Model Performance
Historical/manual baseline results
Selected Model
Gradient Boosting
Highest mean 3-fold CV F1
Mean CV F1
0.8889
Standard deviation: 0.1571
Final Validation F1
0.6667
8 untouched validation records
Data Split
22 / 8
Training / final validation
Model Comparison
3-fold evaluation
Metric details: Accuracy, precision, recall, and F1 results from cross-validation for each supported model.
Validation Results
Gradient Boosting
Metric details: Final performance values calculated from the untouched validation records.
Confusion Matrix
Model comparison
Logistic Regression
Mean F1: 0.7778
16
True Low
0
False High
2
False Low
4
True High
Random Forest
Mean F1: 0.5556
16
True Low
0
False High
3
False Low
3
True High
Gradient Boosting
Mean F1: 0.8889
16
True Low
0
False High
1
False Low
5
True High
Confusion Matrix
Gradient Boosting
6
True Low
0
False High
1
False Low
1
True High
Matrix details: Dark cells are correct classifications; error cells contain false-high and false-low classifications.
Tuned Model Parameters
Training-only GridSearchCV settings
Learning Rate0.03
Max Depth1
N Estimators50
Best mean CV F10.8889
View final validation classification report
| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| Low/Moderate Risk | 0.8571 | 1.0000 | 0.9231 | 6 |
| High Risk | 1.0000 | 0.5000 | 0.6667 | 2 |
| Macro average | 0.9286 | 0.7500 | 0.7949 | 8 |
| Weighted average | 0.8929 | 0.8750 | 0.8590 | 8 |
Disaster Model Usage Notes
Task, selected model, and validation scope
Correct task:
This is a Low/Moderate versus High disaster-risk classification model, not a peso-damage regression model.
This is a Low/Moderate versus High disaster-risk classification model, not a peso-damage regression model.
Selected model:
Gradient Boosting obtained the highest mean cross-validation F1 score among the three candidates.
Gradient Boosting obtained the highest mean cross-validation F1 score among the three candidates.
Validation:
The final score came from 8 untouched records. Results support scenario analysis and are not guaranteed future forecasts.
The final score came from 8 untouched records. Results support scenario analysis and are not guaranteed future forecasts.
Best Model
Gradient Boosting
Tuned final classifier
Validation Accuracy
68.18%
F1 Score: 0.5333
Validation Recall
0.5714
High-risk detection rate
Train / Test Records
173 / 44
Final validation split
Model Comparison
Evaluation results for the three flood-risk classifiers
Metric details: Accuracy, precision, recall, and F1 results for each supported flood classifier.
Validation Results
Gradient Boosting
Metric details: Final scores calculated from hold-out records that were not used to fit the model.
Feature Importance
Top Gradient Boosting flood classification features
Barangay name length
0.2781
Municipality avg. landslide score
0.1920
Landslide score
0.1593
Municipality: Samal
0.0669
Municipality: City Of Balanga
0.0627
Municipality landslide-prone count
0.0501
Municipality: Abucay
0.0399
Municipality total barangays
0.0395
Validation Results
Gradient Boosting
| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| Low Risk | 0.7857 | 0.7333 | 0.7586 | 30 |
| High Risk | 0.5000 | 0.5714 | 0.5333 | 14 |
| accuracy | - | - | 0.6818 | 44 |
| macro avg | 0.6429 | 0.6524 | 0.6460 | 44 |
| weighted avg | 0.6948 | 0.6818 | 0.6869 | 44 |
Confusion Matrix
Model comparison
Logistic Regression
Accuracy: 70.05% | F1: 0.3299
136
True Low
14
False High
51
False Low
16
True High
Random Forest
Accuracy: 65.44% | F1: 0.4000
117
True Low
33
False High
42
False Low
25
True High
Gradient Boosting
Accuracy: 65.90% | F1: 0.4032
118
True Low
32
False High
42
False Low
25
True High
Confusion Matrix
Gradient Boosting
22
True Low
8
False High
6
False Low
8
True High
Matrix details: Dark cells are correct classifications; error cells contain false-high and false-low classifications.
Final Train/Test Distribution
Development and untouched validation records
Flood Model Usage Notes
Screening purpose, validation errors, and probability output
Use for screening:
The model can prioritize barangays for awareness and monitoring, but final decisions should still be validated by technical staff.
The model can prioritize barangays for awareness and monitoring, but final decisions should still be validated by technical staff.
Watch false negatives:
6 high-risk validation records were predicted as low risk, so low-risk output should not remove a barangay from monitoring.
6 high-risk validation records were predicted as low risk, so low-risk output should not remove a barangay from monitoring.
Focus on high probability:
Barangays with higher predicted probability should be shown first in risk awareness and preparedness recommendations.
Barangays with higher predicted probability should be shown first in risk awareness and preparedness recommendations.