The visualization shows t-SNE embeddings of tree species classified by a DualInceptionTime model trained on time series from Sentinel-2 and Planet imagery.
Correct predictions are shown as circles, misclassifications as crosses. Colors correspond to species.
To support error analysis, class-wise centroids were computed using a kernel density estimation (KDE) on the 3D t-SNE embeddings of correctly classified samples. For each class, the centroid corresponds to the point with the highest local density, capturing the most representative embedding region. Then, for each misclassified point, the three nearest centroids are identified using Euclidean distance, enabling interpretation of inter-class confusion and semantic similarity in the learned latent space.
Bressant, C., Wenger, R., Puissant, A., Herrault, P-A., Data-driven Explainability of Urban Tree Species Classification from Deep Learning Models and Satellite Image Time Series, To be submitted.
Latil, M., Wenger, R., Michéa, D., Puissant, A., Forestier, G., Urban tree species time series classification using multimodal satellite imagery, IEEE JURSE 2025, Accepted.