Table 1 The performance of deep learning and standard machine learning models for predicting mosquito age classes from the same or alternate insectaries, with and without dimensionality reduction (DR) and transfer learning
Models | Dimensionality reduction (DR) technique | Training data sources | Transfer learning | Base Model runtime | Transfer learning runtime | Predictions for age of mosquitoes from same insectary (Ifakara) -Test accuracy (%) | Predictions for age of mosquitoes from alternate insectary (Glasgow)—unseen data accuracy (%) |
|---|---|---|---|---|---|---|---|
CNN-1 | No DR | Ifakara | No TL | 7.2 s/iteration | N/A | 99 | 46 |
CNN-2 | No DR | Ifakara | 2% (33 of 1635) | 7.2 s/iteration | 1 min | 99 | 100 |
CNN-3 | No DR | Ifakara | 5% (82 of 1635) | 7.8 s/iteration | 2 min | 99 | 96 |
MLP-1 | PCA | Ifakara | No TL | 6.5 s/iteration | N/A | 91 | 58 |
MLP-2 | t-SNE | Ifakara | No TL | 1 s/iteration | N/A | 84 | 58 |
MLP-3 | PCA | Ifakara | 2% (33 of 1635) | 0.8 s/iteration | 35 s | 91 | 97 |
MLP-4 | PCA | Ifakara | 5% (82 of 1635) | 0.7 s/iteration | 51 s | 91 | 96 |
MLP-5 | t-SNE | Ifakara | 2% (33 of 1635) | 0.7 s/iteration | 47 s | 83 | 50 |
MLP-6 | t-SNE | Ifakara | 5% (82 of 1635) | 0.7 s/iteration | 49 s | 83 | 55 |
XGB-1 | No DR | Ifakara | No TL | 645 s/iteration | N/A | 92 | 48 |
XGB-2 | No DR | Ifakara | 2% (33 of 1635) | 975 s/iteration | 1 s | 92 | 98 |
XGB-3 | No DR | Ifakara | 5% (82 of 1635) | 861 s/iteration | 1 s | 92 | 98 |
XGB-4 | PCA | Ifakara | No TL | 60 s/iteration | N/A | 90 | 48 |
XGB-5 | t-SNE | Ifakara | No TL | 66 s/iteration | N/A | 68 | 55 |
XGB-6 | PCA | Ifakara | 2% (33 of 1635) | 54 s/iteration | 1 s | 90 | 98 |
XGB-7 | PCA | Ifakara | 5% (82 of 1635) | 54 s/iteration | 2 s | 90 | 97 |
XGB-8 | t-SNE | Ifakara | 2% (33 of 1635) | 60 s/iteration | 1 s | 81 | 43 |
XGB-9 | t-SNE | Ifakara | 5% (33 of 1635) | 60 s/iteration | 1 s | 82 | 49 |