LeafMosaic: a pipeline for exploring community‐level plant classification from ecological images

LeafMosaic is a lightweight, end-to-end pipeline for exploring community‐level plant classification from ecological images. It tiles a high-resolution image into small patches, extracts deep features via pretrained CNNs, projects those features with UMAP, discovers unsupervised clusters, and finally classifies each tile with a pretrained model. The result is both unsupervised cluster outlines and supervised species predictions overlaid back on your original image. Features Tiling: break a large image into 100×100 px patches Filtering: discard mostly‐empty/white tiles (> 50% background) Feature Extraction: CNN-based: use DenseNet or VGG backbones (ImageNet weights) SVD-based: lightweight Singular Value Decomposition for fast feature extraction Dimensionality Reduction: UMAP in 2D and 3D for verification Clustering: HDBSCAN on 2-D embeddings to find visual groups Classification: Keras model predicts dominant species per tile (proprietary training set and transfer learning enabled) Overlays: draw colored outlines for clusters and top predictions back onto the original image, with a clear legend