Evaluasi Pergeseran Domain Zero-Shot dan Latensi Klasifikasi Pipeline Hibrida Fitur Terseleksi Penyakit Daun Padi

Abstract

Most image-based rice leaf disease classification models are evaluated using datasets captured under controlled laboratory environments, leaving their operational behaviour on real-world agricultural field images largely unverified. This study reports an empirical zero-shot domain transfer evaluation and classification latency profiling of a hybrid feature-fusion pipeline (configuration E5). The architecture combines deep representations extracted from EfficientNet-B0 on segmented leaf images (1,280 dimensions) with 73 mask-based handcrafted features, standardized via z-score and filtered to the top 128 features using the ANOVA F-score before classification by a Radial Basis Function Support Vector Machine (SVM-RBF, C = 5.0). Trained once on 4,410 deduplicated public images spanning five classes, the frozen model was directly applied without retraining or fine-tuning to 80 locally collected field images covering two mapped disease classes (Bacterial Blight and Brown Spot, 40 samples each). Results reveal a severe performance degradation: accuracy fell from 99.64% ± 0.18% in cross-validated in-domain evaluation to 18.75% (15 of 80 images; 95% Wilson score confidence interval: 11.7–28.7%), while Macro-F1 dropped from 99.47% ± 0.28% to 13.64%. Perturbation experiments adding Gaussian noise across four intensity levels (σ = 0, 0.05, 0.10, and 0.20) shifted the out-of-distribution accuracy marginally from about 18.75% to about 17.5%; because the baseline accuracy was already near the five-class random-guess floor, this flat trajectory reflects a floor effect rather than model robustness. Isolated profiling of the classification step alone—a single SVM prediction on a precomputed 128-feature vector on CPU—took 0.78 ± 0.54 ms (P95: 1.69 ms; 1,278.5 FPS) over 100 repetitions, and 0.70 ± 0.21 ms (P95: 1.24 ms; 1,421.0 FPS) in a repeat verification. Because this figure excludes leaf and lesion segmentation, handcrafted feature extraction, and the EfficientNet-B0 forward pass, it does not represent end-to-end latency. Underlying causes of the domain transfer failure are systematically examined as research hypotheses