To evaluate and compare multiple machine learning strategies, we tested several model configurations, including gradient-boosted decision trees, random forests, and neural networks.
Among these, the XGBoost-based approach demonstrated the best overall performance and computational efficiency. For this reason the PoF system uses gradient-boosted decision trees that iteratively improve prediction accuracy by correcting errors made in previous iterations. We found that this method provides an efficient optimization framework for large-scale, global applications.
The model is trained using a classifier setup, where a positive event is defined as the occurrence of an active fire detection within a grid cell on a given day.
Overall, the XGBoost configuration provided the most robust and interpretable results, balancing accuracy, computational cost, and operational suitability.
import pandas as pd
import os
import time
import joblib
import matplotlib.pyplot as plt
import xgboost as xgb
from xgboost import XGBClassifier, plot_importance
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, confusion_matrix, roc_auc_score, RocCurveDisplay
Load your data created from the previous notebook
df = pd.read_parquet("./data/training_data.parquet")
print("Loaded data shape:", df.shape)
print("Columns:", list(df.columns))Define your features and target
target_col = "AF"
feature_cols = [c for c in df.columns if c not in ["AF"]]
X = df[feature_cols]
y = df[target_col]
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=42, stratify=y
)
min(y_train)| Parameter | Description |
|---|---|
objective="binary:logistic" | Specifies that this is a binary classification problem. The model outputs probabilities between 0 and 1 using a logistic function. |
tree_method="hist" | Uses the histogram-based algorithm to speed up training, especially for large datasets. It groups feature values into discrete bins instead of processing each unique value. |
n_estimators=300 | The number of boosting rounds (trees) to build. Higher values can improve accuracy but may increase training time or risk overfitting. |
max_depth=8 | Maximum depth of each decision tree. Deeper trees can model more complex patterns but may overfit the data. |
learning_rate=0.1 | Shrinks the contribution of each tree. Lower values make learning slower but more robust; typically balanced with n_estimators. |
subsample=0.8 (optional) | Fraction of the training data used for each tree. Prevents overfitting and increases generalization. Default is 1.0 (use all data). |
colsample_bytree=0.8 (optional) | Fraction of features (columns) randomly sampled for each tree, adding diversity and reducing overfitting. |
eval_metric="logloss" | Evaluation metric for binary classification. Measures the model’s prediction accuracy in terms of probability calibration (lower is better). |
random_state=42 (optional) | Sets the random seed for reproducibility. Ensures consistent results when running the code multiple times. |
n_jobs=-1 | Uses all available CPU cores for parallel computation, speeding up training. |
# -----------------------------
# TRAIN CLASSIFIER
# -----------------------------
print("Training XGBoost Binary Classifier...")
start_time = time.time()
model = XGBClassifier(
objective="binary:logistic",
tree_method="hist",
n_estimators=300,
max_depth=8,
learning_rate=0.1,
# subsample=0.8,
# colsample_bytree=0.8,
eval_metric="logloss",
# random_state=42,
n_jobs=-1
)
model.fit(X_train, y_train)
end_time = time.time()
print(f"✅ Training completed in {end_time - start_time:.2f} seconds")
# -----------------------------
# EVALUATE MODEL
# -----------------------------
y_pred = model.predict(X_test)
y_proba = model.predict_proba(X_test)[:, 1]
print("\nClassification report:")
print(classification_report(y_test, y_pred, digits=3))
roc_auc = roc_auc_score(y_test, y_proba)
print(f"ROC-AUC: {roc_auc:.3f}")
# Confusion matrix
print("\nConfusion Matrix:")
print(confusion_matrix(y_test, y_pred))
# -----------------------------
# PLOT ROC CURVE & FEATURE IMPORTANCE
# -----------------------------
os.makedirs("./outputs", exist_ok=True)
plt.figure(figsize=(6, 5))
RocCurveDisplay.from_estimator(model, X_test, y_test)
plt.title("ROC Curve")
plt.savefig("./outputs/POF_ROC.png", dpi=300)
plt.show()
plt.close()
plt.figure(figsize=(10, 6))
plot_importance(model, max_num_features=20, importance_type="gain")
plt.title("Feature Importance (Gain)")
plt.tight_layout()
plt.savefig("./outputs/POF_importance.png", dpi=300)
plt.show()
plt.close()
# -----------------------------
# SAVE MODEL
# -----------------------------
out_model = "./data/POF_model.joblib"
joblib.dump(model, out_model, compress=3)
print(f"Model saved → {out_model}")