To address the drawbacks of traditional granular fertilizer detection methods—such as high cost, low efficiency, and heavy labor demand—this study presents an intelligent detection system based on deep learning. The system integrates two core functions: automatic analysis of fertilizer deposition patterns and real-time flow rate detection in centrifugal spreaders. Its goal is to improve fertilization accuracy, optimize the structure of spreading devices, and enable closed-loop control in variable-rate application. The hardware system includes a custom calibration mat, image acquisition unit, control terminal, and processing core for image capture, storage, and control. On the software side, two deep learning models are developed. The first is the F-TPP (Fertilizer-Swin Transformer PANet Partition) instance segmentation model, which combines Swin Transformer and PANet, and optimizes anchor scales to improve segmentation of small and densely distributed particles. The second is an improved lightweight YOLOv5s-seg model for real-time discharge monitoring. It uses HGNetV2 as the backbone, incorporates AKConv for size adaptability, and adds the iRMB attention mechanism to improve shape sensitivity. The detection head is also refined to reduce parameters and expand the receptive field. Experimental results show that the F-TPP model achieves a precision of 93.8%, recall of 96.3%, and F1 score of 0.95, outperforming models like Mask R-CNN, SOLO, and YOLOv5. The improved YOLOv5s-seg model achieves 45 FPS, 95.9% mAP, a model size of 6.66 MB, and a computational cost of 17.8 G. Under fertilizer outlet openings of 17–25 mm and a flow range of 15.13–54.83 g/s, the system reached a correlation coefficient of 0.9984 between measured and actual mass, with a maximum error of 5.07%.In conclusion, the proposed system provides a lightweight, accurate, and real-time solution for granular fertilizer monitoring, supporting intelligent variable-rate fertilization in precision agriculture.