Abstract: Sidewall expansion dominates ephemeral gully erosion (EGE) once the channel bed incises to a less- or non-erodible layer. Channel widening mechanisms and prediction remain inadequately addressed due to the process’s complexity and episodic nature. We conducted laboratory-simulated runoff scouring experiments with 18 treatments: three inflow rates (0.000 25,0.000 67,and 0.001 08m³·s^-1), three slope gradients (1%, 5%, and 11%), and two typical soils (Atwood and Forestdale). Based on photogrammetry technology and deep learning methods, an automatic recognition network model for eroded channel boundaries and an associated automatic channel width output module were established, with DeepLabV3+ as the underlying framework. This model enabled the acquisition of time-series segmented images of sidewall boundary recognition (spatial resolution: 1 mm) and channel width distributions along the slope. Empirical relationships linking channel width with inflow rate and slope gradient were subsequently developed for the two soils. The results show that channel width variation exhibits a “two-stage” characteristic: power-function rapid growth in the initial stage followed by gradual stabilization process. Sediment discharge and widening rate generally increase with increasing inflow rate and slope gradient but do not entirely increase consistently with channel length stream power. Sidewall expansion Atwood slopes responds more rapidly to increases in slope gradient, whereas inflow rate has a more pronounced effect on Forestdale slopes. The model achieves a boundary recognition accuracy exceeding 90%, with a spatial resolution of 0.001m and temporal resolution of 3s, overcoming the low efficiency bottleneck of conventional methods and enabling efficient automatic extraction of eroded channel morphological parameters from images.The channel width prediction model has a coefficient of determination (R²)>0.9 and Nash-Sutcliffe efficiency coefficient (E_NS)>0.6. The methods and results of this study provide a methodological and theoretical basis for the monitoring and prediction of sidewall expansion erosion.