Roof material classification is a key task for applications such as urban heat island studies, energy modeling, and resource management. This study focuses on leveraging machine learning (ML) algorithms to achieve accurate roof material classification, despite the challenges posed by noisy training data. BD TOPO, a widely used database providing roof material information at the building footprint level, is known for its inaccuracies and incompleteness. Recognizing the potential of ML models to produce reliable classifications even when trained on imperfect data, we designed a framework to improve upon BD TOPO’s baseline accuracy.

The methodology includes advanced preprocessing steps to enhance input data quality. For each building footprint in BD TOPO, statistical metrics—mean, median, and standard deviation—were extracted from multi-seasonal Pleiades imagery (winter and summer). These features were corrected for the satellite's off-nadir acquisition angle to ensure alignment with building boundaries. Additional spectral indices (e.g., NDVI and GNDVI) and band ratios were derived to capture material-specific properties and seasonal variability.

To address the inherent noise in the BD TOPO labels, we employed XGBoost and Random Forest, two algorithms known for their robustness, to label inaccuracies. Our hypothesis is that training these models on noisy yet extensive datasets will result in classifications that are more accurate and consistent than the original BD TOPO labels.

While results are pending, we expect this approach to effectively distinguish roof materials and enhance the accuracy of urban mapping. The proposed framework demonstrates the potential of combining multi-seasonal data and ML techniques for scalable and reliable roof material classification, with broader applications in sustainability and urban planning