Improving Urban Tree Species Classification with High Resolution Satellite Imagery and Machine Learning

In the context of climate change, reducing heatwave and air pollution are major challenges by using nature-based solutions. Urban greening helps to limit heat islands and promote resilience and trees also offer many other advantages in terms of making our cities more sustainable. This study explores the potential of multi-resolution imagery (Pléiades, PlanetScope and Sentinel-2) to map urban tree species in the Strasbourg Urban Area, France. We propose two object-oriented approach, one based on theorical crown buffer and the other on a watershed segmentation using a Random Forest algorithm to map the ten most representative tree species. Our results showed promising improvements in the overall accuracy and F1-Score of urban tree species classification using very high resolution imagery (Pléiades), high spatial resolution time series (PlanetScope and Sentinel). The method also successfully demonstrated a logical distribution of errors based on the trees intrinsic characteristics (species, distribution pattern, location) and the effect of data pre-processing. All these findings underscore the complexity of tree species classification in urban environments and suggest areas for future methodological enhancements.