Impact of Preprocessing and Segmentation Techniques on Machine Learning and Deep Learning Models for Rice Crop Disease Classification: a Systematic Evaluation with Statistical Validation
Аннотация
Rice crop diseases cause 20–40% annual yield losses worldwide, threatening food security. While Machine Learning (ML) and Deep Learning (DL) show promise for automated disease detection, the individual and combined impacts of preprocessing and segmentation on classification performance remain poorly quantified across different model architectures. This study presents a systematic evaluation framework analyzing how preprocessing and segmentation influence classification across six classifiers (Convolutional Neural Network (CNN), ResNet-18, MobileNetV2, Support Vector Machine (SVM), Random Forest (RF), CNN + SVM) using five benchmark rice disease datasets under four experimental scenarios: (i) baseline, (ii) preprocessing only, (iii) segmentation only, and (iv) combined preprocessing and segmentation. Four preprocessing techniques (Contrast Limited Adaptive Histogram Equalization (CLAHE), Gaussian blur, histogram equalization, gamma correction) and five segmentation methods (U-Net, threshold, watershed, GrabCut, edge detection) were evaluated using task-based metrics with rigorous statistical validation (paired t‑tests, Bonferroni correction α = 0.0083, Cohen’s d, power analysis). Results demonstrate model-dependent effects. Deep learning models showed accuracy improvements with preprocessing for ResNet-18 specifically (+32.24%, p < 0.001, d = 2.34, the only result surviving Bonferroni correction), while CNN (+6.64%, d = 0.82), and MobileNetV2 (+9.08%, d = 1.12) showed large effect sizes that did not reach the corrected significance threshold at n = 5 datasets and should be read as suggestive rather than confirmed. Conversely, traditional ML models degraded. SVM −9.66% (d = −0.76) and RF −14.66% (d = −1.03). U‑Net segmentation provided high expert-rated quality (4.3/5, Fleiss κ = 0.78) and an average classification improvement of +8.2%. Combining CLAHE preprocessing with U‑Net segmentation achieved its largest accuracy gain for ResNet-18 on the Mendeley dataset (up to 100%) at an average processing time cost of 192 ms versus a 45 ms classification only baseline. These findings provide evidence-based guidelines for developing robust agricultural classification systems, demonstrating that optimal preprocessing-segmentation pipelines are model-dependent with clear accuracy-efficiency tradeoffs.