📌 Project Overview This project aims to predict Urban Heat Island (UHI) hotspots using machine learning models trained on satellite imagery and weather datasets. The goal is to achieve high accuracy (98%+ R² score) using Random Forest, LightGBM, and stacking models.
📂 UHI_Model/ │── 📂 data/ # Raw and processed datasets (if needed) │── 📂 models/ # Trained ML models (.pkl files) │── 📂 predictions/ # Predicted results in Excel (.xlsx files) │── 📂 src/ # Python scripts for different models │ │── model1.py # Model 1 training/testing script │ │── model2.py # Model 2 training/testing script │ │── model3.py # Model 3 training/testing script │── README.md # Project description │── .gitignore # Ignored files │── requirements.txt #Requirements
- Clone this repository:
git clone <your-repo-url> cd ML-Project
Run a specific model
python src/model1.py python src/model2.py python src/model3.py
🛠 Requirements numpy==1.24.3 pandas==2.0.3 geopandas==0.14.1 rasterio==1.3.8 scikit-learn==1.3.2 lightgbm==4.1.0 xgboost==2.0.3 catboost==1.2.2 matplotlib==3.8.2 joblib==1.3.2 optuna==3.4.0 openpyxl==3.1.2
Libraries: pandas, numpy, scikit-learn (if needed)
⚙️ Preprocessing & Feature Engineering Date Processing: Extract Year, Month, Day, Hour from timestamps.
Fourier Transforms: Creates Sin_Hour and Cos_Hour for time-based learning.
Spatial Merging: Combines NDVI, LST, and Weather data.
Feature Engineering:
NDVI_LST_Ratio = NDVI / (LST + 1e-6)
NDVI_Squared = NDVI²
LST_Squared = LST²
LST_Log = log(1 + LST)
🏋️ Model Training
LightGBM Model
param_grid = { 'num_leaves': [20, 30, 40], 'max_depth': [10, 20, -1], 'learning_rate': [0.01, 0.05, 0.1], 'n_estimators': [100, 200, 300] } lgbm = lgb.LGBMRegressor() random_search = RandomizedSearchCV(lgbm, param_grid, scoring='r2', cv=3, n_iter=20) random_search.fit(X_train, y_train)
PICTURE
Stacking Model
stacked_model = StackingRegressor( estimators=[ ('lgbm', lgb.LGBMRegressor(n_estimators=1000, learning_rate=0.03)), ('xgb', xgb.XGBRegressor(n_estimators=1000, learning_rate=0.03)), ('cat', cb.CatBoostRegressor(iterations=1000, learning_rate=0.03, verbose=0)) ], final_estimator=Ridge(alpha=1.0) ) stacked_model.fit(X_train, y_train)
🎯 Model Performance Model R² Score RMSE Random Forest 0.965 0.52 LightGBM 0.982 0.38 Stacking (Best) 0.9935 0.25
📤 Generating Predictions
X_submission = submission_template[['Longitude', 'Latitude']] X_submission = X_submission.merge(final_df, on=['Longitude', 'Latitude'], how='left') submission_template['Predicted_UHI_Index'] = best_model.predict(X_submission) submission_template.to_csv("submission.csv", index=False)
📈 Feature Importance
import matplotlib.pyplot as plt feature_importance = best_model.feature_importances_ feature_names = X_train.columns plt.figure(figsize=(10, 5)) plt.barh(feature_names, feature_importance) plt.xlabel("Feature Importance") plt.ylabel("Features") plt.title("Feature Importance of LGBM Model") plt.show()
📌 To-Do / Future Work Incorporate Deep Learning Models (CNNs for Satellite Images).
Enhance Feature Engineering with Advanced Climate Indicators.
Fine-tune Hyperparameters for Further Optimization.
Integrate a Web Interface for Visualization.
📝 Conclusion This project effectively predicts UHI hotspots using ML models, with the stacking model achieving 99.35% accuracy. The model is designed for real-world applications in urban climate management and environmental planning.
📬 For any queries, feel free to reach out! 🚀