Road Shade Mapping Using Car Sensors

Sammanfattning

This thesis presents a comprehensive study on the development of a shade mapping model for road segments using illuminance measurements collected from a fleet of vehicles. Two distinct machine learning approaches were investigated: (i) classification models that assign discrete labels (e.g., shaded versus non-shaded) and (ii) regression models that predict continuous shading levels to capture subtle variations in environmental conditions. Extensive exploratory data analysis and data processing ensured that the raw sensor data were well-prepared for subsequent model development. In the classification case, several algorithms were evaluated using a dataset constructed from hand-labeled examples. For instance, the Label Spreading algorithm achieved an accuracy of 90%, while other classification methods reached approximately 97% accuracy, accompanied by robust precision and recall metrics. Due to the limited volume of hand-labeled data, semi-supervised learning techniques were explored to incorporate unlabeled data; however, the inherent challenges of manual labeling constrained the potential benefits of these approaches. Conversely, the regression models were trained on an abundant sensor dataset. Two regression methods were compared: Random Forest and Histogram Gradient Boosting. The Random Forest model consistently outperformed Histogram Gradient Boosting by achieving lower error rates and a higher coefficient of determination (with an R2 value of 0.81), though at the cost of a longer training time. Cross-validation confirmed that the regression models generalize well to unseen data, demonstrating that where ample data exist, regression is effective at capturing the complex variability of road shading. Overall, the generated shade map shows strong potential for practical applications. When integrated with road surface temperature models, it can enhance the detection of icy or slippery road conditions, thereby contributing to improved road safety and more efficient maintenance strategies. Future work will prioritize refining regression models through enhanced data processing, improved feature selection, and rigorous hyperparameter optimization, while critically reassessing labeling strategies for classification tasks.