To date, relatively few studies have examined flood sediment processes on the Loess Plateau of China at a small watershed scale using long-term datasets with high temporal resolution. In addition, research on predictive sediment modelling has also been inadequate. This study investigates sediment sources and availability by analysing hourly-scale water-sediment data from long-term site-specific observations. It integrates hysteresis analysis of suspended sediment concentration and discharge (SSC-Q) with an improved differential hysteresis index (HIK). A revised sediment connectivity index (RIC) was developed by incorporating source-sink landscape indices and gully density facilitating the visualisation of sediment transport pathways and potential erosion risk zones. A set of specific sediment yield (SSY) modelling schemes suitable for the Loess Plateau was explored using feature-selection algorithms with machine learning. The findings revealed that complex hysteresis loops accounted for 64.67% of the total sediment transport, making them the dominant sediment transport mechanism in the region. Sediment movement between the sources and sinks of the watershed are highly interconnected, originating from both scouring of distant slopes and the erosion of proximal stream beds and agricultural fields. Deep Neural Networks (DNN) demonstrate a higher prediction accuracy for SSY, especially when using variables selected by the input Boruta Feature Selection Algorithm (Boruta) algorithm. The findings provide insights into understanding flood sediment processes in small watersheds on the Loess Plateau. This has also contributed to improved river flow management and erosion control in small watersheds of the Loess Plateau.