https://github.com/in-c0/drought-predictor-neural-network
Classification and Regression model trained on climate features extracted from the ERA5 global dataset
https://github.com/in-c0/drought-predictor-neural-network
Last synced: 2 months ago
JSON representation
Classification and Regression model trained on climate features extracted from the ERA5 global dataset
- Host: GitHub
- URL: https://github.com/in-c0/drought-predictor-neural-network
- Owner: in-c0
- Created: 2024-10-21T12:05:38.000Z (7 months ago)
- Default Branch: main
- Last Pushed: 2024-10-21T12:45:44.000Z (7 months ago)
- Last Synced: 2025-01-26T09:29:20.869Z (4 months ago)
- Size: 61.5 KB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# drought-predictor-neural-network
A drought predictor (Classification & Regression model) trained on climate features selectively extracted from the ERA5 global dataset over the period of 41 years from 1979 to 2020. Code is excluded for copyright reasons.
### Input Variables
| Variable | Description |
|----------|--------------------------------------------------|
| mn2t | Minimum temperature at 2 meters (°K) |
| msl | Mean sea level pressure (Pa) |
| mx2t | Maximum temperature at 2 meters (°K) |
| q | Specific humidity (kg kg⁻¹) |
| t | Average temperature at Pressure level 850 hPa (°K) |
| t2 | Average temperature at 2 meters (°K) |
| tcc | Total cloud cover (0-1) |
| u | U wind component at pressure level 850 hPa (m s⁻¹) |
| u10 | U wind component at 10 meters (m s⁻¹) |
| v | V wind component at pressure level 850 hPa (m s⁻¹) |
| v10 | V wind component at 10 meters (m s⁻¹) |
| z | Geopotential (m² s⁻²) |
| month | 1 to 12 |
| year | 1979 to 2020 |
| grid ID | The ID of the grid cell |
| SPI | Standardised Precipitation Index (unitless) |
### Target Variable
The drought index SPI (Standardised Precipitation Index) is used here as a proxy to characterise the intensity of drought caused by precipitation deficiency. Very low SPI values suggest intense drought, while very high values indicate very wet conditions.
SPI is our target variable in the regression task, and we will predict it based on the climate variables.
In the classification task, we will calculate a binary target variable ’Drought’ from SPI.
The ’Drought’ variable will be 1 to indicate the occurrence of drought and 0 to indicate no drought. We will apply a threshold of -1 to SPI, where values below or equal to this threshold indicate drought conditions (i.e. Drought = 1); otherwise, there is no drought (i.e. Drought = 0).
time series plot of SPI for a single grid cell is shown in Fig. 3. Periods where SPI falls below or is equal to -1 indicate periods of drought.
### Preprocessing
```
Loading data
======= Climate_SPI.csv =======
Row, Col: (15120, 16)
year month u10 v10 mx2t mn2t tcc t2 msl t q u v z SPI grid_ID
0 1979 1 -2.958586 -0.634959 309.023720 297.283986 0.291988 302.992399 100796.767708 299.970613 0.008493 -4.568461 -2.839551 459.991748 -0.715037 303
1 1979 10 -0.049360 -0.113635 302.851483 289.459347 0.192154 296.133314 101346.796371 293.335081 0.004813 -2.372457 -3.721755 904.832294 -0.246504 303
2 1979 11 -0.404197 -0.341808 308.174168 294.707882 0.306703 301.354577 101135.556510 298.360025 0.005545 -2.794565 -3.715220 751.517519 -0.921090 303
3 1979 12 -1.089172 -0.600055 310.344489 296.948569 0.327175 303.544371 100999.285534 300.545961 0.006324 -3.429895 -3.687178 622.386996 -0.917537 303
4 1979 2 -2.160727 -0.715109 303.806692 294.093896 0.502479 298.759323 101125.883929 297.114917 0.009923 -3.892438 -2.943332 778.958111 0.631926 303
===============================
Excluding year and grid_ID from data
======= Climate_SPI.csv =======
Row, Col: (15120, 14)
month u10 v10 mx2t mn2t tcc t2 msl t q u v z SPI
0 1 -2.958586 -0.634959 309.023720 297.283986 0.291988 302.992399 100796.767708 299.970613 0.008493 -4.568461 -2.839551 459.991748 -0.715037
1 10 -0.049360 -0.113635 302.851483 289.459347 0.192154 296.133314 101346.796371 293.335081 0.004813 -2.372457 -3.721755 904.832294 -0.246504
2 11 -0.404197 -0.341808 308.174168 294.707882 0.306703 301.354577 101135.556510 298.360025 0.005545 -2.794565 -3.715220 751.517519 -0.921090
3 12 -1.089172 -0.600055 310.344489 296.948569 0.327175 303.544371 100999.285534 300.545961 0.006324 -3.429895 -3.687178 622.386996 -0.917537
