Table 9 Ablation study for overfitting mitigation techniques.
From: Advanced deep learning framework for soil texture classification
Configuration | Accuracy | F1-Score | Kappa | AUC | Std. Dev (accuracy) | Notes |
|---|---|---|---|---|---|---|
Baseline (No Augmentation, No Regularization) | 0.942 | 0.853 | 0.897 | 0.943 | ± 0.0064 | High variance, signs of overfitting |
+ Dropout Only | 0.956 | 0.867 | 0.912 | 0.954 | ± 0.0050 | Reduced overfitting, slightly smoother training |
+Dropout + BatchNorm | 0.965 | 0.879 | 0.921 | 0.961 | ± 0.0038 | Improved generalization, better convergence |
+ Early Stopping | 0.972 | 0.886 | 0.934 | 0.970 | ± 0.0031 | Prevented late-epoch drift, stabilized training |
+Data Augmentation | 0.978 | 0.892 | 0.944 | 0.977 | ± 0.0027 | Increased diversity, improved class recall |
+ All Techniques (Proposed ATFEM) | 0.981 | 0.896 | 0.948 | 0.981 | ± 0.0022 | Optimal setting, robust and generalizable |
