<![CDATA[Data-centric approaches have gained attention in plant disease classification; however, a systematic evaluation of underperforming ‘hard classes’ remains limited. This study proposes a four phase pipeline comprising granular error diagnosis, no-reference image quality assessment, class-targeted augmentation, and an ablation study to improve hard-class robustness without increasing model complexity. Using the New Plant Disease dataset and a lightweight MobileNetV3-Small backbone, we first established a baseline. Based on this baseline performance, we identified 12 hard classes (defined as those with either F1 < 0.96 or recall < 0.96 on the baseline model) with a mean F1 of 0.9260. The optimal configuration (categorical cross-entropy, no class weighting, baseline head) raised the mean hard-class F1 to 0.9668, corresponding to an absolute improvement of +4.08 percentage points, while maintaining global test accuracy at 98.08%. A two-tier design with three random seeds confirmed robustness (mean hard-class F1 = 0.9652 ± 0.0020). Unexpectedly, after data enrichment, inverse frequency class weighting degraded hard-class F1 (to 0.9621), and the default focal loss parameters offered no additional benefit over plain cross-entropy (0.9647). The MobileNetV3-Small showed comparable performance to the heavier EfficientNetB0 under our evaluation protocol, with no statistically significant difference detected (p = 0.089; limited power due to n = 3). Tomato Target Spot (C35) remained the most persistent bottleneck (F1 ≈ 0.90). Future work includes explainable AI for error analysis and real‑field validation.]]>
Copyrights © 2026
