Table 3.
Classification results on imbalance test set of N status by using different modeling methods
| Modeling methoda | Training datab | Classifiersc | Test set (μ = 6.76)Score (%) | ||
|---|---|---|---|---|---|
| WAPd | MAPd | WRd | |||
| Conventional | Labeled data | PLSDA | 50.3 | 46.7 | 55.1 |
| Labeled data | LDA | 51.8 | 50.1 | 55.2 | |
| Resampling + Labeled data | PLSDA | 49.2 | 52.0 | 53.3 | |
| Resampling + Labeled data | LDA | 60.2 | 54.5 | 60.9 | |
| Ensemble learning | Resampling + Labeled data | Adaboost | 58.8 | 52.0 | 57.2 |
| Resampling + Labeled data | XGBoost | 56.7 | 55.3 | 58.6 | |
| Resampling + Labeled data | RFC | 58.4 | 60.2 | 55.0 | |
| Self-training | Resampling + HSIs data | PLSDA | 54.6 | 51.8 | 63.1 |
| Resampling + HSIs data | LDA | 66.7 | 58.7 | 64.0 | |
| Resampling + HSIs data | AdaBoost | 62.6 | 54.7 | 61.0 | |
| Resampling + HSIs data | XGBoost | 61.7 | 56.4 | 61.8 | |
| Resampling + HSI data | RFC | 67.8 | 62.0 | 65.2 | |
a The conventional method represents supervised learning based method.
b The HSIs data means hyperspectral images data, including labeled mean spectrum and unlabeled pixels data.
c PLSDA, RFC, LDA, AdaBoost, and XGBoost are the abbreviations of partial least squares discriminant analysis, random forest classifier, linear discriminant analysis, adaptive boosting, and extreme gradient boosting methods respectively.
d The WAP, MAP, and WR are evaluation metrics of classification tasks.