4 2 -2.160727 -0.715109 303.806692 294.093896 0.502479 298.759323 101125.883929 297.114917 0.009923 -3.892438 -2.943332 778.958111 0.631926
===============================
Adding 'Drought' column (0 or 1) based on SPI
======= Climate_SPI.csv =======
Row, Col: (2240, 15)
month u10 v10 mx2t mn2t tcc t2 msl t q u v z SPI Drought
9 7 -1.078143 0.902460 292.153382 278.805876 0.079997 285.057087 102283.450857 284.523974 0.003543 -2.473988 -1.114550 1695.127406 -1.099545 1
14 11 -0.753911 -0.576591 307.029682 293.860886 0.184344 300.264339 101172.854167 297.258291 0.005441 -3.071376 -3.726942 793.013513 -1.262581 1
20 6 -0.819211 0.867296 291.798410 280.497951 0.213060 285.672475 102054.188281 285.072712 0.005043 -2.170747 -0.765514 1537.134654 -1.606841 1
23 9 -0.170930 0.643260 301.150361 285.577702 0.054355 293.081444 101787.791667 291.516460 0.003437 -2.213144 -2.092395 1312.511619 -1.105385 1
37 10 0.034087 0.872630 302.460314 288.187859 0.148251 295.186908 101527.905746 292.523171 0.003713 -2.111258 -2.117658 1073.673213 -1.362772 1
===============================
Roughly 14.81% of the data has been labeled as 'Drought'.
The samples are equally distributed across 12 months over the entire dataset.
Month distribution:
month
1 1260
2 1260
3 1260
4 1260
5 1260
6 1260
7 1260
8 1260
9 1260
10 1260
11 1260
12 1260
Training set size: 10584 samples (70.0%)
Validation set size: 2268 samples (15.0%)
Test set size: 2268 samples (15.0%)
Total: 15120 samples
Replacing 'month' with 'cos_month' and 'sin_month'...
Training set:
======= Climate_SPI.csv =======
Row, Col: (10584, 16)
u10 v10 mx2t mn2t tcc t2 msl t q u v z Drought month_normalised cos(month_normalised) sin(month_normalised)
6552 -1.479215 -0.349176 308.946537 296.729571 0.200873 302.857344 101032.544010 298.684343 0.007421 -3.278758 -2.823981 694.846833 0 0.000000 1.000000e+00 0.000000
10290 1.215222 1.341728 292.337389 282.794098 0.507675 287.565900 101984.340104 284.021999 0.005509 -0.961927 -0.816961 1463.148627 0 1.570796 6.123234e-17 1.000000
1098 -1.049909 0.459411 301.584464 289.574852 0.202775 295.267790 101702.975000 292.449862 0.006014 -2.335011 -1.833616 1250.147030 0 1.570796 6.123234e-17 1.000000
4525 -0.587676 -0.438277 301.028648 288.192378 0.286398 294.396633 101463.419607 290.270405 0.005417 -2.975901 -3.688355 1019.604987 0 4.712389 -1.836970e-16 -1.000000
4313 -1.335883 0.822051 304.262825 291.893402 0.207135 298.119280 101405.180696 293.967737 0.006583 -3.529105 -1.983720 983.308748 0 1.047198 5.000000e-01 0.866025
===============================
Validation set:
======= Climate_SPI.csv =======
Row, Col: (2268, 16)
u10 v10 mx2t mn2t tcc t2 msl t q u v z Drought month_normalised cos(month_normalised) sin(month_normalised)
6706 1.839478 -0.081112 289.473332 278.865858 0.314341 283.883452 101895.747732 282.425177 0.003976 -0.153143 -1.877165 1323.057953 0 3.665191 -8.660254e-01 -0.500000
10093 0.897742 0.280632 295.250206 284.041292 0.427203 289.746081 101453.679940 284.966471 0.004620 -1.996432 -2.744730 945.190482 0 4.712389 -1.836970e-16 -1.000000
4169 -2.536848 0.726182 303.497044 293.499535 0.409161 298.526789 101526.751512 294.385840 0.007723 -4.536848 -2.243372 1103.272760 0 1.047198 5.000000e-01 0.866025
4660 -1.417957 0.659064 302.306398 289.817686 0.286100 296.016343 101274.662388 293.764555 0.007624 -2.855455 -1.091930 848.800585 0 0.523599 8.660254e-01 0.500000
5675 1.310550 0.547354 294.728937 282.397130 0.147330 288.758396 101700.329687 284.766466 0.004109 -1.696079 -2.510932 1189.587807 1 4.188790 -5.000000e-01 -0.866025
===============================
```
### Transformations
List of transformations applied:
- Removal of columns `year` and `grid_ID`
- Cyclic encoding of `month`, replacing it with `cos_month` and `sin_month`
- Standardisation of `u10`, `v10`, `mx2t`, `mn2t`, `tcc`, `t2`, `msl`, `t`, `q`, `u`, `v`, `z` using `StandardScaler`
- Skewness shifting using `PowerTransformer` (yeo-johanson) transformation on `q`, `tcc`, `v10`, `u`
### Building the model
- A sequential model
- A single dense hidden layer (ReLU)
- Dropout = 0.3
- 100 epochs
- A single output layer (sigmoid)
- Loss function = Binary cross-entropy
- Batch sizes = [64, 128, 256]
- Optimisers = ['SGD', 'Adam']
- sgd_learning_rate = 0.01
- adam_learning_rate = 0.001
- With SGD, the model with batch size 128 seems to be performing the best with lowest loss.
