Rice leaf disease is a major factor in reducing agricultural productivity because the visual similarity of symptoms between diseases complicates accurate diagnosis at an early stage. Most previous studies used an end-to-end Convolutional Neural Network (CNN) approach with a Softmax layer, which has limitations in modeling complex and overlapping feature distributions. This study aims to develop and evaluate a Hybrid Two-Stage Classification framework for rice leaf disease classification by separating the feature extraction and classification processes. The first stage uses EfficientNetB3 optimized through a full fine-tuning strategy to generate feature representations that are adaptive to the visual characteristics of rice diseases. The second stage applies XGBoost as a gradient boosting-based classifier on the extracted feature vectors. Testing was carried out using the RiceLeafs dataset consisting of four classes: Brown Spot, Healthy, Hispa, and Leaf Blast. Experimental results show that a full fine-tuned EfficientNetB3 with a standard Softmax classifier achieves 75% accuracy. Furthermore, the proposed hybrid approach, which substitutes the Softmax head with XGBoost, successfully raises the overall accuracy to 83% with a macro-F1 of 0.81. These findings indicate that the integration of deep feature extraction and ensemble-based classification can improve class separation precision and classification stability in image-based plant disease diagnosis.