Landscape patterns influence ecological processes. The purposes of this paper are to find out the nonlinear correlation mechanism between landscape pattern characteristics and total ecosystem service (TES) under the background of the coordination between ecosystem services, clarifying the dominant factors and optimal driving threshold, to ensure the effective implementation of ecological decision-making on the premise of complying with nature. The methods employed include the difference comparison method, self-organizing map (SOM), and random forest-partial dependence plots model, and compared the results of random forest (RF) model to that of the gradient boosting decision tree (GBDT) model and the ordinary least squares (OLS) model to verify the model’s reliability. The results show that: (i) the eco-efficient synergetic zones of Fujian rovince are mainly distributed in the high-altitude area, and the TES value within it shows the trend of west＞east. (ii) The influence degree of landscape characteristics of the cultivated, woodland, and construction land on TES shows a trend of area > shape > spatial layout. Among them, the influence of the landscape area of cultivated and woodland is higher than that of their patch area, and the influence of the patch area of grassland and construction land is higher than that of their landscape area. (iii) The influence of landscape pattern factors on TES has an optimal driving threshold interval and a sensitive regulation interval. The optimal driving threshold ranges of landscape indices of cultivated land are percentage of Landscape (PLAND) < 40%, largest patch index (LPI) < 30%, 1.923.5. The optimal driving threshold ranges of landscape indices of woodland are PLAND > 52%, LPI > 40%, LSI>2.5, PD>5. The optimal driving threshold range of grassland LPI is < 19%. The optimal driving threshold range of construction land PLAND is <9%, and when the construction land area accounts for more than 10%, its negative impact on TES no longer increases significantly with the expansion of construction land. It is concluded that the correlation between TES and landscape characteristics factors is nonlinear, and the coupling process has obvious stages, which can not be fitted by a single linear regression equation. In the comparison of two machine learning models and a linear regression model, the RF model has the best fitting effect, and the machine learning models are all superior to the linear model. The methodologies and insights from the study are broadly applicable and can inform further research in various regions and scales.