Advancing roof material mapping: a multi-class deep learning approach on very-high-resolution aerial imagery
TL;DR
This study establishes a methodological baseline for multi-class roof material classification from very-high-resolution RGB imagery acquired over Namur, Belgium, and provides insight into prediction reliability by identifying areas associated with uncertain or implausible classifications.
Abstract
Abstract. Accurate classification of roof materials is important for urban planning, environmental monitoring, and circular economy applications. We investigate deep learning approaches for the multi-class classification of 12 roof materials using very-high-resolution (5 cm) red-green-blue (RGB) aerial imagery acquired over Namur, Belgium, representative of data commonly available in operational contexts. Under limited training data conditions, a classification framework combining contrast enhancement, texture-oriented pre-processing, lightweight convolutional neural networks, and ensemble modeling is evaluated. The proposed approach achieves an F1-score of 0.80±0.01, with class-wise F1-scores ranging from 0.68±0.04 to 0.93±0.05. Model variability and class confusion are reduced. Spatial uncertainty analysis further provides insight into prediction reliability by identifying areas associated with uncertain or implausible classifications. Despite limited training data and increased class complexity, the results remain within the range of state-of-the-art performance. The study establishes a methodological baseline for multi-class roof material classification from very-high-resolution RGB imagery.