Data-driven interpretability of urban tree species classification from deep learning models and Satellite Image Time Series

Urban trees provide critical ecosystem services, yet species-level information remains limited due to the cost of conventional field inventories and the general lack of data on trees located on private land. Recent advances in remote sensing and deep learning offer new opportunities for large-scale tree species classification, particularly through Satellite Image Time Series (SITS). In this study, we evaluate the performance of InceptionTime-based models trained on Sentinel-2 and PlanetScope time series, used individually and jointly via a fusion-based architecture (Dual-InceptionTime), to classify 10, 15, and 20 urban tree species in Strasbourg, France. We demonstrate the benefit of combining the complementary spatial, spectral, and temporal properties of Sentinel-2 and PlanetScope data. The fused model achieved high performance, reaching up to 0.727 ± 0.004 accuracy for 10 species, and 0.656 ± 0.005 for 20 species. In contrast, models trained on each sensor independently showed similar results. We also analysed the consistency of predictions across model configurations to investigate whether recurrent patterns could be observed in classification errors. Based on these observations, we developed a data-centered interpretability framework that links classification outputs to phenological, morphological, and environmental variables computed for each tree. Using Generalised Linear Mixed Models (GLMMs), we found that intra-species phenological anomalies, inter-species similarity, and crown changes significantly contributed to classification errors. Also, trees located in dense urban environments were more likely to be correctly classified. These findings support the relevance of multi-source SITS and data-driven interpretability as a foundation for accurate and interpretable urban tree species mapping tailored to end-user needs.