Comprehensive benefit evaluation of conservation tillage based on BP neural network in the Loess Plateau

•Compared impacts of 5 tillage systems on Calcic Cambisol erosion in Loess Plateau.•Measured soil physicochemistry, microbial diversity, and soybean yield over 9 y.•Ascomycota and Basidiomycota abundances correlated with soybean yield.•Ascomyta and Gemmatimonadetes correlated with available K and N, respectively.•BP neural network showed that ridge-furrow plastic mulching optimal for this region. Soil erosion has caused substantial nutrient losses in the hilly-gully area of the Loess Plateau in China. Consequently, traditional tillage provides little ecological benefit there, and conservation tillage is required instead. To identify a suitable conservation tillage mode for this region, we set up no-tillage (NT), straw mulching (SM), plastic mulching (PM), and ridge-furrow plastic mulching (RPM) systems, used conventional tillage (CT) as a control, and monitored all systems for nine years. We measured and compared relative soil physicochemical properties, fungal and bacterial community diversity and composition, and microbial responses to the various environmental factors. Based on the measured soil quality indicators, we used a back propagation (BP) artificial neural network to perform a comprehensive evaluation of the various treatments to identify the conservation tillage mode that provided the greatest ecological benefits. Long-term RPM was the most highly efficacious at preventing water erosion. Relative to CT, conservation tillage methods significantly increased soil available N and available K levels and, by extension, the soybean grain yields. Neither fungal nor bacterial alpha diversity indices were significantly different among the tillage treatments, but the relative abundances of the dominant bacterial and fungal phyla significantly differed among treatments, and these relative abundances were significantly correlated with both yield, soil available K, and available N contents. Based on the results of our other analyses, bulk density, water and soil conservation capacity, available N, available K, and yield were selected as the appropriate BP neural network model inputs. We then used the output, the comprehensive evaluation value of each cultivation mode, to generate comprehensive simulations. The regression coefficient of the expected and the actual output value reached 0.99. The model disclosed that all types of conservation tillage (except NT) were superior to traditional tillage and that RPM conferred the greatest comprehensive benefits. Bas