The decline in the productivity of clove (Syzygium aromaticum) plantations is often caused by delays in treating nutrient deficiencies due to subjective manual detection methods and the high visual similarity among symptoms. This study aims to develop a hybrid deep learning system that integrates MobileNetV2 and a Graph Convolutional Network (GCN) to automatically identify Nitrogen, Phosphorus, and Potassium deficiencies, as well as Healthy conditions. Input images are first preprocessed and extracted using a pre-trained MobileNetV2 into a 256-dimensional normalized dense embedding vector. The structural relationships and visual similarities between samples are then mapped into a neighborhood graph using the K-Nearest Neighbors algorithm with specific parameters to form a normalized adjacency matrix. The final classification is performed by a two-layer GCN network with residual skip connections. Evaluation of 92 test images demonstrates highly effective model performance, with uniform accuracy, precision, recall, and F1-score reaching 94.57%. Although the final accuracy of this MobileNetV2-GCN hybrid model is equivalent to that of a single MobileNetV2 architecture, the integration of GCN is substantively proven capable of modeling similarity relations (similarity graphs) between image samples while maintaining the stability of feature clustering in the latent space.