The root system plays a crucial role in optimizing fertigation (fertilizer and water application) for improved crop productivity and enhanced nitrogen (N) and water use efficiency. However, optimization remains challenging due to the difficulty of accurately characterizing root dynamics, particularly in field conditions. Automated continuous root phenotyping can provide information to address the gap. This study investigated the response of bell pepper roots to N application and soil moisture-based irrigation, while testing a convolutional neural network (CNN)-based model for estimating total root length (TRL). The experiment was conducted in a greenhouse using 11 automated minirhizotron (MR) systems with two treatments: 100 ppm N (control) and 50 ppm N + triggered irrigation (low N). Irrigation was triggered at a VWC of 8.4%, at 20 cm depth. A total of 32,990 MR images were analyzed using the CNN-based model, alongside 1,087 manually annotated images using Rootfly for model improvement. Low N promoted higher TRL and greater root length density at 30 cm depth, and continuous TRL measurements indicated a greater increase over time in the low N treatment compared to the control. Control achieved ~31% higher yield, although root growth and NUE were enhanced under low N. The model improved TRL estimation accuracy from R² = 0.96 to 0.98 on images collected at 14–21 day intervals, and from R² = 0.89 to 0.97 for continuous TRL after model retraining. The findings show that automated root phenotyping can reliably capture root responses, supporting precision agriculture by providing real-time root growth information.
