From our project, we had some major findings. It was found that the values of different parameters of air depend on the latest past records (few days to a month) and not on many previous months. While retrieving real-time values through API for different parameters, sometimes, null or zero values occur. This might be due to malfunctioning of the sensors or inappropriate weather conditions. Zero or very less values might also occur at night because of the fact that certain parameters like O3 mix with other chemical compounds to form other compounds and consequently their value reduces. No2 and SO2 are also sometimes interacting and hence their abrupt values. The raw data is much easier to understand through visualizations for a common man. Also, lockdown is expected to be the effective alternative measure to be implemented for controlling air pollution.
Figure 1.3 Screenshot of fetched data.
Table 1.2 Precision, recall, and F1-score.
| Classes | Precision | Recall | F1-Score |
| Moderate | 1.0 | 0.99 | 0.99 |
| Poor | 1.0 | 0.95 | 0.97 |
| Satisfactory | 0.98 | 1.0 | 0.99 |
| Severe | 1.0 | 1.0 | 1.0 |
| Very Poor | 1.0 | 1.0 | 1.0 |
| Avg/total | 0.99 | 0.99 | 0.99 |
| Final Accuracy: 0.9893 | |||
Table 1.3 MAE and RMSE scores for different epochs.
| Test MAE for 1 | 8.864 |
| Test RMSE for 1 | 12.122 |
| Test MAE for 2 | 17.996 |
| Test RMSE for 2 | 35.390 |
| Test MAE for 3 | 23.820 |
| Test RMSE for 3 | 35.938 |
| Test MAE for 4 | 6.021 |
| Test RMSE for 4 | 9.269 |
Figure 1.4 Predicted values in Bengaluru in December, 2017.
Figure 1.5 Predicted values in Bengaluru in June, 2020.
Figure 1.6 Predicted values in New Delhi in December, 2017.
Figure 1.7 Predicted values in New Delhi in June, 2020.
Table 1.4 MAE scores for LSTM hyper parameters.
| Batch size | Epochs | NO2 | O3 | PM10 | PM2.5 | SO2 |
| 10 | 10 | 22 | 52 | 142 | 64 | 14 |
| 24 | 100 | 17 | 22 | 142 | 52 | 13 |
| 15 | 100 | 13 | 19 | 139 | 51 | 13 |
| 8 | 10 | 16.8 | 25.4 | 124 | 44.8 | 13 |
| 6 | 10 | 13 | 25 | 119.7 | 44 | 13 |