nor data-driven hydrological models have fully captured the effects of human activities on the distribution of blue and green water resources in space and time. Here we construct a distributed machine learning model for monthly blue and green water resources, which is trained and calibrated for the Yellow River Basin (YRB) in China, and validated and tested for the transferability to similar climatic zones in the case for Colorado River Basin (CRB) in the United States. The modeling thoroughly accounts for the influence of human activities, incorporating 5 scales (grid, county, city, province, and cluster), 4 algorithms, and 2 model integration methods (Stacking and Bayesian). The R² values reached 0.84 and 0.97 for blue and green water models, respectively, during the test period in the YRB. The corresponding high modeling accuracy maintained with R² values of 0.72 and 0.97 when transferred to the CRB. The model performed better in regions with higher human activity intensity. Precipitation and spatial encoding are respectively the most sensitive feature variables for the green water and blue water models, while nighttime lights and population density are respectively the most significant human activity-related features. The study highlights the non-negligible impacts of socioeconomic factors on spatiotemporal dynamics of blue and green water resources, and the feasibility of machine learning modeling.