- With Adam, the model with batch size 64 seems to be performing the best, and it achieves the lowest validation loss (~0.309) within 100 epochs. For both optimisers, the models with batch size 256 is converging more slowly and seems to require more than 100 epochs for it to yield results matching the other batch sizes.
The best model is with Adam with batch size 64, with the minimal validation loss 0.3048010468482971
The accuracy stabilizes around 86% for both training and validation sets. The validation accuracy initially appears slightly higher than training accuracy in the first 10-20 epochs, but falls afterwards to values slightly lower than training accuracy. The accuracy seems to stabilise around that point, though there is some fluctuation. I expect that there will be no significant increase in accuracy beyond this, indicating that the model has mostly converged at around 86%.
### Optimisation
The 14 input features we have can be grouped by relevance. My hypothesis was that:
- `q` (humidity) and `tcc`(total cloud cover) are likely the most relevant factors to precipitation/droughtness,
- `cos(month_normalised)` and `sin(month_normalised)` might have a strong correlation due to seasonal nature of climate data,
- `t`, `t2`, `mx2t`, `mn2t` are all related to temperature, and using these all may be redundant,
- `msl` (mean sea level), `u` (wind), `v` (wind), `u10` (wind), `v10` (wind) will have to be strongly correlated to each other, including temperature, in order to be an indicator of drought (i.e. for sea water to evaporate and form a rain cloud that can move quickly enough to reach the Murray-Darling basin area),
In addition to this grouping, we will train the model with these additional subsets:
- humidity + total cloud cover (`q`, `tcc`)
- humidity + seasonality (`q`, `cos(month_normalised)`, `sin(month_normalised)`)
- humidity + total cloud cover + seasonality (`q`, `tcc`, `cos(month_normalised)`, `sin(month_normalised)`)
- humidity + total cloud cover + seasonality + geopotential (`q`, `tcc`, `cos(month_normalised)`, `sin(month_normalised)`, `z`)
- temperature + msl + wind (`t`, `t2`, `mx2t`, `mn2t`, `msl`, `u`, `v`, `u10`, `v10`)
- temperature (average only) + msl + wind (`t`, `msl`, `u`, `v`, `u10`, `v10`)
- temperature + msl + wind + total cloud cover (`t`, `t2`, `mx2t`, `mn2t`, `msl`, `u`, `v`, `tcc`)
- temperature + msl + wind + total cloud cover + seasonality (`t`, `t2`, `mx2t`, `mn2t`, `msl`, `u`, `v`, `tcc`, `cos(month_normalised)`, `sin(month_normalised)`)
- temperature (average only) + msl + wind + humidity + total cloud cover + seasonality (`t`, `msl`, `u`, `v`, `q`, `tcc`, `cos(month_normalised)`, `sin(month_normalised)`)
- all 14 features
The best subset is: temp_avg_msl_wind_humidity_cloud_seasonality with validation accuracy: 0.86816579
### Evaluation
- Accuracy: 0.81569665 (Balanced: 0.74676501)
- Precision: 0.42084942
- Recall: 0.64880952
- F1-Score: 0.51053864
## Regression model
The goal is to predict the intensity of the drought, represented by ‘SPI’. We applied the same transformation and building configuation, except that the output layer uses linear activation and MSE for the loss function. (Metrics MAE)
- Balanced Accuracy on New Data: 0.74614713
- Precision on New Data: 0.44197685
- Mean Absolute Error (MAE): 0.49075953
- Pearson Correlation Coefficient: 0.78394778
## Rooms for Improvement
- The model converges too early; could use deeper or more complex networks
- Using Recurrent layer instead of Dense
- Experimenting with additional layer(s),
- Experimenting with different threshold value,
- F1-score or AUC-PR