Attribution of Eco-hydrological changes based on coupled SWAT-ML method

Vegetation is an important part of terrestrial ecosystem, and the vegetation growth condition is closely related to hydro-meteorological elements. Accurate simulation of ecohydrological elements is an important guarantee to maintain the security of ecosystem. Building physical models based on mechanistic processes such as the Soil and Water Assessment Tool (SWAT) is a solid way to understand the ecohydrological processes, but the simulation of vegetation growth is not accurate enough to interpret the entire complexity of ecohydrological processes. Data-driven machine learning models can efficiently and accurately identify the relationship between vegetation and hydrometeorological elements. Coupling distributed hydrological models and machine learning models is beneficial to improve the ecohydrological simulation accuracy, and to provide support for maintaining ecosystem security.A watershed ecohydrological simulation framework was constructed by coupling SWAT and six machine learning methods in the headwater basin of the Yangtze River, called Jinsha River basin, China. Firstly, we established a SWAT model to get the temporal and spatial patter of hydro-meteorological factors including soil moisture, runoff, evapotranspiration, temperature and precipitation in the watershed by using meteorological factors from the gauging stations. Then Pearson correlation coefficients was utilized to identify factors that are more relevant to vegetation growth based on the lagged response of vegetation changes to hydro-meteorological factors. We also applied machine learning models to construct the regression relationship between climatical factors and two indicators reflecting vegetation growth, which are normalized difference vegetation Index (NDVI) and solar-induced chlorophyll fluorescence (SIF), achieving the prediction of vegetation growth status. Based on this framework, the ecohydrological elements data series from 1965-2014 were completed in the monitoring data sparse area to conduct a long time series and sequential analysis. Finally, trend analysis and partial correlation analysis were used to explore the variation characteristics of ecohydrological elements and their relationships with climate factors.The results show that (1) the SWAT model can simulate the runoff process well of the whole Jinsha River basin (R2>0.84, NS >0.68), and the machine learning model can well estimate the SIF of the whole Jinsha River basin (NS>0.98, MSE<0.0003) and NDVI (NS=0.98, MSE=0.0005) in the upstream. (2) The vegetation type in the middle and downstream of the Jinsha River is mainly woodland, and the NDVI index has oversaturation phenomenon, so machine learnings can produce large biases, while the SIF data do not have such phenomenon, which is a better indicator to characterize the vegetation growth. (3) The trends and drivers of ecohydrological elements have obvious regional, and seasonal differences, and in general, temperature is the main driver of vegetation and precipitation is the main driver of runoff. This research built a new method to simulate ecohydrological processes in a spatio-temporal continuum, providing a strong support for ecohydrological evolution analysis.